View Full Version : R520 Infomania
So let's go back the other way for some speculation (no, I'm not changing my mind). We haven't talked about memory much of late. Let's give "the smart money" it's due and say it is a 24-pipe beastie along conventional r3/4xx lines with SM3.0 bolted-on. Where are they going to get 900mhz DDR (roughly 50% increase from X850 top-end) in quantity to feed that critter?
Or is Rambus XDR the answer?
Or a 512-bit bus (which, I think, everyone agrees ain't gonna happen --but then I think there was generally shock when Parhelia and then R300 showed up with it. . and then when NV30 didn't).
Because if you say 'no' to all three, you are left back where I am (See, I said I wasn't changing my mind. :D ) Altho, to be fair, that is more "Baumann School of Eyeblink Semaphore Translation". Also "finally, revenge of the TBDRs". I should also add that once when Wavey had hay-fever I was convinced that ATI was going to buy Sony to get the licensing rights to NV's tech. :wink:
digitalwanderer
23-Mar-2005, 00:39
Why do you think it's only 24 pipes? :|
Megadrive1988
23-Mar-2005, 00:48
At the same time this question of Rambus tech is pretty intriguing. Wouldn't it be interesting if NVidia peeled away from GDDR tech to implement XDR in 2006 on PC parts, having gained experience with this tech by working on PS3?
Jawed
G70 is using Rambus's yellowstone memory bus.
G70=PS3
source ? :?
kemosabe
23-Mar-2005, 00:55
So let's go back the other way for some speculation (no, I'm not changing my mind). We haven't talked about memory much of late. Let's give "the smart money" it's due and say it is a 24-pipe beastie along conventional r3/4xx lines with SM3.0 bolted-on. Where are they going to get 900mhz DDR (roughly 50% increase from X850 top-end) in quantity to feed that critter?
Or is Rambus XDR the answer?
Or a 512-bit bus (which, I think, everyone agrees ain't gonna happen --but then I think there was generally shock when Parhelia and then R300 showed up with it. . and then when NV30 didn't).
Because if you say 'no' to all three, you are left back where I am (See, I said I wasn't changing my mind. :D ) Altho, to be fair, that is more "Baumann School of Eyeblink Semaphore Translation". Also "finally, revenge of the TBDRs". I should also add that once when Wavey had hay-fever I was convinced that ATI was going to buy Sony to get the licensing rights to NV's tech. :wink:
LOL, I can tell you are enjoying this thread. Well I'm sticking to the uninspiring scenario of 16 pipes with performance boost coming from the mishmash of sources already discussed, not the least of which is the higher core clock.
And I will consolidate that conjecture by speculating that Mobility R520 will be hot on the heels of its desktop version and that its flagship version will also feature 4 quads and high clocks. It's no coincidence that you hear very little Mobility X800 hype. This is the market that continues to boom (http://www.theregister.co.uk/2005/03/21/euro_notebook_market_jan_05/) and will soon surpass desktop and where ATI has recently been experiencing slow erosion of market share. That 0.09u low-k advantage (lower power, less heat) is going to have to pay big dividends somewhere, and I'm betting ATI tries to drive OEM notebook sales in a big way in 2005 with its new architecture.
Why do you think it's only 24 pipes? :|
Me, I don't think it is even 24. :D "The smart money" seems to think it is 24p/8vs. They seem to be basing this on doing some math from NV40 (because of SM3.0) and the 300-350m transistors number. NV40 (220m) x 1.5 (24 pipes) = 330m transistors.
LOL, I can tell you are enjoying this thread.
Hell, yes. :) And watch how fast I shift if we get a credible report of XDR or a 512-bit bus on R520. :wink: That latter is interesting, because for all the reports of how nasty it would be to do (which, you might recall, the same things were said to be prohibitive for 256-bit pre-Parhelia, when tiny little Matrox pulled it off), the savings of being able to use 450mhz memory at 512MB quantities has got to be pretty impressive to balance against the cost of implementing the wider bus.
But then you wind back to, "and what are the scenarios that require all that bandwidth compared to today?" and 16 pipes looks better again. The Conventional Wisdom (24p, insane bandwidth) is always the safe, easy way to go --right up until it isn't.
Unknown Soldier
23-Mar-2005, 06:11
Well Speculation has it like this.
[Speculation]
R520
300-350 Million Transistors
24 pipes
600 Mhz Core
700 Mhz DDR3
[Dave's hints]
300 Million
16 pipes
--------------------------------------
What I want to know is this.. When is the NDA up for both ATI and Nvidia???
US
Is the reason the GF6 series takes such a hit with Far Cry's HDR because of more/larger memory calls, or b/c of a core limitation (FP blending essentially halves the GPU's "fillrate," IIRC)?
Also, is it possible ATi will dedicate a larger proportion of those 300-350M transistors to onboard cache? Is a large onboard cache really necessary condsidering video card's outrageous bandwidth, or does the cache's lower latency become more important with branching and more shader calculations?
Chalnoth
23-Mar-2005, 07:21
From what I understand, GPU's typically use vastly less cache than CPU's, due to the way they handle data. And I don't think we'll ever get solid information about the onboard caches. That info's never been released to the public, as far as I can remember.
not knowing hat the r520 will bring to the table sucks because it screws up any upgrade option i have.
wireframe
23-Mar-2005, 08:23
not knowing hat the r520 will bring to the table sucks because it screws up any upgrade option i have.
Why is that? Or was this a desperate attempt at getting someone in-the-know to spill the beans to help out poor Hovz? :P
Hey! Wait a minute! You don't know what the next Nvidia part will bring either, do you? So it seems you have already chosen ATI. That should make the upgrade option very simple. ;)
....no. not at all.
as far as i know r520 is 'suppose' to launch sometime b4 nvidias next chip. i currently have a p4 2.6c(which is absolute garbage for any game) and a 9800 pro. buying an agp motherboard now would be retarded, and yet if i buy a pci express motherboard i have to buy a new video card right now. see the bind?
wireframe
23-Mar-2005, 09:21
....no. not at all.
as far as i know r520 is 'suppose' to launch sometime b4 nvidias next chip. i currently have a p4 2.6c(which is absolute garbage for any game) and a 9800 pro. buying an agp motherboard now would be retarded, and yet if i buy a pci express motherboard i have to buy a new video card right now. see the bind?
I'm all too familiar with that bind. If I knew that the x850XT is where the game is at for the next 12 months I would buy one along with a PCIe mainboard right now. The fact that I have a 6800 Ultra now, hence my upgrade options are more limited, coupled with all this talk about R520 is helping me hold back. There is no PCIe mainboard that really excites me for the socket 939 platform (or other, really) yet so this is probably a good thing.
FWIW, I was using a P4 2.6C with my 6800 Ultra before switching over to an Athlon 64 3500+ and there really is not a huge difference. Graphics is still a bottleneck and it seems those games that really push the CPU are almost insatiable. Sure, it helps in benchmarking when you want to stretch the peaks and soften the gullies, but in actual gameplay it's not a vast difference. That means that if you want more performance right now and don't want to buy a whole new system, you could get a good experience by upgrading your Radeon 9800 and maybe even your 2.6C to a 3.4.
Megadrive1988
23-Mar-2005, 09:26
ok lets say that both R520 and R580 have 16 pixel pipelines (and 8 vs)
if the R580 is going to have triple the TMUs, or texture sampling or ALUs or whatever, per pipe, compared to the R420 and R480, then maybe R520 will have twice what the R420 and R480 have, thus 2/3 of the the R580 will have.
does that make any sense, DaveB and Jawed ?
....no. not at all.
as far as i know r520 is 'suppose' to launch sometime b4 nvidias next chip. i currently have a p4 2.6c(which is absolute garbage for any game) and a 9800 pro. buying an agp motherboard now would be retarded, and yet if i buy a pci express motherboard i have to buy a new video card right now. see the bind?
I'm all too familiar with that bind. If I knew that the x850XT is where the game is at for the next 12 months I would buy one along with a PCIe mainboard right now. The fact that I have a 6800 Ultra now, hence my upgrade options are more limited, coupled with all this talk about R520 is helping me hold back. There is no PCIe mainboard that really excites me for the socket 939 platform (or other, really) yet so this is probably a good thing.
FWIW, I was using a P4 2.6C with my 6800 Ultra before switching over to an Athlon 64 3500+ and there really is not a huge difference. Graphics is still a bottleneck and it seems those games that really push the CPU are almost insatiable. Sure, it helps in benchmarking when you want to stretch the peaks and soften the gullies, but in actual gameplay it's not a vast difference. That means that if you want more performance right now and don't want to buy a whole new system, you could get a good experience by upgrading your Radeon 9800 and maybe even your 2.6C to a 3.4.
i get about 28 fps in ut2k4 onslaught and 50 or so in 8+ dms. an athlon 64 isnt going to be a noticeable improvement? i find that hard to believe.
wireframe
23-Mar-2005, 09:44
i get about 28 fps in ut2k4 onslaught and 50 or so in 8+ dms. an athlon 64 isnt going to be a noticeable improvement? i find that hard to believe.
OK, UT24K is one of those titles that can use the CPU and, especially, an AMD one (seems to hate NetBurst). However, in my experience this was only with Onslaught and Assault with bots. So in this particular title you could probably get much better performance under some circumstances by only upgrading to a faster CPU (preferrably Athlon 64).
This is off topic so we should probably leave it there. We could start a new thread or you could PM me if you want to pick me for numbers in going from 2.6c to a 3500+.
caboosemoose
23-Mar-2005, 09:53
ok lets say that both R520 and R580 have 16 pixel pipelines (and 8 vs)
if the R580 is going to have triple the TMUs, or texture sampling or ALUs or whatever, per pipe, compared to the R420 and R480, then maybe R520 will have twice what the R420 and R480 have, thus 2/3 of the the R580 will have.
does that make any sense, DaveB and Jawed ?
No it doesn't because the numbers are 16-1-1-1 for R520 and 16-1-3-1 for R580 so whatever the third figure refers to is most definitely x3 on R580.
ok lets say that both R520 and R580 have 16 pixel pipelines (and 8 vs)
if the R580 is going to have triple the TMUs, or texture sampling or ALUs or whatever, per pipe, compared to the R420 and R480, then maybe R520 will have twice what the R420 and R480 have, thus 2/3 of the the R580 will have.
does that make any sense, DaveB and Jawed ?
I don't think the middle digit of the codenames has any relevance.
I'm still trying to work out how ATI can come up with the codename RV410. Why would something based on R420 get a lower number?
Anyway, "I'm just zis guy, you know"? Or as Manuel would say "I know nuhthinq!"
And DaveB prefers to keep us guessing. Apparently it improves the advertising revenue for B3D...
Jawed
Dave Baumann
23-Mar-2005, 10:11
Actually, some people just aren't listening!
@caboosemoose, are there any indications about new/improved features in that document? E.g. an improved AF or AA implementation? Thanks!
Anyone consider the possibility that ATI plans to make it a clock ramped 16 pipe chip initially but has an extra quad or these "fat pipes" ready depending on nvidia's approach. That is what they did with the R420 after all.
caboosemoose
23-Mar-2005, 10:41
All the interesting info I have on R520 and R580 (and it's presented as is, take it or leave it) I've pretty much already shared.
caboosemoose
23-Mar-2005, 10:51
Anyone consider the possibility that ATI plans to make it a clock ramped 16 pipe chip initially but has an extra quad or these "fat pipes" ready depending on nvidia's approach. That is what they did with the R420 after all.
Not really, we're all interested in the chip being produced - whether bits of it are switched off at launch isn't really the point - plus, it's hard to be sure that ATI originally intended to launch R420 with a quad disabled. I guess they might have done if NV40 had been a total blow out, but despite the failure of the FX that was never terribly likely.
Here's my off-the-top-of-my-head go for R520. I'm really guessing for a whole bunch of it.
Just under 300M transistors
Hiring of Jeff Fu to create a decent cooler :lol:
4 fragment quads
8 vertex units
New logic connecting the two
Two vec4 and two scaler ALUs per fragment unit
All fragment ALUs have access to the texture sampler for texture addressing
Vertex unit has four separate ALUs (2 vec4 and 2 scalar) and all can use the texture unit
Accelerated DST ops
Fragment and vertex units both have a bilinear texture sampler
One-to-one mapping (but through a crossbar) for output fragments into ROPs
Fully programmable AA sample patterns
Up to four loops through the ROPs for AA
Tweaked angle-adaptive AF
Ability to force an AF level for the entire scene
NPOT texture support
New buffer design for fragment loopback (buffer more fragments)
256-bit memory bus and support for new DRAM densities and speeds
Tweaked/New Hyper-Z to reject more pixels per clock
No more F-Buffer
More granularity in the memory crossbar
:!:
Dave Baumann
23-Mar-2005, 11:23
Not entirely sure "crossbars" will be the order of the day here.
Well, probably more a fragment switch than a crossbar, but still new logic compared to R420. Still a crossbar for the memory controller, though.
Dave Baumann
23-Mar-2005, 11:54
Not necessarily.
How about you guys stop talking cryptic and just put the real spec. in my PM box.
thanks and have nice day
caboosemoose
23-Mar-2005, 12:03
How about you guys stop talking cryptic and just put the real spec. in my PM box.
thanks and have nice day
Yeah, stop talking cryptic:
"Just under 300M transistors
Hiring of Jeff Fu to create a decent cooler Laughing
4 fragment quads
8 vertex units
New logic connecting the two
Two vec4 and two scaler ALUs per fragment unit
All fragment ALUs have access to the texture sampler for texture addressing
Vertex unit has four separate ALUs (2 vec4 and 2 scalar) and all can use the texture unit
Accelerated DST ops
Fragment and vertex units both have a bilinear texture sampler
One-to-one mapping (but through a crossbar) for output fragments into ROPs
Fully programmable AA sample patterns
Up to four loops through the ROPs for AA
Tweaked angle-adaptive AF
Ability to force an AF level for the entire scene
NPOT texture support
New buffer design for fragment loopback (buffer more fragments)
256-bit memory bus and support for new DRAM densities and speeds
Tweaked/New Hyper-Z to reject more pixels per clock
No more F-Buffer
More granularity in the memory crossbar "
I wish someone would pin their colours to the mast instead of vagaries like this.
Broken Hope
23-Mar-2005, 13:15
....no. not at all.
as far as i know r520 is 'suppose' to launch sometime b4 nvidias next chip. i currently have a p4 2.6c(which is absolute garbage for any game) and a 9800 pro. buying an agp motherboard now would be retarded, and yet if i buy a pci express motherboard i have to buy a new video card right now. see the bind?
I'm all too familiar with that bind. If I knew that the x850XT is where the game is at for the next 12 months I would buy one along with a PCIe mainboard right now. The fact that I have a 6800 Ultra now, hence my upgrade options are more limited, coupled with all this talk about R520 is helping me hold back. There is no PCIe mainboard that really excites me for the socket 939 platform (or other, really) yet so this is probably a good thing.
FWIW, I was using a P4 2.6C with my 6800 Ultra before switching over to an Athlon 64 3500+ and there really is not a huge difference. Graphics is still a bottleneck and it seems those games that really push the CPU are almost insatiable. Sure, it helps in benchmarking when you want to stretch the peaks and soften the gullies, but in actual gameplay it's not a vast difference. That means that if you want more performance right now and don't want to buy a whole new system, you could get a good experience by upgrading your Radeon 9800 and maybe even your 2.6C to a 3.4.
i get about 28 fps in ut2k4 onslaught and 50 or so in 8+ dms. an athlon 64 isnt going to be a noticeable improvement? i find that hard to believe.
I get a minimum of around 40fps in onslaught, usually I hit the online cap of 85, DM I usually never go below 60 and usually it's 100fps+
That's with an A64 3500+ and an X800XT. 1152x854 res 4xAA 16xAF
....no. not at all.
as far as i know r520 is 'suppose' to launch sometime b4 nvidias next chip. i currently have a p4 2.6c(which is absolute garbage for any game) and a 9800 pro. buying an agp motherboard now would be retarded, and yet if i buy a pci express motherboard i have to buy a new video card right now. see the bind?
Gosh, I was still considering my 2.4c@2.88 kinda mid-rangie. Played D3, FC, and HL2 on it without any problems at 12x10. YMMV.
"ebay is your friend" on those interimye deals if you feel you need to MOVE RIGHT NOW. Just don't buy top-o'-the-market on the graphics card so you only take a moderate hair-cut on resale. I mean, do you really want to pay the premium to get ahold of these new cards the first month or two anyway?
digitalwanderer
23-Mar-2005, 14:49
Actually, some people just aren't listening!
What? :|
silence
23-Mar-2005, 15:29
Actually, some people just aren't listening!
What? :|
i bet digi is all ears. :lol:
DegustatoR
23-Mar-2005, 15:42
3 TMUs per pipe are just strange today if you ask me. If that's just todays TMUs and not some kind of new ALUs able to do something more than just sampling textures then it's a wasted silicon... I'll believe it only when i see it...
Chalnoth
23-Mar-2005, 16:07
It's not strange. It's down right idiotic.
r580 may also be the first part supporting fp filtering it may has soemthing to do with it
digitalwanderer
23-Mar-2005, 16:27
What is all this weird "stoastic AA" or whatever whispering I'm hearing about? :|
You hear whispering? I don't think they will have stochastic AA. Sireric said its only an quality improvement to current algorithm if you do 16 or more samples
digitalwanderer
23-Mar-2005, 16:35
You hear whispering?
Yeah, I can't shake this bloody headcold to save me arse!
I don't think they will have stochastic AA.
What IS "stochastic AA"?
DegustatoR
23-Mar-2005, 16:36
r580 may also be the first part supporting fp filtering it may has soemthing to do with it
It may, yes. But i find it hard to explain how one FP16-bi/tri-able TMU can produce 3 FX16-bi/tri texels... The question is are we talking about tripling the number of texture samples taken per tick or we're talking about tripling something else (f.e. transistors number)?
Another possibility is if ATI's found a way to do texture sampling in a math ALUs so this texture sampling is just a 'side product' and doesn't really cost anything in transistor numbers (hard to believe it actually).
You hear whispering?
Yeah, I can't shake this bloody headcold to save me arse!
I don't think they will have stochastic AA.
What IS "stochastic AA"?
being able to do fully random sample patterns (jittering) per pixel
if you google it you will get many information about it
digitalwanderer
23-Mar-2005, 16:42
Thanks tEd!
I did try googling it, but got a bit too much info...."being able to do fully random sample patterns (jittering) per pixel", sums it up very nicely and in a way I can understand.
Thank you.
So I'm hearing rumors of a new AA, would it have to be hardware based? I thought the R3xx/4xx series had programable AA patterns or some such. :|
Tim Murray
23-Mar-2005, 16:43
What is all this weird "stoastic AA" or whatever whispering I'm hearing about? :|
you sure do know how to summon me to a thread.
and no, it won't have stochastic AA, because it needs AT THE VERY LEAST 16 samples. and I don't think we're going to see 16x AA that is usable in this generation. and you really need more samples than that, because I've heard that 16x stochastic won't really look much better than ATI's current 6x (and in many cases will actually look worse).
digitalwanderer
23-Mar-2005, 16:47
you sure do know how to summon me to a thread.
It's a gift. 8)
Don't you have school or something, or are you on spring break too? And what's this about 16 sample patterns and why can't it have 'em?
So I'm hearing rumors of a new AA, would it have to be hardware based? I thought the R3xx/4xx series had programable AA patterns or some such. :|
I don't know anything about a new AA algorithm but i would strongly assume it would be HW based if there is one.
r3xx/r4xx do have programmable AA as they can have different sample patterns per frame and within the 12*12 sample grid. It can't do different sample patterns per pixel
wireframe
23-Mar-2005, 16:55
and no, it won't have stochastic AA, because it needs AT THE VERY LEAST 16 samples. and I don't think we're going to see 16x AA that is usable in this generation. and you really need more samples than that, because I've heard that 16x stochastic won't really look much better than ATI's current 6x (and in many cases will actually look worse).
Let's forget the connection between this rumor and the R520 for a minute and focus on the stochastic AA sampling. I don't understand why you would want this for MSAA if it requires 16+ samples. At 16 samples you are looking very good with a fixed and ordered sampling pattern already, what added value would a stochastic sampler possibly give you?
The only benefit I can think of for a stochastic sampler is in texture sampling whereby introducing a random element you can minimize systematic aliasing in the form of moire. Or perhaps you were already talking about x16 AF and AA was a typo or just a generlized term (normally when I see AA I think MSAA or SSAA)?
Am I completely off base here?
ShootMyMonkey
23-Mar-2005, 17:01
and no, it won't have stochastic AA, because it needs AT THE VERY LEAST 16 samples.
More than that is that stochastic sampling means very small scale variations from the scale of a regular grid sampling, and if you're doing straight rasterization, where does the resolution refinement for those tiny variations come from? With a raytracer, it makes sense because the only thing in the way of fine-grained variance is fp precision. With a rasterizer, the best I can think of would be extreme supersampling of a tile in local storage prior to issuing out to the framebuffer and randomly selecting among a fraction of the supersampled pixels for each resulting final pixel. At which point, you beg the question of, if you're applying such extreme supersampling anyway, where's the benefit to stochastic AA anyway? Unless you're really deathly concerned about the axial bias, I don't see the point.
r580 may also be the first part supporting fp filtering it may has soemthing to do with it
It may, yes. But i find it hard to explain how one FP16-bi/tri-able TMU can produce 3 FX16-bi/tri texels... The question is are we talking about tripling the number of texture samples taken per tick or we're talking about tripling something else (f.e. transistors number)?
Even if those 3tmu number is true we don't know what kind of capability one of those tmu has
Besides r580 being the first (ati)part to do fp filtering it may also be the first (ati)part supporting bilinear filtering on vertex textures?
btw tomorrow ati will have conference call for the 2nd quarter. Maybe some new information coming along with it
We just reached that "one level down in detail below where I am actually comfortable that I'm not being a total fool", so just pat me on the head and smile if this is too godawful. But I was under the impression that currently they can do single-cycle for 2xaa and loop back once or twice more to do 4x and 6x. Is that right? What would it take transistor-wise, or more units-wise, to make 4x single-cycle? (this would be the part where I'm deathly afraid I just asked a stupid question). Would it be a worthwhile thing to do, particularly if you were thinking of taking advantage of the bump-up to 512mb to add 8xaa (which I don't know they are thinking of doing). From a marketing perspective, 8xaa and (again, possible faux pas alert), "our 4xaa is as fast as the other guys 2xaa, and our 8xaa is as fast as their 4xaa!" might be nice. Might it even bring 16x Temporal into play as a reasonable possibility performance-wise?
caboosemoose
23-Mar-2005, 19:41
We just reached that "one level down in detail below where I am actually comfortable that I'm not being a total fool",
Yeah, I know that level. Unfortunately it keeps bloody moving. Remember when all you need to know was pipelines and texture units, maybe a bit of filtering nous and you were in business. Bleh, all this ALU, co-issue, ROPs and accelerated DST ops just doesn't jive with me.
ShootMyMonkey
23-Mar-2005, 20:11
One of the nice things about a tech forum is that you can say something that's totally off base and someone is probably going to be there who knows better. I don't particularly mind if I say something that might be completely wrong about something with which I'm not entirely familiar, because that way, someone's likely to chime in and I'll get to learn something new in the process.
Chalnoth
23-Mar-2005, 21:34
We just reached that "one level down in detail below where I am actually comfortable that I'm not being a total fool", so just pat me on the head and smile if this is too godawful. But I was under the impression that currently they can do single-cycle for 2xaa and loop back once or twice more to do 4x and 6x. Is that right? What would it take transistor-wise, or more units-wise, to make 4x single-cycle?
Basically, you'd add in z-test units. Don't know how many more transistors this would require, but I doubt it would be that useful.
That is to say, with longer shaders, you won't even notice the fact that it's taking a few cycles to output a pixel. Since longer shaders are only going to get more common, this isn't a huge concern.
The only reason I say 512-bit is because of the idea of GDDR4 (coming real soon now, end of the year) and the bandwidth crunch that can only get worse with the next generation VPUs. If R580 has all that texturing power (3xTMUs per pipeline? One a floating point unit?) it's gonna need an awful lot of bandwith.
At the same time this question of Rambus tech is pretty intriguing. Wouldn't it be interesting if NVidia peeled away from GDDR tech to implement XDR in 2006 on PC parts, having gained experience with this tech by working on PS3?
Jawed
I remember, last year there were speculations about new memory interface for R520
http://www.beyond3d.com/forum/viewtopic.php?t=17663
and
http://www.beyond3d.com/forum/viewtopic.php?t=16376
Oh, you left out the best one: http://www.beyond3d.com/forum/viewtopic.php?t=19155
If I weren't so lazy and enjoyed speculating more than working at it, I'd collect all the juicy Baumann hints & speculations over the last six months into a single topic post and see what we see. Of course, they may not have all been about R520, and it can on occasion be deucedly difficult to tell when he's hinting and when he's engaging "in the great game" himself. Nevertheless.
Nice find geo. I remember, Reverend getting all emotional for no apparent reason. Blimey.
That Token Ring thingy looks fantastic don't it? Could Fast14 tech make this run at suitably hairy speeds?
So, would this be a component of a unified shader architecture, or could it come earlier in R520?
Jawed
Kaleidoscopes usually turn in a circle, don't they? :P Wavey once dropped that code-name out of the blue into a thread whose subject was R520 memory usage.
Then there is this: http://www.beyond3d.com/forum/viewtopic.php?t=19891 which no one ever commented on at the time. May be no relation, or may be power-saving related, or hell, I dunno. . .
EDIT: Errp! Now that I read to the bottom on that patent it might very well be a better way to increase yield by speed-binning chips more accurately. Tho wouldn't you expect the TSMC's of the world to be doing that kind of thing rather than ATI?
Then there is this: http://www.beyond3d.com/forum/viewtopic.php?t=19891 which no one ever commented on at the time. May be no relation, or may be power-saving related, or hell, I dunno. . .
http://www.beyond3d.com/forum/viewtopic.php?p=452121
Kaleidoscopes usually turn in a circle, don't they? :P Wavey once dropped that code-name out of the blue into a subject on R520 memory usage.
Hell, I don't care if it's just you and me on this one. Brilliant!
Then there is this: http://www.beyond3d.com/forum/viewtopic.php?t=19891 which no one ever commented on at the time. May be no relation, or may be power-saving related, or hell, I dunno. . .
As such, there exists a need for a method and apparatus that allows for determining a processing speed of an integrated circuit, wherein the integrated circuit may be on a wafer having a plurality of integrated circuits, wherein the performance speed is determined independent of process, voltage and temperature measurements and may be based on the execution of a system application on the integrated circuit. Moreover, there also exists a need for determination of processing speeds for integrated circuits of the plurality of integrated circuits to improve the binning process.
Seems to me that this is a little circuit on each die that allows the dies to be speed-binned really easily after they've been cut. This gives ATI a finer-grained speed binning capability instead of binning on a per-wafer basis.
Jawed
Then there is this: http://www.beyond3d.com/forum/viewtopic.php?t=19891 which no one ever commented on at the time. May be no relation, or may be power-saving related, or hell, I dunno. . .
Commented in this thread (http://www.beyond3d.com/forum/viewtopic.php?t=19991). It's only about speed binning of chips.
Just to add... if you look at the date of rwolf's post and his (incorrect - no offense, rwolf) line of speculation, then compare to the Inquirer's article a couple of days later. ;)
Then there is this: http://www.beyond3d.com/forum/viewtopic.php?t=19891 which no one ever commented on at the time. May be no relation, or may be power-saving related, or hell, I dunno. . .
http://www.beyond3d.com/forum/viewtopic.php?p=452121
Yeah, thanks. I just edited my original on reading all the way to the bottom and came to the same conclusion as that thread. Ah well, another cool theory slain --the asynch part, that is, which Dave didn't like in the first place. :cry:
Heh, ATi was supposed to offer stochastic AA way back with the 8500, weren't they? IIRC, Anand's 8500 p/review should mention stochastic. (I guess the F-buffer got filed into that same to-do folder. :P)
I also thought nV only did 2x AA per cycle (so 4x required two goes), while ATi could do up to 6x AA in a single cycle. Wrong again?
Chalnoth
24-Mar-2005, 01:32
I also thought nV only did 2x AA per cycle (so 4x required two goes), while ATi could do up to 6x AA in a single cycle. Wrong again?
All multisampling hardware out there today only does 2x AA per cycle.
DegustatoR
24-Mar-2005, 02:26
Even if those 3tmu number is true we don't know what kind of capability one of those tmu has
What kind of capability can it has? To do 3 point samples, 2 bi and one tri? :)
Besides r580 being the first (ati)part to do fp filtering it may also be the first (ati)part supporting bilinear filtering on vertex textures?
Are we counting VS3 TMUs in this number too? Anyway, extending TMUs abilities (adding FP filtering and filtering in VST, new forms and levels of aniso maybe?) can hardly be described as x3 compared to previous generation. That just doesn't make any sense.
DegustatoR
24-Mar-2005, 02:29
I also thought nV only did 2x AA per cycle (so 4x required two goes), while ATi could do up to 6x AA in a single cycle. Wrong again?
Even if ATI could do it (and they can't) memory bandwidth wouldn't allow them to do it.
Kaleidoscopes usually turn in a circle, don't they? :P Wavey once dropped that code-name out of the blue into a subject on R520 memory usage.
Hell, I don't care if it's just you and me on this one. Brilliant!
Technical, I'm not. I do have some experience in collating tidbits. Why, on this very thread (second post down): http://www.beyond3d.com/forum/viewtopic.php?t=21341&postorder=asc&start=273
So, I don't think "its just you and me on this one". :wink:
Simon F
24-Mar-2005, 08:42
What IS "stochastic AA"?
Search for "Distributed Raytracing" by Cook, Porter and Carpenter (of LucasFilm (Pixar?)). That's the paper that really kicked it all off back in 1984.
Thanks tEd!
I did try googling it, but got a bit too much info...."being able to do fully random sample patterns (jittering) per pixel", sums it up very nicely and in a way I can understand.
Actually, you don't want fully random sample locations. You want to use a minimum distance Poisson Disk sampling pattern. Google time again?
That's the wrong tree to bark up :P
That's the wrong tree to bark up :P
Don't spoil it, I was enjoying the tangent that they were going off at :P
Kaleidoscopes usually turn in a circle, don't they? :P Wavey once dropped that code-name out of the blue into a subject on R520 memory usage.
Hell, I don't care if it's just you and me on this one. Brilliant!
I was thinking this isn't just that a kaleidoscope turns in a circle, but all those pretty coloured fragments turn with it too.
Technical, I'm not. I do have some experience in collating tidbits. Why, on this very thread (second post down): http://www.beyond3d.com/forum/viewtopic.php?t=21341&postorder=asc&start=273
So, I don't think "its just you and me on this one". :wink:
Yeah but Baumann doesn't count. He's the source of the agony.
Anyway, hats off to you for your tidbit collating.
Jawed
Ailuros
24-Mar-2005, 09:51
We just reached that "one level down in detail below where I am actually comfortable that I'm not being a total fool", so just pat me on the head and smile if this is too godawful. But I was under the impression that currently they can do single-cycle for 2xaa and loop back once or twice more to do 4x and 6x. Is that right? What would it take transistor-wise, or more units-wise, to make 4x single-cycle?
Basically, you'd add in z-test units. Don't know how many more transistors this would require, but I doubt it would be that useful.
That is to say, with longer shaders, you won't even notice the fact that it's taking a few cycles to output a pixel. Since longer shaders are only going to get more common, this isn't a huge concern.
Depends what the target is after all. If a designer would want usable 8x sparse MSAA (across the board), then looping 4x times 2xAA samples isn't exactly ideal, especially if heavy stenciling comes into play.
R4xx seems to lose =/>70% with 3 loops (6xAA) + dynamic shadows.
If on the other hand it is being considered that a 4x sample density is adequate, there's of course no real need to increase the amount of samples per cycle. I'd rather speculate that we might remain at 2x samples per cycle for the foreseeable future and it might change some time in the more distant future when more "exotic" algorithms might make their appearance.
I wish there would be a way to get "fill-rate free" Supersampling. I tried in lower resolutions 16xSSAA + AF on the GF and my jaws dropped from the texture quality...errrr just don't ask about performance :roll:
silence
24-Mar-2005, 10:03
its fun to read 17 pages of "infomania" to get no info at all..... :lol:
Unknown Soldier
24-Mar-2005, 10:46
Hey .. we all suckers for punishment :D
US
How about spilitting the fragment pipes into two sets that work on the same two quads and do higher FSAA modes the E&S way? Not sure what this requires in terms of additional hardware though. Sure the performance hit would be massive, but if they're re-using R3x0 and want to wet a few pants...
digitalwanderer
24-Mar-2005, 13:27
That's the wrong tree to bark up :P
It don't make the barking any less fun! :P
I still don't see how adding branching to R3x0 would be an easy task. especially since they already had branching capable hardware.
And Huddy mentioned they'll have "fast branching hardware"
I still don't see how adding branching to R3x0 would be an easy task. especially since they already had branching capable hardware.
And Huddy mentioned they'll have "fast branching hardware"
That's a good point. It would seem a sure bet they spent considerable more trannies on that in R520 than NV did in NV40. Is that a caching thing primarily?
DegustatoR
24-Mar-2005, 14:29
If R520 will have 300 millions transistors with 16 pipes and fast PS3 branching and NV-whatever's-next-gen will have 300 millions transistor with 32 pipes and slow PS3 branching what chip do you think will be faster overall? :)
PatrickL
24-Mar-2005, 14:50
Depends, you are talking about real chips or wet dreams ? :twisted:
If R520 will have 300 millions transistors with 16 pipes and fast PS3 branching and NV-whatever's-next-gen will have 300 millions transistor with 32 pipes and slow PS3 branching what chip do you think will be faster overall? :)
Nvidia in the short term due to few titles having PS3 branching. And I doubt ATI would spend that much effort on something without immediate effect (see their reasoning for not including PS3 at all on the R4XX line...unless they lied for their reasoning :twisted: )
Honestly though, I doubt they'd focus solely on fast branching.
If R520 will have 300 millions transistors with 16 pipes and fast PS3 branching and NV-whatever's-next-gen will have 300 millions transistor with 32 pipes and slow PS3 branching what chip do you think will be faster overall? :)
The comparison was to NV40. I would be very surprised if NV didn't beef up their branching as well next-gen.
Plus, I'll kick off a mantra sure to be seen many times in months to come "You can't really compare NVs 90nm part to ATI's R520 because NV had an extra six months --different generations, really; let's wait to see what Rxxx looks like."
That milestone is way early, but wotthehell. :wink:
digitalwanderer
24-Mar-2005, 15:52
Depends, you are talking about real chips or wet dreams ? :twisted:
Wet dreams, we're still speculating pretty heavily here. ;)
kemosabe
24-Mar-2005, 16:35
If R520 will have 300 millions transistors with 16 pipes and fast PS3 branching and NV-whatever's-next-gen will have 300 millions transistor with 32 pipes and slow PS3 branching what chip do you think will be faster overall? :)
That's more like a soaked dream if you're thinking of all those trannies and 32 pipes on 0.11u. :lol: That's far more likely a 2006 objective for NVDA.
ATI's Q2 conference call just ended a couple of hours ago. They didn't shed as much light on R520 as some analysts were hoping, but Orton did mention that TSMC's 0.09u yields and capacity are looking pretty good and that the R520 generation will begin ramping in the next quarter. He also confirmed that their multi-VPU technology is on its way. So this spring/summer the pressure shifts back to NVDA again. Don't you love seesaws? :twisted:
DegustatoR
24-Mar-2005, 17:35
That's more like a soaked dream if you're thinking of all those trannies and 32 pipes on 0.11u. :lol: That's far more likely a 2006 objective for NVDA.
32 pipe NV40 is surely possible on 90nm. NV'll have "a product" on 90nm in 2005 ;) 2006 will bring us WGF2-graphics which will have other targets to meet besides being just "faster" than previous SM3 generation.
but Orton did mention that TSMC's 0.09u yields and capacity are looking pretty good and that the R520 generation will begin ramping in the next quarter.
Well, considering that ALL NV's NV4x chips will soon be done at TSMC (NV44, NV43, NV42 and NV48) i really don't understand why do you guys are so sure that ATI's 90nm part will have a big lead over NV's 90nm part? ;)
Anyway, i believe that we shouldn't expect anything like "free PS3 branching" from R520. It well may be more efficient than NV4x's, but it surely won't be free.
So the question of 16 pipes with 300 transistors remains -- i just don't believe that ATI's gonna do such complex pipelines in R520 which is based on relatively simple pipelines of R300. 24 or 32 pipes is more believable.
Well, considering that ALL NV's NV4x chips will soon be done at TSMC (NV44, NV43, NV42 and NV48) i really don't understand why do you guys are so sure that ATI's 90nm part will have a big lead over NV's 90nm part? ;)
I suppose that depends on whether that 110nm NV refresh shows up one of these days soonish. Do you expect NV to put that out and then follow with a 90nm topper in a month or two?
kemosabe
24-Mar-2005, 18:12
Well, considering that ALL NV's NV4x chips will soon be done at TSMC (NV44, NV43, NV42 and NV48) i really don't understand why do you guys are so sure that ATI's 90nm part will have a big lead over NV's 90nm part? ;)
I was rather talking about a 90nm part vs. a 110nm part for the current upgrade cycle this spring. Not to mention that many people here are fond of reminding us that ATI has yet to incorporate FP32 and SM3.0 into any of its high-end designs (which is of course true). It also bears mentioning that NVDA has yet to successfully design an ASIC on the low-k process (the de facto 0.09u standard). Last official word from Jen-Hsun was that it was "dangerous". :wink:
So, I don't think "its just you and me on this one". :wink:
Yeah but Baumann doesn't count. He's the source of the agony.
Heh. I suspect Wavey suffers too in a myriad of ways. I doubt he got a 4" binder with low-level schematics and 200,000 word narrative of R520 in it 6 months ago. That he's had a few more hints than us is not quite the same as having it laid out. He said as much upstream. That would be worse for me, cause then I really would be grinding at "unlocking the mysteries" and feeling thick if I couldn't or turned out wrong. Then, of course, there is the frustration of "No, you fools, that's not what I meant!" and "Huh, they got it and *still* don't believe me!"
Evildeus
25-Mar-2005, 07:42
IMHO, a R520 with 320m transistors should have the same size as a R420 and a G70 (or whatever is the name) on 0.11 with 300m should have the same size as a NV40. Ati should still have an advantage in this matter.
But well, don't forget that the passage from FP24 to FP32 should take +33% transistors according Ati...
DegustatoR
25-Mar-2005, 10:34
I suppose that depends on whether that 110nm NV refresh shows up one of these days soonish. Do you expect NV to put that out and then follow with a 90nm topper in a month or two?
Not sure about a month or two, but i'm not sure that we can expect R520 availability before autumn either. I'm thinking that NV's 90nm part will be announced sometime around September with real availability closer to holyday season. Until then it's gonna be 110nm part vs R520.
I was rather talking about a 90nm part vs. a 110nm part for the current upgrade cycle this spring.
I'm not sure that R520 will be available this spring. Summer -- maybe, but Autumn looks more realistic.
It also bears mentioning that NVDA has yet to successfully design an ASIC on the low-k process (the de facto 0.09u standard). Last official word from Jen-Hsun was that it was "dangerous".
I sure hope you don't think that NV will be so scared of low-k that it'll stop it's die shrinks on 110nm :D
I was rather talking about a 90nm part vs. a 110nm part for the current upgrade cycle this spring.
I'm not sure that R520 will be available this spring. Summer -- maybe, but Autumn looks more realistic.
By my read, what Orton told the financial analysts yesterday adds up to preview/release in 2nd 1/2 of April or early/mid-May and shipping in volume during June. Now, he didn't come right out and say that, but that's the way it added up to me when I look at what he did say and the financial predictions they made. I think it would also add up that they'll have more than one sku available in that time frame, as they told the analysts to expect their third quarter (March thru May) to be up a little. . .and fourth quarter (June thru August) *to double*. You can't get a doubling done in a just few days, so I'm thinking that as of yesterday at least they are thinking they are going to have all or almost all of that quarter to sell those new parts in volume.
And that's really "the late end" anaylsis. Obviously we're nearly end of March already. I suspect (certainly don't know, however) that the somewhat up third quarter (March-May) isn't about R520-class, but rather about getting their mid-range and AGP house in order.
kemosabe
25-Mar-2005, 17:13
geo, just to clarify. Orton said that overall Q3 revenues would be slightly down relative to Q2. And he was specifically referring to the consumer segment (cellphones, DTV) when he mentioned the doubling of revenues in Q4, although he also expects the PC segment to perform "better than seasonal" for that quarter. In any case, he has already stated that "significant revenues" from the R5XX generation would be a reality by the summer, so I am with you in thinking that by the time the fall rolls around, it's going to be a top-to-bottom R5XX family. Everyone seems content to conclude that availability hiccups in one generation must necessarily continue into the next. I think that's a perception Orton is in fact counting on.
Thanks for the clarification. Yeah, I think we're on the same page.
Most of the analyst reports on the CC that I've read seem to think that ATI was being conservative, a lot of them expect ATI to beat their own forecasts.
Most of the analyst reports on the CC that I've read seem to think that ATI was being conservative, a lot of them expect ATI to beat their own forecasts.
Are you trying to tell us you think we're leaning a bit late? :) I'm being conservative too.
Given that the next financial analysts call would be late June, and what he told them about the fourth quarter (June-Aug), I think that would be a very unpleasant call if he can't tell them that they are shipping in volume when that call takes place. So, absent a last minute setback, I think worst-case at least some of our diligent visit-50-websites-if-necessary-cost-is-no-object friends will be smirking at the rest of us in June sometime.
kemosabe
25-Mar-2005, 18:08
Most of the analyst reports on the CC that I've read seem to think that ATI was being conservative, a lot of them expect ATI to beat their own forecasts.
Yes, this is the popular line of thinking for some analysts like Jonathan Hykawy who follows the company closely. He thinks that R430 will bring home the bacon in the current quarter until R520 starts shipping. I'm still not convinced that X300/X600 are safe in those OEM slots going into the spring refresh. Not sure whether it will be a typical refresh considering that most OEMs have only been shipping their PCI-E systems for a couple of months, but I expect a bit more market share loss to 6200/6600 in both desktop and mobile before the new architecture takes over.
Despite two analysts asking about R520, Orton said nothing other than the yields being to their satisfaction. He sounds determined to minimize the leaking. :(
Despite two analysts asking about R520, Orton said nothing other than the yields being to their satisfaction. He sounds determined to minimize the leaking. :(
Well, he did say "broad range of 90nm products".
And this:
“Right in front of us is the emergence of a new class of visual processing platforms. Over the next few months we expect to launch our next generation of gpu’s based on the new 90nm technology. Stay tuned and watch for the next major leap in visual realism come to reality.”
Geo: “Next major leap of visual realism” sounds kind of hyperbolic –or does he know something we don’t? That doesn’t sound like performance –that sounds qualitative, and in the absence of an api change. Hrmph! Is that a "SM3.0 didn't count until *we* did it" kind of observation, typical CEO-geewhiz-boilerplate, or something else?
Hellbinder
25-Mar-2005, 20:44
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
Actually, I was *hoping* it meant that. :D Now tell us what it is, HB. :wink:
kemosabe
25-Mar-2005, 21:22
And you would be very wrong.
:twisted:
caboosemoose
25-Mar-2005, 23:52
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
No it doesn't. It's just the usual guff about a new product. I'm not criticising Orton for saying that kind of thing - that's his job, but I have to say i think it's ludicrous to infer anything whatsoever about the R5xx series from his statements.
DegustatoR
26-Mar-2005, 00:24
No it doesn't. It's just the usual guff about a new product. I'm not criticising Orton for saying that kind of thing - that's his job, but I have to say i think it's ludicrous to infer anything whatsoever about the R5xx series from his statements.
Shoo! You're spoiling it!! :twisted:
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
No it doesn't. It's just the usual guff about a new product. I'm not criticising Orton for saying that kind of thing - that's his job, but I have to say i think it's ludicrous to infer anything whatsoever about the R5xx series from his statements.
Ah, crap --then it's not 90nm!? :shock:
:wink:
And you would be very wrong.
:twisted:
Ah yes! A timely reminder of that unseemly outburst of raw emotion. :lol: Not to worry, I don't get emotionally invested in such hopes. A little sumtim' along those lines would be nice --after all, they spent those transistors somewhere-- but I'm not counting on a world-shaker.
Ailuros
26-Mar-2005, 09:42
This is of course a rather weird example, but it might pose food for thought:
http://www.xbitlabs.com/images/video/2004-27gpu2/Games/flatout_1600_candy.gif
image taken from:
http://www.xbitlabs.com/articles/video/display/2004-27gpu2_27.html
I can see a X850XT PE with 8.6 GPixels/sec and 37.7 GB/sec and a GF6800U@SLI with 2*6.4 = 12.8 GPixels/sec and 35.2 GB/sec per GPU having a ~22% difference. The application isn't of course something special shader-wise and there's always the PCI-E bandwidth in the way between boards in a SLi setup, yet even with those aspects in mind a 50% higher theoretical fill-rate gains less than half of that more performance.
I'd expect that on any upcoming high end board wherever f.e. shader or AF density isn't high, the performance advantage to be quite lower than the difference in theoretical raw fill-rate.
Possibly it's just me, but when I look at a graph like that I find myself thinking it is the end of history for performance at those IQ settings. In other words, I say to myself we don't more performance so much as we need new settings to spend that hypothetical new performance on. But then, maybe that's just me.
wireframe
26-Mar-2005, 16:34
Possibly it's just me, but when I look at a graph like that I find myself thinking it is the end of history for performance at those IQ settings. In other words, I say to myself we don't more performance so much as we need new settings to spend that hypothetical new performance on. But then, maybe that's just me.
I hope it's just you (being crazy and all... :P). I read many posts about how powerful the current top of the line cards are and the arguments about how at that level you really shouldn't discriminate too much between them in terms of fram rate and instead focus on things like quality. I, for one (I hope I am not alone about this one now...), find that even though the current generation is twice as fast as the previous, and this is supposed to be some shock or treat, they are still not powerful enough even for current games.
I become even more confused when I read about people talking about CPU bottlenecking and how far ahead the GPUs are. This is total garbage if you ask me. Sure, you want a fast CPU to avoid those dips where the CPU is the bottleneck and when you benchmark you don't want those lows, but if by upping the resolution and AA you find your frame rate plummeting then you certainly can use faster video hardware.
In Far Cry, for example, the shader stuff seems to be the breeze for something like a 6800 Ultra. It's when you have tons of foliage that things get rough. It may be a combination of things, but this seems to suggest that the very basics of rendering performance are still lacking.
I am one of those people who hasn't been visually impressed by the latest games. Far Cry and Painkiller did something for me, but this came down to Far Cry being huge and Painkiller being artistic. Doom 3 and Half Life 2 are just "meh" in my opinion, with too much 'legacy' in them to qualify as a true next-gen the way we want and expected them to be. I also think that shader usage in todays titles are a puny fraction of what they could, should, and will be. But that's probably just me ;)
I suppose I could summarize my view like this: if the 6800 Ultra had been twice as powerful across the board then we would have had reason for pause and to pick our jaws up from the floor like we have those times when real jumps in 3D rendering happend, whether they were in terms of hardware or software.
Chalnoth
26-Mar-2005, 16:57
I suppose I could summarize my view like this: if the 6800 Ultra had been twice as powerful across the board then we would have had reason for pause and to pick our jaws up from the floor like we have those times when real jumps in 3D rendering happend, whether they were in terms of hardware or software.
Well, the 6800 Ultra already was essentially double the performance of everything else out there when it was announced. What more do you want?
tahrikmili
26-Mar-2005, 23:14
Well, the 6800 Ultra already was essentially double the performance of everything else out there when it was announced. What more do you want?
The jump R300 made over R200/NV25
digitalwanderer
26-Mar-2005, 23:26
I finally got some definate and hard info on the R520!
It's gonna be red, count on it. 8)
DegustatoR
26-Mar-2005, 23:41
Well, the 6800 Ultra already was essentially double the performance of everything else out there when it was announced. What more do you want?
The jump R300 made over R200/NV25
I tend to think that the jump from NV35 to NV40 SLI was FAR bigger than the one you're talking about.
tahrikmili
26-Mar-2005, 23:55
Well, the 6800 Ultra already was essentially double the performance of everything else out there when it was announced. What more do you want?
The jump R300 made over R200/NV25
I tend to think that the jump from NV35 to NV40 SLI was FAR bigger than the one you're talking about.
NV35 was a joke, let's leave that POS card out of the discussion. The jump from R360 (which was around at the time of NV35) to NV40 SLI wasn't as big as it was from NV25 to R300. The difference in IQ and improvement in performance was simply astounding. NV40 offers nothing in IQ over R360.
DegustatoR
27-Mar-2005, 00:11
NV35 was a joke, let's leave that POS card out of the discussion.
Nope, it wasn't. And i see no reason to leave it out of any discussion.
The jump from R360 (which was around at the time of NV35) to NV40 SLI wasn't as big as it was from NV25 to R300.
Did you check the benchies? May i suggest i'd go and do it?
The difference in IQ and improvement in performance was simply astounding. NV40 offers nothing in IQ over R360.
NV25 to R300 -- there was nothing "astounding" in quality difference with the possible exeption of MSAA.
NV40 offers FP32, SM3 and 8xS (which is better FSAA mode quality-wise than any mode available on Radeons).
Anyway, that's offtopic and begin to sound religious. I won't continue this talk.
Ailuros
27-Mar-2005, 00:13
Possibly it's just me, but when I look at a graph like that I find myself thinking it is the end of history for performance at those IQ settings. In other words, I say to myself we don't more performance so much as we need new settings to spend that hypothetical new performance on. But then, maybe that's just me.
I warned that the graph is about a relatively simplistic simulation game; future more shader intensive games (than the current crop of games like Hl2, Far Cry etc.) will most likely be not as "generous" when it comes to performance.
IMHO games are currently in a transitional period of hybrid dx7/9.0 features. Once real dx9.0 appear in the future I believe the demands in fill-rate will be unbelievably high.
The jump from R360 (which was around at the time of NV35) to NV40 SLI wasn't as big as it was from NV25 to R300.
How much faster was the R300 over the NV25 in the end? I recall an ~up to 250% proportion and I'm sure it wouldn't be too hard to find cases where even the 6800U will be by the same degree (if not more) faster than a R360, let alone two NV40's in SLi; always of course depending on the application and which IHV everyone wants to favour in the end.
The difference in IQ and improvement in performance was simply astounding. NV40 offers nothing in IQ over R360.
What does the R4xx offer in contradiction compared to R360 then? Something below "nothing" maybe?
The will be also IQ improvements for SM3.0 and float HDR capable cards in the foreseeable future; even more once ATI's R5xx family of desktop VPUs arrives.
tahrikmili
27-Mar-2005, 00:14
NV40 offers FP32, SM3 and 8xS (which is better FSAA mode quality-wise than any mode available on Radeons).
There is yet to be observed a single occasion where FP32 and SM3 can make any IQ improvements that SM2 can not handle, and the 8xSSAA mode simply crawls it's comparable to NV3x's FP32. :roll:
Whatever.
tahrikmili
27-Mar-2005, 00:17
How much faster was the R300 over the NV25 in the end? I recall an ~up to 250% proportion and I'm sure it wouldn't be too hard to find cases where even the 6800U will be by the same degree (if not more) faster than a R360, let alone two NV40's in SLi; always of course depending on the application and which IHV everyone wants to favour in the end.
To quote your graph:
http://www.xbitlabs.com/images/video/2004-27gpu2/Games/flatout_1600_candy.gif
75fps vs 29fps -> ~250%
Heh.. My question remains - does the NV4x actually bring with it the huge IQ improvements R300 brought with MSAA & a 256-bit memory bus? If we can enjoy high levels of anti-aliasing today, it's because of the R300 architecture. And what exactly has the NV40 architecture contributed to gaming IQ-wise, so far? A big zero. FP32 and SM3.0 paths that actually do jack..
What does the R4xx offer in contradiction compared to R360 then? Something below "nothing" maybe?
I never argued the R4xx line offered anything significant over the R3xx line. This generation compared to the last generation brought with it very little new usable features and certainly no observable IQ improvements, whereas the R200/NV2x -> R300 transition was a huge leap in BOTH performance and IQ.
Well, the 6800 Ultra already was essentially double the performance of everything else out there when it was announced. What more do you want?
The jump R300 made over R200/NV25It was about the same, wasn't it? Double the performance of the previous high end.
I'm not sure SLI is an appropriate comparison for single cards (especially WRT cost), but SLI is attractice in its ability to break through the current manufacturing limits (next-gen performance this gen). IIRC, though, SLI wasn't debuted at the same time as NV40, so it's still more appropriate to compare it to a single NV40, IMO.
wireframe
27-Mar-2005, 00:24
In cosidering the jump to NV40 (or the non-jump as some may seem to claim) you may want to remember the frame buffer blending capabilities of the NV40. Without reiterating the "HDR can be done through shaders" let's keep in mind that this is a real improvement. In fact, this is what I had in mind when I made my "I want more" post. Forget Far Cry in 1600x1200 4xFSAA 8xAF. Think about it in terms of "r_HDDRendering 1" (that's the command to activate HDR rendering in Far Cry, for those not familiar). This has a huge impact on performance. You cannot use FSAA in this mode, but even if you could it would be terribly slow. Let's say a NV40 gives you a comfortable Far Cry experience in HDR at 1024x768 with AF.
Another thing to take away from this is that R420 could be 4 times the frequency that it is and it still does not have this very important capability. It will be a non-player in titles like that (Spliter Cell: Chaos Theory, for example). What's the point of just pushing more frames in "old" games? Once real shader heavy games hit and use HDR (and don't think for a second that Half Life 2 is a shader intensive game) it will be pitiful. You will need a NV40 SLI to get the effect and an R420 won't even run it.
This is why it is important that R520 delivers major performance, so that this revolution can happen. Sure, it probably won't 'just happen', but this is the thing that needs to happen.
EDIT:
Hopefully Orton's comments about increased visual realism is exactly about this. SM 3.0 for sure (which implies 32FP) and HDR. They already have the 3Dc... heh.
EDIT 2:
And stop with this "double this generation". It's not about running Quake 3 at twice the frame rate. Run Quake 4 with HDR and more shaders than you can count and then we have something to benchmark. Double the frame rate ceases to be important at certain critical inflection points. It's a whole new game.
wireframe
27-Mar-2005, 02:12
I just did something that was both fun and horrible at the same time. I booted up Half-Life 2 and forced it into DirectX 8.1 mode (mat_dxlevel 81). I had never seen this before, having been running this title only on a Geforce 6800 Ultra using the default dxlevel 90. My goodness, where to begin?
It looks exactly the same! I don't see why people are even making an issue and benchmarking bonanza out of this, showing horrible numbers for the FX 5900 in DX 9.0. Who cares, if it looks the same running DX 8.1? That is assuming that an FX 5900 can run this game comfortably using this 8.1 mode, of course. Well, I have an idea why this is made into an issue, but that would require us all to sit down in a circle wearing tinfoil hats and chanting some pretty peculiar stuff.
If you didn't know what to specifically look for in HL2/CS:Source between DX81 and DX90, would you be able to tell which is which without flipping back and forth? Try it. Just be careful, I am not asking if you can see a difference, some speculars will give that away, the question is if you can see the "dx9 high fidelity" in there. I don't feel the difference even knowing there is one and knowing the mode it is running in.
The point here is that Half-Life 2, as an example, was and is hyped up to be some sort of revolutonalry "DirectX 9 high-fidelity game." Why, then, does it look the same when running under DirectX 8.1? The answer is obvious. It is not a next-gen title in anything but marketing and some clever cvars. Now, turn on HDR in Far Cry and I am pretty sure you will get that feeling that you are seeing something new. Something good. Something that would be considered worthy of being called "the next level." Now, it's important to realize that Far Cry is not fully equipped to run in HDR mode, some levels show it better than others. It is not really intended as a play mode, but it shall have to suffice for 3D addicted enthusiasts such as ourselves. It shows great promise.
Disclaimer: this is not to put down HL2 as a game or subliminally elevate Doom 3...it all stinks blah blah blah.... ;)
It looks exactly the same!
I disagree (http://www.beyond3d.com/forum/viewtopic.php?t=16232). I was more than a little surprised at how poor the DX8 path looked, as if it had been artificially crippled.
In cosidering the jump to NV40 (or the non-jump as some may seem to claim) you may want to remember the frame buffer blending capabilities of the NV40. Without reiterating the "HDR can be done through shaders" let's keep in mind that this is a real improvement. In fact, this is what I had in mind when I made my "I want more" post. Forget Far Cry in 1600x1200 4xFSAA 8xAF. Think about it in terms of "r_HDDRendering 1" (that's the command to activate HDR rendering in Far Cry, for those not familiar). This has a huge impact on performance. You cannot use FSAA in this mode, but even if you could it would be terribly slow. Let's say a NV40 gives you a comfortable Far Cry experience in HDR at 1024x768 with AF.
Another thing to take away from this is that R420 could be 4 times the frequency that it is and it still does not have this very important capability. It will be a non-player in titles like that (Spliter Cell: Chaos Theory, for example). What's the point of just pushing more frames in "old" games? Once real shader heavy games hit and use HDR (and don't think for a second that Half Life 2 is a shader intensive game) it will be pitiful. You will need a NV40 SLI to get the effect and an R420 won't even run it.
This is why it is important that R520 delivers major performance, so that this revolution can happen. Sure, it probably won't 'just happen', but this is the thing that needs to happen.
EDIT:
Hopefully Orton's comments about increased visual realism is exactly about this. SM 3.0 for sure (which implies 32FP) and HDR. They already have the 3Dc... heh.
EDIT 2:
And stop with this "double this generation". It's not about running Quake 3 at twice the frame rate. Run Quake 4 with HDR and more shaders than you can count and then we have something to benchmark. Double the frame rate ceases to be important at certain critical inflection points. It's a whole new game.
thank you
In cosidering the jump to NV40 (or the non-jump as some may seem to claim) you may want to remember the frame buffer blending capabilities of the NV40. Without reiterating the "HDR can be done through shaders" let's keep in mind that this is a real improvement. In fact, this is what I had in mind when I made my "I want more" post. Forget Far Cry in 1600x1200 4xFSAA 8xAF. Think about it in terms of "r_HDDRendering 1" (that's the command to activate HDR rendering in Far Cry, for those not familiar). This has a huge impact on performance. You cannot use FSAA in this mode, but even if you could it would be terribly slow. Let's say a NV40 gives you a comfortable Far Cry experience in HDR at 1024x768 with AF.
For me IQ improving features are only worth it, if they don't require me to turn off AA, cause I absolutely hate jaggies. So I'm waiting for frame buffer blending plus AA with usable frame rates. I hope I don't have to wait until WGF 2 for that!
Anyway, we're quite a bit OT here. Any news about R520? :D
Subtlesnake
27-Mar-2005, 10:08
The point here is that Half-Life 2, as an example, was and is hyped up to be some sort of revolutonalry "DirectX 9 high-fidelity game." Why, then, does it look the same when running under DirectX 8.1? The answer is obvious. It is not a next-gen title in anything but marketing and some clever cvars. Now, turn on HDR in Far Cry and I am pretty sure you will get that feeling that you are seeing something new. Something good. Something that would be considered worthy of being called "the next level." Now, it's important to realize that Far Cry is not fully equipped to run in HDR mode, some levels show it better than others. It is not really intended as a play mode, but it shall have to suffice for 3D addicted enthusiasts such as ourselves. It shows great promise.
HDR is going to be added with an upcoming level release by Valve. I guess that will cause it to suddenly become a next-gen title :)
Ailuros
27-Mar-2005, 10:11
To quote your graph:
75fps vs 29fps -> ~250%
wireframe is actually right in that this whole thing is ridiculous. If I'd handpick though a couple of applications it would look like that:
http://www.xbitlabs.com/images/video/2004-27gpu2/Games/hellhole_1600_candy.gif
Heh.. My question remains - does the NV4x actually bring with it the huge IQ improvements R300 brought with MSAA & a 256-bit memory bus? If we can enjoy high levels of anti-aliasing today, it's because of the R300 architecture. And what exactly has the NV40 architecture contributed to gaming IQ-wise, so far? A big zero. FP32 and SM3.0 paths that actually do jack..
As I said anything concerning float HDR or SM3.0 will show up to a bigger degree in upcoming games, once both vendors support those functionalities. Actually from my side I'm wishing even better usability on ATI's upcoming R520, like for instance the ability to combine multisampling with float HDR.
Albeit of admittedly low to very low usability the NV4x line offers the hybrid 8xS (4*8 EER) and 16x (8*8 EER) modes. In an older game that's loaded with alpha tests, an image quality conscious user will most certainly value those modes. Here the level of antialiasing is actually higher than the competition; whereby of course the lower degree of usability is a valid negative side.
If NVIDIA deserves any criticism for it's AA algorithm then it's the lack of programmability and hardware based gamma correction functionalities. In fact minus the minor changes since the NV2x line, the AA algorithm is basically the same since then.
I never argued the R4xx line offered anything significant over the R3xx line. This generation compared to the last generation brought with it very little new usable features and certainly no observable IQ improvements, whereas the R200/NV2x -> R300 transition was a huge leap in BOTH performance and IQ.
Vendors do in my book the best they can for a specific timeframe. NV took that leap from the NV1x-->NV2x transition in the past, while switching to multisampling. Albeit I'll be the first to get enthusiastic about significant antialiasing algorithmic improvements, I somewhat have the feeling that ATI will rest for the time being even in the R5xx line on those R3xx laurels for that department.
Personally I expect significant changes for either/or or ideally both with the first crop of WGF2.0 class accelerators.
Ailuros
27-Mar-2005, 10:19
Considering Far Cry's float HDR implementation I've rather mixed feelings about it. While there are spots that it looks great, there are also instances that it looks rather bad and that probably because it looks more like an afterthought than anything else.
Despite that, I still think that float HDR if properly implemented can deliver a significant IQ improvement; future accelerators though need a lot more than NV40 for that department.
The point here is that Half-Life 2, as an example, was and is hyped up to be some sort of revolutonalry "DirectX 9 high-fidelity game." Why, then, does it look the same when running under DirectX 8.1? The answer is obvious. It is not a next-gen title in anything but marketing and some clever cvars. Now, turn on HDR in Far Cry and I am pretty sure you will get that feeling that you are seeing something new. Something good. Something that would be considered worthy of being called "the next level." Now, it's important to realize that Far Cry is not fully equipped to run in HDR mode, some levels show it better than others. It is not really intended as a play mode, but it shall have to suffice for 3D addicted enthusiasts such as ourselves. It shows great promise.
HDR is going to be added with an upcoming level release by Valve. I guess that will cause it to suddenly become a next-gen title :)
its going to be funny seeing how much better bloom everywhere makes half life 2 look
HDR is going to be added with an upcoming level release by Valve. I guess that will cause it to suddenly become a next-gen title :)
its going to be funny seeing how much better bloom everywhere makes half life 2 look
:roll:
caboosemoose
27-Mar-2005, 15:56
I just did something that was both fun and horrible at the same time. I booted up Half-Life 2 and forced it into DirectX 8.1 mode (mat_dxlevel 81). I had never seen this before, having been running this title only on a Geforce 6800 Ultra using the default dxlevel 90. My goodness, where to begin?
It looks exactly the same! I don't see why people are even making an issue and benchmarking bonanza out of this, showing horrible numbers for the FX 5900 in DX 9.0. Who cares, if it looks the same running DX 8.1? That is assuming that an FX 5900 can run this game comfortably using this 8.1 mode, of course. Well, I have an idea why this is made into an issue, but that would require us all to sit down in a circle wearing tinfoil hats and chanting some pretty peculiar stuff.
If you didn't know what to specifically look for in HL2/CS:Source between DX81 and DX90, would you be able to tell which is which without flipping back and forth? Try it. Just be careful, I am not asking if you can see a difference, some speculars will give that away, the question is if you can see the "dx9 high fidelity" in there. I don't feel the difference even knowing there is one and knowing the mode it is running in.
The point here is that Half-Life 2, as an example, was and is hyped up to be some sort of revolutonalry "DirectX 9 high-fidelity game." Why, then, does it look the same when running under DirectX 8.1? The answer is obvious. It is not a next-gen title in anything but marketing and some clever cvars. Now, turn on HDR in Far Cry and I am pretty sure you will get that feeling that you are seeing something new. Something good. Something that would be considered worthy of being called "the next level." Now, it's important to realize that Far Cry is not fully equipped to run in HDR mode, some levels show it better than others. It is not really intended as a play mode, but it shall have to suffice for 3D addicted enthusiasts such as ourselves. It shows great promise.
Disclaimer: this is not to put down HL2 as a game or subliminally elevate Doom 3...it all stinks blah blah blah.... ;)
You obviously aren't terribly discerning or haven't been looking properly - the difference between some of the water shaders, dx9 and dx8.1, in HL2 is night and day. Granted some of the other shaders are essentially identical.
digitalwanderer
27-Mar-2005, 16:04
Did I mention that I talked to Patti and got OFFICIAL CONFIRMATION that it will be red? :|
(Ok, actually I didn't as I haven't even LOOKED at me e-mails inbox for the last week and a half....I get that way sometimes. *HRM* )
Dave Baumann
27-Mar-2005, 17:02
Did I tell you that the board will look remarkably like a board not from ATI?
digitalwanderer
27-Mar-2005, 17:11
Did I tell you that the board will look remarkably like a board not from ATI?
Really? You mean I really should find out what color it is?
I'm on it, thanks. 8)
Did I tell you that the board will look remarkably like a board not from ATI?
Milestone check: "Credible source makes comment that implies he's actually seen one of these critters, or at least a picture of one."
One of these days I really am going to do that "milestone list". It'd be fun.
Chalnoth
27-Mar-2005, 18:29
Did I tell you that the board will look remarkably like a board not from ATI?
Dave, you're hilarious.
Did I tell you that the board will look remarkably like a board not from ATI?
And is your analysis that the reason for this is aesthetic, functional, or both?
There were a couple of R520 boards at CeBIT. I would imagine Dave saw one and noted the obvious similarity to a Diamond Virge 3D.
Dave, did you notice a King & Queen motif on the back of the PCB? What's that all about - supertiling?
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible. I am not sure about the reports of samples reaching 700MHz core - there was a badly photoshop'd pic floating about that's reduced the credibility of this rumour. They will push 650MHz+ after a metal-only respin which should be back soon (if it isn't already).
G70/G80 are coming in late Q3/Q4 earliest so in the meantime nV may offer dual NV48s - not 100% on that designation, but apparently whatever it is has ~10% more transistors than NV40, is based on 0.11u and capable of 500-550MHz. I don't quite understand how they're going to pitch it.
2nd-hand info so add salt liberally. On the other hand, I don't think it's a waste of time to try and completely figure out what those 16-1-1-1/16-1-3-1 descriptions mean... ;)
digitalwanderer
27-Mar-2005, 20:35
There were a couple of R520 boards at CeBIT. I would imagine Dave saw one and noted the obvious similarity to a Diamond Virge 3D.
Dave, did you notice a King & Queen motif on the back of the PCB? What's that all about - supertiling?
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible. I am not sure about the reports of samples reaching 700MHz core - there was a badly photoshop'd pic floating about that's reduced the credibility of this rumour. They will push 650MHz+ after a metal-only respin which should be back soon (if it isn't already).
G70/G80 are coming in late Q3/Q4 earliest so in the meantime nV may offer dual NV48s - not 100% on that designation, but apparently whatever it is has ~10% more transistors than NV40, is based on 0.11u and capable of 500-550MHz. I don't quite understand how they're going to pitch it.
2nd-hand info so add salt liberally. On the other hand, I don't think it's a waste of time to try and completely figure out what those 16-1-1-1/16-1-3-1 descriptions mean... ;)
Thank you MuFu!!!! :shock:
Tweaker
27-Mar-2005, 20:40
There were a couple of R520 boards at CeBIT. I would imagine Dave saw one and noted the obvious similarity to a Diamond Virge 3D.
what similarity do you mean, IIRC there were some expansion slots on Diamond Virge
digitalwanderer
27-Mar-2005, 20:41
http://www2b.biglobe.ne.jp/~tajiri/dorcom/win2k/Stealth_3D.jpg
:|
Chalnoth
27-Mar-2005, 22:46
what similarity do you mean, IIRC there were some expansion slots on Diamond Virge
I seem to remember that, around that time, you could typically elect to purchase a video card with or without memory expansion slots. I'm 99.9% certain this was the case with the Trio64 (I believe this was the S3 card immediately preceding the Virge), but not completely certain where the Virge stood.
Did I mention that I talked to Patti and got OFFICIAL CONFIRMATION that it will be red? :|
(Ok, actually I didn't as I haven't even LOOKED at me e-mails inbox for the last week and a half....I get that way sometimes. *HRM* )
Don't you love being ignored?
Tweaker
27-Mar-2005, 23:14
I believe this was the S3 card immediately preceding the Virge
Yes, that's right.
Btw, it was only joke. Someone deleted MuFu's latest post. Probably MuFu himself, damn it.. :D
digitalwanderer
27-Mar-2005, 23:52
Don't you love being ignored?
Only by my family and friends, when it happens with strangers I get annoyed. :?
Oh, Jaysus, I just had post traumatic shock flashbacks of adding memory chips individually (and bending the f***ing pins!). But I think that was one generation earlier on a Trident VLB.
Oh, Jaysus, I just had post traumatic shock flashbacks of adding memory chips individually (and bending the f***ing pins!). But I think that was one generation earlier on a Trident VLB.
I wish I had photographic evidence from when me and dad doubled the video ram on our first computer, from 1K to 2K. SRAM chips piggy-backed and soldered leg to leg, bent legs (in different planes), patch wires everywhere, cut tracks...
Ah, the early 80s...
Jawed
kemosabe
28-Mar-2005, 00:59
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible. I am not sure about the reports of samples reaching 700MHz core - there was a badly photoshop'd pic floating about that's reduced the credibility of this rumour. They will push 650MHz+ after a metal-only respin which should be back soon (if it isn't already).
650MHz+? Is that all? What do they pay you the big bucks around here for? :wink:
BTW, I don't recall seeing that pic. Are the photoshop jobs under NDA too? :lol:
Sorry, did I say MHz? I meant GHz.
digitalwanderer
28-Mar-2005, 01:09
It's the 1.6Ghz memory that is worrying me a bit.
I'm sure they can do it, but can they supply it. :?
Ailuros
28-Mar-2005, 01:36
It's the 1.6Ghz memory that is worrying me a bit.
I'm sure they can do it, but can they supply it. :?
Obviously not with the first samples; in any case that's a milestone both IHVs will face while setting new fill-rate records.
There were a couple of R520 boards at CeBIT. I would imagine Dave saw one and noted the obvious similarity to a Diamond Virge 3D.
what similarity do you mean, IIRC there were some expansion slots on Diamond Virge
Hmm! So would one be justified in reading those tea leaves as a bigger connector, and thus greater card-to-card bandwidth, for ATI MVP than NV SLI?
martrox
28-Mar-2005, 01:53
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible. I am not sure about the reports of samples reaching 700MHz core - there was a badly photoshop'd pic floating about that's reduced the credibility of this rumour. They will push 650MHz+ after a metal-only respin which should be back soon (if it isn't already).
Inquirer here we come.......... :roll:
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible.
Memory clock, or memory *bus* clock? :D
Nah, that's gotta just be my pet theory and I don't know dick.
The prophecy has been fulfilled with a perfect accuracy.
20 Pages based solely on non substantial infos.
Nevertheless, some tidbits, being either a joke or some speculative materials, will make the L'Inq headlines for the next week.
Amen.
:P
It's all in this 20 pages I tells thee. And MuFu's latest mutterings can only have come from one or two sources :lol: At least I now know to trust most of what he says.
It'll be nice to revisit this thread in a few months and see what was wrong and what wasn't. Regardless, Fudo would be mad not to read what's in here...
21 pages of 'info' and i walk away knowing absolutly nothing(facts) about r520 other than it will support sm 3.0 :x
Tim Murray
28-Mar-2005, 03:36
The prophecy has been fulfilled with a perfect accuracy.
20 Pages based solely on non substantial infos.
Nevertheless, some tidbits, being either a joke or some speculative materials, will make the L'Inq headlines for the next week.
Amen.
:P
I'm still thinking it'll be 1.3 times your prophecy.
21 pages of 'info' and i walk away knowing absolutly nothing(facts) about r520 other than it will support sm 3.0 :x
Hahah! Sounds like this is definitely a normal web-forum! ;)
The truly important thing we've established in this thread is that MuFu has been "busy getting a life" and is not, repeat not strapping on his Electro-Boy suit in secret to sneak around the cyber byways sniffing out tasty tidbits. Nosirree. Way to go on that Real Life thing, MuFu! We're all jealous. :lol:
Well, and grateful too. Don't stop. :wink:
kemosabe
28-Mar-2005, 06:45
Yup, his R520 life is off to a late but glorious start. Can't wait till he sets his sights on R600 and really gets his life in gear. :lol:
silence
28-Mar-2005, 10:02
could r520 look like nv30? 8) ..... if its coming out late summer, it will be available just in time to do some cleaning in my backyard....those damned leaves... :wink:
Heard some mutterings today. R520 memory clock is freaky high - 1.6GHz+ possible.
Come to think of it, could this be a forward-looking refresh thing? I'm remembering the contretemps over R300 as to what it shipped with vs what it would support. . .
digitalwanderer
28-Mar-2005, 16:56
Wait a minute, when the hell did MuFu get a life and is that allowed?
I thought we had rules against that here... :?
21 pages of 'info' and i walk away knowing absolutly nothing(facts) about r520 other than it will support sm 3.0 :x
Oh come one. We know it will be "fast" compared to whats out today. And we know it may or may not be red and it may not look like other ATI boards. See oddles of info :)
digitalwanderer
28-Mar-2005, 18:31
Is anyone else kind of disapointed that none of this has made Le Inq yet? :(
21 pages of 'info' and i walk away knowing absolutly nothing(facts) about r520 other than it will support sm 3.0 :x
Oh come one. We know it will be "fast" compared to whats out today. And we know it may or may not be red and it may not look like other ATI boards. See oddles of info :)
Besides, the subject title doesn't promise "info", it promises "infomania", and there has been plenty of that! The difference between fire and firefly. . .
. . .and that it has a connector that is implied to be significantly bigger than NV's SLI connector. . .
digitalwanderer
28-Mar-2005, 18:57
And we know it may or may not be red
Actually, I DO know the answer to that one now....I'm just waiting on permission to print from me ATi insider before I can go public with the BIG NEWS. 8)
trinibwoy
28-Mar-2005, 19:29
Is anyone else kind of disapointed that none of this has made Le Inq yet? :(
I've actually never seen anything on the Inq that was taken from B3D. Guess I haven't been around long enough?
21 pages of 'info' and i walk away knowing absolutly nothing(facts) about r520 other than it will support sm 3.0 :x
Oh come one. We know it will be "fast" compared to whats out today. And we know it may or may not be red and it may not look like other ATI boards. See oddles of info :)
20% faster? 50%? thasts what i want to know so i can hurry up and upgrade.
digitalwanderer
28-Mar-2005, 21:15
Hot off the frontpage of Elite Bastards (http://www.elitebastards.com/#news9685) we bring you official news of the R520:
I HAVE OFFICIAL WORD ON THE R520!!!!! :shock:
This just in:
I am happy to confirm that the card will be red....now which exact tone of red I can't say, that is a highly guarded secret....
Developing... :twisted:
Official confirmation. 8)
digitalwanderer
28-Mar-2005, 21:54
Wtg, "Scoop". :)
They don't call me "Colour Commentator" for nothing ya know. 8)
IMHO games are currently in a transitional period of hybrid dx7/9.0 features. Once real dx9.0 appear in the future I believe the demands in fill-rate will be unbelievably high.
Winding back a hair for a tangent --can someone give me or point me at an explanation for why the bolded would be true? I thot I more or less understood what 9.0 brings to the table, but I'm not making the leap to "unbelievably high" fill-rate demands as a consequence. Actually, I understood the opposite to be the case --that as we become shader-throuput-bound that fillrate & bandwidth are less likely to be the bottleneck for awhile. Isn't that one of the basis for Wavey's "16 pipes" theories?
Fill rate should be less of an issue as more shader intensive games come out, A perfect example is Half-Life2 I can crank the AA and resolution and not loose that many frames. In Doom3 which isn't as shader intensive but more fill rate bound AA takes a larger hit.
wireframe
29-Mar-2005, 00:21
Fill rate should be less of an issue as more shader intensive games come out, A perfect example is Half-Life2 I can crank the AA and resolution and not loose that many frames. In Doom3 which isn't as shader intensive but more fill rate bound AA takes a larger hit.
But can you prove that this is a shader limitation and not just a CPU/programming bottleneck?
Fill rate should be less of an issue as more shader intensive games come out, A perfect example is Half-Life2 I can crank the AA and resolution and not loose that many frames. In Doom3 which isn't as shader intensive but more fill rate bound AA takes a larger hit.
Well, yes, that's what I thot too. But I respect Ailuros opinion enuf to ask what he meant by that rather than assume he's full of it. :wink:
Would FP blending and HDR rendering require more fill rate or more processing power like shaders do?
Ailuros
29-Mar-2005, 02:02
IMHO games are currently in a transitional period of hybrid dx7/9.0 features. Once real dx9.0 appear in the future I believe the demands in fill-rate will be unbelievably high.
Winding back a hair for a tangent --can someone give me or point me at an explanation for why the bolded would be true? I thot I more or less understood what 9.0 brings to the table, but I'm not making the leap to "unbelievably high" fill-rate demands as a consequence. Actually, I understood the opposite to be the case --that as we become shader-throuput-bound that fillrate & bandwidth are less likely to be the bottleneck for awhile. Isn't that one of the basis for Wavey's "16 pipes" theories?
What if those 16 SIMD channels are clocked at 700MHz? That's 11.2 GPixels/sec and almost twice as much as on a 6800U. What about even further in the future with even more quads and similar clockspeeds?
Looking again at my stuff, I probably should have worded it differently and should have used something like "internal shader fill-rate" or something like: how would you think a NV40 will perform in a UE3 engine based game compared to a possible 8 quad monster @ 600MHz for example?
Considering bandwidth on the other hand here's a pic from one of the GPU gems whitepapers to be found on NVIDIA's site:
http://users.otenet.gr/~ailuros/estimate.jpg
Albeit just a prediction, I can see bandwidth scaling almost in parallel with transistor counts for the next decade. Clock speeds are estimated here to see a more modest increase, but then again it doesn't mention how quads are estimated to scale either.
Ailuros
29-Mar-2005, 02:18
Fill rate should be less of an issue as more shader intensive games come out, A perfect example is Half-Life2 I can crank the AA and resolution and not loose that many frames. In Doom3 which isn't as shader intensive but more fill rate bound AA takes a larger hit.
But can you prove that this is a shader limitation and not just a CPU/programming bottleneck?
HL2 is IMO both GPU and CPU bound (it falls exactly in the hybrid dx7/9.0 category I mentioned before); as for Doom3 the mystery for the performance drops with AA enabled, lies in the MSAA + stencil shadows combination (I'd rather say that Doom3 is both CPU - think of how exactly it renders those dynamic shadows - and stencil fill-rate bound).
kemosabe
29-Mar-2005, 02:39
Some hardware sites (http://www.babelfish.altavista.com/babelfish/trurl_pagecontent?lp=de_en&trurl=http%3a%2f%2fwww. 3dcenter.de%2fartikel%2f2005%2f03-21_a.php) are sticking with the 6-quad R520 scenario, but I don't quite get their theory that RV515/530 might be fabbed on anything other than 0.09u.
And they would be very wrong.
:lol:
Possibly I'm getting light-headed, but I think I begin to understand. Usually when talking to someone like Ailuros I have that feeling like I had in Scotland that the other fellow is talking my language and I *ought* to be able to make sense of it, so why can't I --quite?
Tho that Gems graph can be assumed to be NV's view of the world and it is an open question if we are on the cusp of another serious divergence between the two IHV's directions. One might even wonder if NV is really "behind" on 90nm for any other reason than to better match up with their perceived bandwidth needs.
It does make sense to me from a transistor-stingy pov to beef up the PS's rather than rely on even more parallelism to take care of it; at least until such time as the ps shaders aren't the limiting factor on how many pixels you can push thru the pipes.
I reread Wavey's original X800 review to help with this, which probably equally explains the light-headedness and the glimmer of understanding. :lol:
Unknown Soldier
29-Mar-2005, 07:20
Wtg, "Scoop". :)
They don't call me "Colour Commentator" for nothing ya know. 8)
:lol:
US
Unknown Soldier
29-Mar-2005, 07:25
The R520 will be 24 pipes but act like a 32 pipe sometimes(ala NV6800)?
24x1
32x0
US --> Oh and Enough already ATI .. release the damn card!
US --> Oh and Enough already ATI .. release the damn card!
Heheh. . . That comment brings up an interesting question:
Would you prefer an paper launch with a much later hardware launch if it meant satisfying our curiosity earlier? ;)
The R520 will be 24 pipes but act like a 32 pipe sometimes(ala NV6800)?
24x1
32x0
That'd mean that 8 of thoses pipelines are different from the 16 others. That's wouldn't make sense.
Would you prefer an paper launch with a much later hardware launch if it meant satisfying our curiosity earlier? :wink:
Personaly, I'd do. :D
Unknown Soldier
29-Mar-2005, 08:37
Heheh. . . That comment brings up an interesting question:
Would you prefer an paper launch with a much later hardware launch if it meant satisfying our curiosity earlier? ;)
But it taped-out last year in November already .. How long must it take before they release the damn thing already!
Personally I think they waiting to sell-out their old stock(X800's) before bring in the new gun .. but If they gonna bring it at a higher cost prices than $500(The price the X800 was going for) then they might as well release it now anyways. If they gonna release it at the current price range of the X800 then they can hold onto it a bit longer.
US --> When is the NDA over??
Ailuros
29-Mar-2005, 09:54
The R520 will be 24 pipes but act like a 32 pipe sometimes(ala NV6800)?
24x1
32x0
US --> Oh and Enough already ATI .. release the damn card!
No idea if it will have 6 quads or not, but that theoretical 32 zixel maximum is already there on R4xx according to one of the whitepapers. NV40 has 16 "double"-ROPs and I doubt that R520 will have more; I can't figure what we'd need more for anyway.
geo,
Another attempt:
R360 = 8*412 = 3296 MPixels/sec
R420 = 16*520 = 8320 MPixels/sec
How high is the difference in performance between those two and especially in shader and/or AF intense scenarios?
I personally expect that trend to continue for the future and I don't think the estimates in that former graph are exaggerated either. 1.9GHz give or take aren't that much for a decade later, nor 10x times the transistor count of a GPU today.
DegustatoR
29-Mar-2005, 12:31
The R520 will be 24 pipes but act like a 32 pipe sometimes(ala NV6800)?
24x1
32x0
That'd mean that 8 of thoses pipelines are different from the 16 others. That's wouldn't make sense.
Actually it would... If those 8 are hybrid PS/VS pipelines from R500 :)
Skinner
29-Mar-2005, 17:08
some XDR goodness:
http://www.devhardware.com/c/a/Memory/Hardware-news-XDR/
scificube
29-Mar-2005, 19:14
Thanks for the article Skinner :)
XDR has always looked good to me...but I have to wonder if Rambus has learned that if they want to be successful you have to sell your products at a reasonable price.
You can't sell at a premium when you're trying to move someone else out of the way. You sell your big stuff at a premium when you're in charge of the show.
As a gamer...I want it so I hope they get it straight.
---------------------------
A bit more back on topic...is the consensus still at R520 will support 3.0 for sure and that's it?
It seems 1.3x the performance of the R420's pipelines on a per pipe comparison of those in the R520 is still in the air...
24 pipes is a bit more solid but still squishy...
A mixture of old pipes and new pipes is pretty shaky...
and then there is the idea performance will not double but the capabilities will increase which seems kinda weird to me unless the new capabilities are only checklist features (SM3.0 comes to mind)...
I just want to be sure I know absolutely nothing about the R520 now :)
Even it's R300->R420->R520 lineage seems to be in question now :?
I hope I'm not alone in feeling lost.
digitalwanderer
29-Mar-2005, 19:17
I hope I'm not alone in feeling lost.
Nah, I feel that way all the time here.
The trick is not to let it get in the way of having some fun. ;)
Well, I'm feeling better about a release date and location being set at least. This time one that doesn't require a rip in the space-time continuum (my deleted post above). I think I'm buying into Ratchet's hint upstream that I was previously leaning late in my estimation. I don't think we're going to make it into May without (p)reviews.
silence
29-Mar-2005, 19:46
23 pages and counting....anyone care to make summery of what we know? :wink:
digitalwanderer
29-Mar-2005, 19:54
23 pages and counting....anyone care to make summery of what we know? :wink:
Sure!
The R520 will be red, but the exact shade of red is still unknown. 8)
silence
29-Mar-2005, 20:02
:shock:
not even shades? :lol:
Dave Baumann
29-Mar-2005, 20:14
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
digitalwanderer
29-Mar-2005, 20:25
(think about what you do with a Kalideoscope aprt of the actual physical action)
Got it, Kalideoscope=tiling....you don't even gotta beat a thicky like me with a hammer to catch that one. :roll:
;)
BTW-The shade of red is very important, and I expect going to surprise a whole bunch of people! :shock:
(think about what you do with a Kalideoscope aprt of the actual physical action)
Got it, Kalideoscope=tiling....you don't even gotta beat a thicky like me with a hammer to catch that one. :roll:
E tu, Digi? :cry:
digitalwanderer
29-Mar-2005, 20:40
E tu, Digi? :cry:
What? What don't ya get? :|
ATI were not willing to publicly comment on their expected forthcoming release of their next generation R520 core despite there being plenty of interest in it. One thing that did transpire during discussions with various ATI employees was that they have taken to heart some of issues that have been perceived in bringing boards to market, to the point where they mentioned in a recent investor conference that their CEO has been spending much time addressing the issues internally in order to be able to turn around new products faster, and as such there may be the possibility that they are not willing to announce R520 until they are much closer to being in volume. This could have some impacts on when we’ll hear announcements from ATI on R520.
Hmm. Looking for a GF4 kinda release?
Well, a kaleidoscope mixes up coloured fragments, randomising them. Also it only samples an angle, which means that only a subset of the coloured fragments are being viewed. The tricky bit has gotta be the 6x (8x, 12x, 16x?) mirroring and re-mirroring of the sampled fragments.
Sounds like anti-aliasing, :lol:
Jawed
Joe DeFuria
29-Mar-2005, 21:06
I'm sticking with my "Kalideoscope is a rebranding / updating of their multimonitor display" technology theory. ;)
I'm sticking with my "Kalideoscope is a rebranding / updating of their multimonitor display" technology theory. ;)
Well, they'd certainly have a marketing job to do with me if they called an anti-aliasing scheme "Kaleidoscope". My immediate mental picture of the results was Picasso paintings.
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
Clarifications, schmarifications.. :p
I think the term "release date" has been a bit of a misnomer in this industry, lately. I think we have observed the phenomena "announcements" and "availability" more often. With reference to Dave's CeBit-article, I therefore predict that the R520 derived PC-part will be announced "When it's done (and sampling in considerable numbers)."
To use a Kaleidoscope(****), you put a rotatable tube with coloured bits of paper and a couple of mirrors to your eye, point the end (which is often made of transluscent plastic) towards a lightsource and turn the barrel of the tube to se wonderfully psychedelic random patterns of colour. The non-physical bit would be "seeing wonderful patterns of colour", and this would make me think the codename would represent something to do with 10bit per channel colour output or something to do with multiple displays (the mirror thing).
Psychedelic colours were also a part of the visual style in the Austin Powers movie from which the term "Mojo" was taken, weren't they? :roll:
Anyways - am I atleast partially wrong? Answer please.. ;)
Kind regards
Kjetil
****
There seems to be two different methods of spelling floating around, "..leido.." and "..lideo..". Mr Baumann prefers "..lideo..", while google prefers "..leido.." with 1700000 over 525, and indeed the original cover inlay for my copy of Srgt. Pepper's heart's Club Band states that the girl in Lucy in the Sky with Diamonds has "Kaleidoscope eyes." This means nothing else than that I'd never miss the opportunity to pick on the spelling of a professional journalist.
Eureka!
http://www.googlism.com/index.htm?ism=kaleidoscope&type=1
I like number 2.
(Sorry for the slight off-topic, I think I spent too much time in the sun today)
Kind regards
Kjetil
CyFactor
29-Mar-2005, 21:46
Anyways - am I atleast partially wrong? Answer please.. ;)
Kind regards
Kjetil
Yes, you are at least partially wrong. :D
Anyways - am I atleast partially wrong? Answer please.. ;)
Kind regards
Kjetil
Yes, you are at least partially wrong. :D
Thank you for the quick and precise reply ;)
K
Ailuros
29-Mar-2005, 22:15
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
If that should be related to the rumoured 300M (I haven't followed every bit of the relevant discussions), I suspect there might be some confusion for others between the console and desktop part. Could be completely wrong though.
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
You create colored patterns, there are a rumor that the R520 will support more than 8 bit per component output (10 bit).
Edit: sorry should have read the rest of the thread before replying, I was not the first with this idea.
R300King!
30-Mar-2005, 01:35
Kalideoscope? Hmmm ..the R520 has 1 pipe and its made to look like there's 32! There's my guess. ..oh and the psychedelic color effect gets you high too! Man, this card is gonna fly off the shelves. Way to go ATI. :lol:
Inane_Dork
30-Mar-2005, 02:22
I bet it's because ATi has a super-secret collision detection library that works exclusively on GPU. Collidoscope.
kemosabe
30-Mar-2005, 03:34
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
Did you just say as little as I thought you said? :lol:
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
Did you just say as little as I thought you said? :lol:
Yeah, well, he did tell us what Kaleidoscope isn't (tho that was a small minority on that one :D ), that he hasn't publicly bought in to the 300-350M number (I don't know that anyone thot he did --we were just trying to reconcile with that number compared to other factors he does seem to be pointing at) and he implied the announce/preview could come well into May and still have the full impact on ATI's fourth quarter that Orton appeared to promise.
Indeed, at the extreme end, arguably if they pull a GF4 release the article that started this thread with the prediction of Computex in Taiwan would not be out of the question from a date perspective.
So, y'know, quite a lot actually by Wavey's standards when he's not in "play with the mouse" mode. :wink:
Personally, I'm not yet ready to give up on an earlier announce date (April), tho we clearly aren't within just a couple weeks (say April 10 or earlier) or you'd *really* expect to have seen certain other Signs and Portents that have not yet appeared. Like, for instance, bombastic annoying teasers on the main page at ati.com.
Tim Murray
30-Mar-2005, 05:37
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
it's AA, isn't it. this means I'm going to have to buy one, doesn't it. I am saddened.
Having dried my eyes over the messy demise of the elegant circle solution, and noting that two out of three of Dave's corrections were polite ways of saying "Geo is full of ***t", I shall reenter the fray on Kaleidoscope. Kaleidoscopes are all about the fragments. Look up the definitions. They blew their transistor budget on a new kick-ass ps core and decided it needed an especial name.
Hellbinder
30-Mar-2005, 06:23
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
No it doesn't. It's just the usual guff about a new product. I'm not criticising Orton for saying that kind of thing - that's his job, but I have to say i think it's ludicrous to infer anything whatsoever about the R5xx series from his statements.
No actually you are wrong.
Khronus
30-Mar-2005, 06:41
I'm staking my claim on Kaleidoscope being their Multi VPU architecture. On a side note, what're the chances of seeing HDMI outputs on the R520 generation? R520 AIW w/ Theater 200 and it's internal audio routing perhaps?
Reverend
30-Mar-2005, 06:42
no it means that the R5XX series has more in it to increase "visual realism" than simply a boost to SM3.
That's what every single video chip since the age of 3D proposes to do ("increase visual realism").
Of course, the R5xx could be a "quantum leap" in this aspect, like what you're alluding to.
I think I speak for all of us here when I say I'd be disappointed if the R520 was just a small leap. :P
Unknown Soldier
30-Mar-2005, 06:48
Just a couple of clarifications: Some people appear to think I've staked a claim as to the number of transistors in R520... I haven't. Also Kalideoscope has nothing to do with the memory bus (think about what you do with a Kalideoscope aprt of the actual physical action). As for the release date, read my CeBit report!
ATI were not willing to publicly comment on their expected forthcoming release of their next generation R520 core despite there being plenty of interest in it. One thing that did transpire during discussions with various ATI employees was that they have taken to heart some of issues that have been perceived in bringing boards to market, to the point where they mentioned in a recent investor conference that their CEO has been spending much time addressing the issues internally in order to be able to turn around new products faster, and as such there may be the possibility that they are not willing to announce R520 until they are much closer to being in volume. This could have some impacts on when we’ll hear announcements from ATI on R520.
That Cebit Report??
Come ATI .. it's APRIL already .. release the monster. :D
US
chavvdarrr
30-Mar-2005, 06:53
:? :oops:
Reverend
30-Mar-2005, 07:20
I think I speak for all of us here when I say I'd be disappointed if the R520 was just a small leap. :P
I'd be disappointed too if NV50 was the same.
Tim Murray
30-Mar-2005, 07:22
http://www.techpowerup.com/reviews/Cebit2005/Day4/5
missed the first post, eh?
im getting rdy to see a lot of disappointed people.
im getting rdy to see a lot of disappointed people.
I have a feeling that will be the case but I hope not.
Reverend
30-Mar-2005, 08:22
im getting rdy to see a lot of disappointed people.
That's a subjective topic. And probably one that wouldn't matter a great deal to the IHVs and software developers.
All developers I know are happy with both NVIDIA and ATI, for the now and for the near future.
"People" in your context mainly (I think) consists of those that want faster fps with AA and AF. They wouldn't recognize real innovation even if they experienced it (and that has nothing to do with AA and AF).
But that's natural.
[edit] In my next RATP post/thread, I will tell you what developers want. Which means what "people" can expect.
PatrickL
30-Mar-2005, 08:45
I doubt about an April launch as ATI.com shows no event planned before May.
im getting rdy to see a lot of disappointed people.
Always the optimist, eh? :P
Personally I don't really care how many factors of x the R520 is faster than the R420, the only way I'd be disappointed is if it didn't have full SM3.0 support (as useless as some people think it is)...
...and maybe some response to SLI, but I think that's more a platform thing and not exclusive R520 territory.
We'll see when the time comes.
Those funny boys at Inq weigh-in:
http://www.theinquirer.net/?article=22202
Of course, how they managed sm3.0 support would be a bit problematical in this scenario. . . 160M x 2 without it. . .:D
Love_In_Rio
30-Mar-2005, 10:58
Those funny boys at Inq weigh-in:
http://www.theinquirer.net/?article=22202
Of course, how they managed sm3.0 support would be a bit problematical in this scenario. . . 160M x 2 without it. . .
No if they are unified pipes...
Blazkowicz
30-Mar-2005, 11:37
Those funny boys at Inq weigh-in:
http://www.theinquirer.net/?article=22202
Of course, how they managed sm3.0 support would be a bit problematical in this scenario. . . 160M x 2 without it. . .:D
maybe they are making up crap from the "24 pipelines that behave like 32", that maybe comes from the "pipe with 1.3x more power" crap, and they assume it will be a 32 pipes part with 24 enabled for a start.
That hardly makes any sense.
BTW I think these "exxtreme 1.3 pipelines" comes from the fact there would be 48 ALUs in the R500 executing up to 64 threads. R500!=R520 of course, but that doesn't stop bad speculations from propagating :)
32 pipes SM3 is still theorically possible, doubling the pixel pipes does not mean doubling the transistor count (look at R420 vs R360)
but still I guess it will have 16 pipes, and if I'm wrong, then 24. (isn't there a point where a huge number of pipes is pointless for small polygons or something?)
Simon F
30-Mar-2005, 12:15
Those funny boys at Inq weigh-in:
http://www.theinquirer.net/?article=22202
Assuming, and this is a BIG assumption, that the 'funny boys' are right and the figure of 17x17 mm is correct, what sort of yield would this have? 1 chip in 10?
chavvdarrr
30-Mar-2005, 12:24
If we are talking about more than 300 million transistors in the chip that is 17 square mm small,
I assume that is typo, 17 sq mm is 4x4 mm :)
maybe 217 or 317 sqmm ? :)
StellaArtois
30-Mar-2005, 12:39
17X17mm I assume he means... = 289mm squared.
DegustatoR
30-Mar-2005, 12:51
NV40/45 has 287 sq mm chip so why are you so afraid of 289? :)
Those funny boys at Inq weigh-in:
http://www.theinquirer.net/?article=22202
*shrug*
Assuming 1.6GHz memory (which sounds high anyway) and 650MHz core, that would mean more than 3x less raw bandwidth relative to the fillrate than a 9700 Pro. Just to put it further in perspective, we thought the 9800XT was bandwidth-starved... a 32-pipe R520 at only 500MHz would require 3.6GHz memory to suffer the same level of depravity.
Of course there's more to this than raw numbers... but not *THAT* much more. I'd like to be suprised, however. :lol:
Plus, we had the 12/16 pipeline nonsense before and look how that turned out. If it has 32 pipes, there will be an SKU with all of them enabled. None of this "a quad in hand" bollocks.
Joe DeFuria
30-Mar-2005, 14:13
Yeah, it doesn't really make sense to me to have that much pixel writing power given bandwidth constraints.
Having "more shader computation ability per pixel per clock" makes much more sense to me.
32 pipes SM3 is still theorically possible, doubling the pixel pipes does not mean doubling the transistor count (look at R420 vs R360)
This didn't come out quite as bad as I thought it would. 160-107=53m, 53m/8*16=106m 160m+106m = 266M Which at 350m still gives you 84m for SM3.0 which seems tightish for 32 pipes but maybe not impossible.
I still don't believe it, however.
Yeah, it doesn't really make sense to me to have that much pixel writing power given bandwidth constraints.
Having "more shader computation ability per pixel per clock" makes much more sense to me.
What, you don't want to be the first kid on your block to run Q3 at over 1,000 fps at 640x480? :lol:
NV40/45 has 287 sq mm chip so why are you so afraid of 289? :)
Well, there is this: http://www.beyond3d.com/forum/viewtopic.php?t=17663
And you might remember pre-R420 that ATI indicated that they were temporarily willing to take a hit on yields with a big'un, but they didn't expect to keep doing it. Of course, most of that was before they got a look at NV40. :)
Assuming 1.6GHz memory (which sounds high anyway) and 650MHz core, that would mean more than 3x less raw bandwidth relative to the fillrate than a 9700 Pro. Just to put it further in perspective, we thought the 9800XT was bandwidth-starved... a 32-pipe R520 at only 500MHz would require 3.6GHz memory to suffer the same level of depravity.
At 256-bit. :) Hey, "he who says A must say B". But I ain't buying yet.
R520 is 90nm remember (or that's what's believed anyway)
Kanyamagufa
30-Mar-2005, 18:55
Regardless of whether or not the rumors are true from our friends at the Inq, it would be pretty freakin' sweet to see 32 pipelines in this baby. It would just be incredible.
This is probably just a stupid idea, but assuming that everything is 32 bit, it might be more interesting to make multiple 128-bit channels to the memory, like 3 or so. Better sustained bandwith and less cascading latencies? And less traces than a single 512-bit bus.
Edit: or how about using a DDR3 interface and using some kind of switch to overlay the signals of 2 DDR2 chips? Thereby doubing the bandwith.
Edit2: or some extra on-package extra memory, as with the Mobility chips? You would only have a (slower?) 256-bit external memory interface.
kemosabe
31-Mar-2005, 01:50
R520 is 90nm remember (or that's what's believed anyway)
That is no longer in question as Dave Orton has publicly removed all doubt.
Kaleidoscopes are all about the fragments. Look up the definitions. They blew their transistor budget on a new kick-ass ps core and decided it needed an especial name.
Come to think of it, if you're going to do the above it would make sense to do it one level up as a direct replacement for SMARTSHADER HD.
Why is this still going?
:shock: :)
I remember reading somewhere that "Kaleidoscope" was a motherboard related feature not graphics.
I remember reading somewhere that "Kaleidoscope" was a motherboard related feature not graphics.
Probably what you are referring to is Anand's piece that said it would be part of the RS600 igp/chipset. Which probably means it isn't MVP-related (I started back there when it first got mentioned some months ago :) ). Doesn't rule out AA or multi-monitor tho, I would think. Or my current pet theory.
vBulletin® v3.8.6, Copyright ©2000-2013, Jelsoft Enterprises Ltd.