R520 info thread #(9!)^9 [Archive]

Arun

30-Sep-2005, 19:33

What an astonishingly boring architecture *shrugs*
What saddens me most is that the G80 is going to be just as boring last I heard. Oh well.

Uttar
EDIT: What Fudo says for the VS can also be said, word for word, for the G70 VS. This is surprisingly lacking any details regarding vertex texturing...
I also kind of doubt this diagram is accurate when it comes to the texture cache, but hey, I'd hate to be proved wrong about that one (and nvidiots would love me to be).

dizietsma

30-Sep-2005, 19:41

Fuad manages to slaughter the info coming out of Ibiza ...... he's obviously got some guy taking snaps from his phone and texting him snippets :)

CMAN

30-Sep-2005, 19:48

From a level of 1-10 (with 1 being my mother and 10 being an engineer at ATI), what do you guys seriously think Fuad's understanding level about how graphics cards work is?

Also, I hope they have availability before November for most of their cards. If the X1600's do not come out until after Thanksgiving, then I would suspect their Christmas season will be disappointing.

Geo

30-Sep-2005, 19:52

What an astonishingly boring architecture *shrugs*
What saddens me most is that the G80 is going to be just as boring last I heard. Oh well.

:shock: Compared to this? http://www.beyond3d.com/reviews/ati/r420_x800/index.php?p=5

What is your baseline against which this new one is "boring"?

BRiT

30-Sep-2005, 19:52

From a level of 1-10 (with 1 being my mother and 10 being an engineer at ATI), what do you guys seriously think Fuad's understanding level about how graphics cards work is?

I say -5 (negative five), putting him somewhere around the level of being your grandmother's pet gerbil.

Demirug

30-Sep-2005, 19:53

Uttar, I don't see any texture unit for the vertexshader at all. nVidia show them and ATI know this.

I don't want to call this a boring architecture. At least is it different and this give me something more to write about. G70 is boring as it is nearly the same as NV4X.

Geo

30-Sep-2005, 19:58

Uttar, I don't see any texture unit for the vertexshader at all. nVidia show them and ATI know this.

Hasn't there been a hint or two around here that this is some kind of hack in R5xx that will allow them to claim compatibility without actually providing dedicated hardware to do it? I thot I saw something that sniffed in that direction. . .

Razor1

30-Sep-2005, 20:00

Hasn't there been a hint or two around here that this is some kind of hack in R5xx that will allow them to claim compatibility without actually providing dedicated hardware to do it? I thot I saw something that sniffed in that direction. . .

If the ring bus is what we think it is the data from a pixel shader can go to the vertex shader with out much latency at all.

Arty

30-Sep-2005, 20:00

I don't want to call this a boring architecture. At least is it different and this give me something more to write about. G70 is boring as it is nearly the same as NV4X.
Noooooo ! G70 has the best architecture ever

:lol:

Demirug

30-Sep-2005, 20:04

If the ring bus is what we think it is the data from a pixel shader can go to the vertex shader with out much latency at all.

There are other ways to do vertex textuering without textureunits in the vertex shader. NV3X supports this. Look for something called "render to vertexbuffer"

Razor1

30-Sep-2005, 20:07

There are other ways to do vertex textuering without textureunits in the vertex shader. NV3X supports this. Look for something called "render to vertexbuffer"

ah yes true

Arun

30-Sep-2005, 20:10

Demirug: Yeah, I realized that about the VS. I was just criticizing Fudo's VS writeup, he could have done a better job at it considering the info he had.
Anyhow, I'm not disagreeing. Architecture-wise, it's less boring than G70 was. Problem is, there's nothing that's really interesting in it. Take the NV40 for example, when it was introduced it brought a lot of ideas to the table: partial MIMD in the PS, L1/L2 texture caches, Vertex Shader Threading to optimize Texturing (*cough* *cough*), free Normalize instruction, and tons of other stuff.

I'm not saying it isn't a potentially good architecture. But I think it isn't worth a zillionth of the hype it has gotten. If anything, it feels more like a R400 derivative to me (note: I said R400, not R420, there's a huge difference here). I'm also under the impression the R520 team has a bunch of the ex-R400 engineers, but I'd love a confirmation on that. And no, I don't believe the R400 had true unified shaders ala R500.

Perhaps I was expecting a bit too many interesting things from this part. Blame the hypers. Honestly, all of these features have been discussed on this forum for nearly 3 years now (R400, once again), with the potential exception of the ringbus (seems like a logical extension of Hybrid Texturing though).

The ONLY two nice things in the R520 seem to be more feature-full Vertex Texturing (I kind of wonder whether it'll be faster than G70 when not using filtering, too) and cheaper Pixel Shader dynamic branching. The G70 Vertex Shader's dynamic branching is also threading-optimized.

Then there's FSAA and AF. I can only hope for a nice surprise there, heh. Call me old-fashioned, call me bitchy, but I'm just not surprised by that architecture. Hopefully the featureset and the performance will impress me more...

Uttar

dizietsma

30-Sep-2005, 20:11

Demirug

30-Sep-2005, 20:18

Uttar, yes it looks a little bit like a R400 design. But as it was never used in a chip they sell it will be new for most of my readers.

Vertex textureing. I am suspect that most games will use FP32 textures for this as nVidia support nothing else. In this case a filter is not a option anyway. The main question will be witch buffer they use to feedback the textureresults in the vertexshader. An internal or the local videoram?

The second question that will say use something about the performace is the size of the register array.

Razor1

30-Sep-2005, 20:19

Can someone with a bigger brain than me explain further on the vertex shading issue. Previously the main focus seems to have been on pixel shading power but this seems to indicate they have massively beefed up the vertex shading.

Why ?

This is actaully confusing me too, even 8 vertex shaders today aren't the bottleneck don't see why they would need 2 instructions per cycle for each vertex shader unit.

Sunrise

30-Sep-2005, 20:25

But I think it isn't worth a zillionth of the hype it has gotten.
Perhaps I was expecting a bit too many interesting things from this part. Blame the hypers.

Call me blind, but i couldn´t see any hype surrounding R520, but that obviously differs from person to person. Being influenced by R400-level tech and some new bus-interface which in itself is a welcome addition, it should rather be seen as some kind of a transition architecture in the PC sector. It´s not like we actually break through limits this time, but considering everyone expected R520 to be heavily R300-ish based i think this is quite a positive surprise to some. R520 looks to be very efficient in what it does, at least that´s my understanding of things.

Those missing "details" about the vertex texturing part however...

Xmas

30-Sep-2005, 20:32

This is actaully confusing me too, even 8 vertex shaders today aren't the bottleneck don't see why they would need 2 instructions per cycle for each vertex shader unit.
That's nothing new, it can issue one vector and one scalar instruction per cycle in the VS, like R300.

The best thing about R520 looks just a little weird.

Arun

30-Sep-2005, 20:46

Uttar, yes it looks a little bit like a R400 design. But as it was never used in a chip they sell it will be new for most of my readers.Very true, I'm talking about it from my perspective here, as well as maybe that of some other Beyond3D regulars.

Vertex textureing. I am suspect that most games will use FP32 textures for this as nVidia support nothing else. In this case a filter is not a option anyway. The main question will be witch buffer they use to feedback the textureresults in the vertexshader. An internal or the local videoram?I wouldn't be so hasty about this. Do you know just how big of a performance hit linear filtering alone is on the G70, if you want to use Vertex Texturing? Let alone bilinear and trilinear... It is true that many uses of Vertex Texturing do not need filtering, but quite a bunch do, too, so being able to use those in realworld situations, in real games, and at good performance, would be a great thing.
If developers need linear or bilinear filtering on their (not-necessarly-FP32) Vertex Textures and ATI supports it natively, trust me, they're going to make sure they can use it in their engine. Even more so considering ATI's range is top-to-bottom, and it replaces their old generation chips in all $100+ market segments.

The second question that will say use something about the performace is the size of the register array.That, and the number of threads, and what their definition of "thread" is. I would assume it to be a quad on the R520, and 3 quads on the RV530, but I'm ready to be proved wrong about this.
Compared to the amount of quads being processed at once in the G70, it seems minuscule, so I'm a bit curious about how comparable it really is.

That's nothing new, it can issue one vector and one scalar instruction per cycle in the VS, like R300.Yep. Same for the NV40, too (not the NV30 afaik, but I'd have to check). And NVIDIA's Vertex Shaders have been thread-based ever since the NV30 (although I'm unsure how much really pre-NV40). Your following statement is making me curious though, Xmas :smile:

Call me blind, but i couldn´t see any hype surrounding R520, but that obviously differs from person to person.I'm talking about architectural hype here; and considering the size of those threads on Beyond3D alone, I would tend to believe it was quite there. I don't remember such an insane amount of dicussions about the G70.

Uttar

Geo

30-Sep-2005, 20:52

Call me old-fashioned, call me bitchy, but I'm just not surprised by that architecture. Hopefully the featureset and the performance will impress me more...

Heh. A friend of mine likes to point out that I use the phrase "a certain tension" to talk about two opposing ideas that both have truth to them but are struggling against each other.

So, it seems to me there is a certain tension in why many of us visit B3D. We don't want to be surprised on release day. . . and are disappointed when we aren't.

Demirug

30-Sep-2005, 20:58

I wouldn't be so hasty about this. Do you know just how big of a performance hit linear filtering alone is on the G70, if you want to use Vertex Texturing? Let alone bilinear and trilinear... It is true that many uses of Vertex Texturing do not need filtering, but quite a bunch do, too, so being able to use those in realworld situations, in real games, and at good performance, would be a great thing.
If developers need linear or bilinear filtering on their (not-necessarly-FP32) Vertex Textures and ATI supports it natively, trust me, they're going to make sure they can use it in their engine. Even more so considering ATI's range is top-to-bottom, and it replaces their old generation chips in all $100+ market segments.

Maybe the current DX Spec will make some problem with this. The Problem is that the documentation say not much about texturefilter in the vertex shader. There are more than one was to read it.

That, and the number of threads, and what their definition of "thread" is. I would assume it to be a quad on the R520, and 3 quads on the RV530, but I'm ready to be proved wrong about this.

If you have the number of threads you need the number of pipeline stages and the number of pixel per thread too. Without this information you can not say if it is good or bad.

Compared to the amount of quads being processed at once in the G70, it seems minuscule, so I'm a bit curious about how comparable it really is.

But the G70 have many quads per thread (aka batch). This is the main reason for the bad dynamic branching performances.

Geo

30-Sep-2005, 21:01

I'm talking about architectural hype here; and considering the size of those threads on Beyond3D alone, I would tend to believe it was quite there. I don't remember such an insane amount of dicussions about the G70.

Some of that goes back to the golden glow around R400, for which Orton can be blamed. In retrospect, it seems clear to me now that some folks, including Wavey, were trying to knock that back quite a bit in the Spring of this year. Seeing R520 now, maybe they knocked it back a little too much, but I think this is a matter of degree that reasonable folks can come to different conclusions on.

We had bupkus on G70, not improved by them changing names on us mid-stream. We don't have much architecturally on G75 or G80. Err, anything at all, except for quads going up? Find some red meat to throw to the piranhas on that front and watch the water churn red/white. :razz:

Sunrise

30-Sep-2005, 21:02

I'm talking about architectural hype here; and considering the size of those threads on Beyond3D alone, I would tend to believe it was quite there. I don't remember such an insane amount of dicussions about the G70.

People like to talk about unreleased and basically new stuff and since G70 is NV47, you can take a guess why there wasn´t as much talk about this one, everything about it was also known long before R520 details (the #-#-#-# bit) surfaced. R400 made quite an impression since it was the holy guarded secret ATi had back then, that also influenced R520 talk quite a bit. Orton is the kind of guy who likes to tease. (That famous R600-interview).

Since i´m also getting older, this all is nothing in comparison to the good ol' pre-NV30 days. Don´t you think ? ;)

Arun

30-Sep-2005, 21:11

Maybe the current DX Spec will make some problem with this. The Problem is that the documentation say not much about texturefilter in the vertex shader. There are more than one was to read it.Good point, I only deal with OpenGL personally so I couldn't say :) (the ARB's Vertex Shading extension definitively has no filtering limitations, in case anyone's wondering)
If you have the number of threads you need the number of pipeline stages and the number of pixel per thread too. Without this information you can not say if it is good or bad.I've got a rough idea about the NV40/G70 latency, so I would expect ATI's to be in that ballpark. Once again, that only makes me curious about ATI's definition of a thread (which mostly is about how many pixels are in one, as you say).
But the G70 have many quads per thread (aka batch). This is the main reason for the bad dynamic branching performances.Good point, I tend not to see batches as threads for obvious reasons, but they indeed are the closest things to Pixel Shading threads in the NV4x/G7x.

As for architectural hype differences, I'd say much of it can be attributed to marketing strategies: NVIDIA wanted to surprise the world with a high-quality refresh coming out of no where, available on launch day. On the other hand, ATI has to play more aggressively here.
And yes, good ole "Are You Ready?" - I still got a GFFX T-Shirt from the guy that produced those videos, damnit! ;) Nice guy, too.

Uttar

Hellbinder

30-Sep-2005, 22:46

It doesnt really look that boring to me...

You have a new complex Scheduling unit, a new Memory Controller, Some complete SM3 vertex (and soon to be noted) Pixel capabilities and some new AA.

This is a lot different than just tacking on SM3 to the R300.

I mean.. what would be exciting? We already know that R600 is going to be an enhanced, refigned Pc-ized version of the Xenos..

Really... there is not much left to add to current GPU technology that would be worth throwing a party over.

Geo

30-Sep-2005, 23:23

Errrm. That "compress/decompress" to the Texture Cache on the diagram. . .is that new hardware support for R520? Or just greater level-of-detail than we've seen before? I'm not seeing anything like that on the X800 diagrams. Z-compress is on the old diagrams, however, if represented slightly differently.

rwolf

30-Sep-2005, 23:26

rwolf

30-Sep-2005, 23:26

rwolf

30-Sep-2005, 23:33

http://www.theinquirer.net/?article=26617

A document seen by the INQUIRER claims that ATI will beat Geforce 7800GTX by as much as ninety per cent in the Fear Demo. ATI claims it will end up almost forty per cent faster in Battlefield 2, while it will top Nvidia's fastest card as much as twentyfive per cent in Far Cry.

In certain games, the Radeon X1800XT PE will end up less than twenty per cent faster than GTX. and we are talking about Splinter Cell and the Painkiller extension.

If the R520 has this level of performance it will be truely amazing. The 7800GTX already has very impressive performance.

Geo

30-Sep-2005, 23:40

http://www.theinquirer.net/?article=26617

If the R520 has this level of performance it will be truely amazing. The 7800GTX already has very impressive performance.

Ay yi yi. At what settings I wonder? I can imagine regular Splinter Cell (without HDR) being bandwidth limited with the 20% coming from that. But HDR? And if it isn't HDR that is making those eye-popping numbers on the other titles, what is?

Edit: Hrrm. Shadows?

EditII: I also feel it necessary to note we've now clearly reached the "Wow!" benchmarks stage in the great progression of "milestones" leading to release. Experience would suggest that the "oh. ick." elements are still to come. And somehow I feel confident there will be one or more, tho hopefully not nearly enuf/intensity to outweigh this more happy stage we just hit. :cool:

ERK

01-Oct-2005, 00:39

http://www.theinquirer.net/?article=26617

If the R520 has this level of performance it will be truely amazing. The 7800GTX already has very impressive performance.
If true, then the efficiency of the 7800GTX is lower than I have been imagining. Or perhaps there are still big surprises in the shader pipeline?

Hellbinder

01-Oct-2005, 00:44

I wonder if the R520 will stomp the crap out of the GTX in Doom now.

It looks like they ahve made some huge performance increases in stencil and soft shadowing.

russo121

01-Oct-2005, 00:53

Is it me or the days are getting longer? :) - still 4 days....

Edit: A document seen by the INQUIRER claims that ATI will beat Geforce 7800GTX by as much as ninety per cent in the Fear Demo. ATI claims it will end up almost forty per cent faster in Battlefield 2, while it will top Nvidia's fastest card as much as twentyfive per cent in Far Cry.

In certain games, the Radeon X1800XT PE will end up less than twenty per cent faster than GTX. and we are talking about Splinter Cell and the Painkiller extension.

Do I smell driver cheating coming again (from Nv of course)? I just hope not!

Geo

01-Oct-2005, 01:09

Oh, let's not go there unless/until. Please?

WaltC

01-Oct-2005, 01:52

I say -5 (negative five), putting him somewhere around the level of being your grandmother's pet gerbil.

Well, I'd be somewhat more charitable and put him at around a 1, or about the same level as CMAN's mother...;)

WaltC

01-Oct-2005, 02:05

http://www.theinquirer.net/?article=26617

If the R520 has this level of performance it will be truely amazing. The 7800GTX already has very impressive performance.

In many cases the 7800GTX is only marginally "faster" than an x850-level board from ATi. I looked at one HL2 HDR frame-rate test from AnandTech today in which the 7800GTX was a whopping 4 fps "faster" than my current, slightly overclocked x800 xt (520MHz.) I mean, to me nV's "next generation" has always seemed only slightly faster than the current ATi generation, and I put "faster" in quotes because I'm not at all sure it would be faster at all with the same level of IQ that ATi provides. 7800x vs. x800x comparisons on the Internet have been very short on IQ analysis to date--almost as if reviewers have just gotten too lazy to bother with IQ anymore (to be charitable about it.)

So, if R5x0 actually is a "next-gen" gpu from ATi as opposed to something just a tad faster than R4x0 (a la 7800 vs. 6800 from nV), then it being much faster should surprise nobody, imo. Being much faster/better is something we should expect from a next-gen part, no? No reason to let the 7800 lower those expectations, seems to me...

wireframe

01-Oct-2005, 02:49

In many cases the 7800GTX is only marginally "faster" than an x850-level board from ATi.
Would these cases possibly be ones that are not the most shader demanding and/or at resolutions around 1600x1200? The 7800GTX rips the X850XT PE a new one even under those conditions in games like Battlefield 2. I think that hints at a lot of performance under the hood of the GTX that may be going unused/unnoticed in many cases. Don't forget that CPU limitation will play a larger part until games push more complex shading to stress that part of the system. I would venture a guess that a game like Half Life 2 is a breeze shader-wise for the GTX and it really comes down to 16 v 16 ROPs on the GTX and the X850XT. I think you really have to crank up the resolution and filtering to see the difference in muscle in these "older" games. Upcoming, newer games like FEAR will probably leave the X850XT on the floor while the GTX moves right along.

I looked at one HL2 HDR frame-rate test from AnandTech today in which the 7800GTX was a whopping 4 fps "faster" than my current, slightly overclocked x800 xt (520MHz.)

EDIT: I am assuming you are talking about Day of Defeat: Source here. Your numbers don't jive with Anand's so if he had some other HDR test could you point that out.

Well, it's certainly more than 4 fps faster than the X850XT than Anand used. I see 15 fps faster under Full HDR and that computes to a 27% increase. However, and more importantly, this differential surfaces mostly because the X850XT has dropped speed with HDR enabled. The GTX is still very near the tested limit (CPU?) with no HDR of ~73 fps for the GTX (the X850XT achieved 70.7 fps here).

Calling this "marginal" is very misleading. It is borderline saying that Windows Solitaire doesn't run much faster with a GTX than a previous generation card. Perhaps you should prepare to use the same explanations for when the X1800XT arrives and all the new performance cannot be easily measured and shown due to limitations elsewhere. :wink:

Junkstyle

01-Oct-2005, 03:09

"The top-of-the-line Radeon X1800XT will be available as of the 5th of November"

arrrgh...ANOTHER MONTH TO WAIT for 512MB? Why cant they ship a 512MB XL? If Nvidia ships a 512MB card before Nov 5th and its faster then bye bye ATI.

wireframe

01-Oct-2005, 03:18

If Nvidia ships a 512MB card before Nov 5th and its faster then bye bye ATI.

Are you saying that personally, as in, you will buy a 512MB GTX if it comes out before the 512MB X1800XT, or are you saying that ATI is "dooooooooomed" ( :razz: ) if this happens?

^eMpTy^

01-Oct-2005, 03:22

anyone else wondering why those card pics from hardware analysis only have single-link dvi ports? I thought R520 and derivatives were supposed to be all dual-link?

Skrying

01-Oct-2005, 03:36

anyone else wondering why those card pics from hardware analysis only have single-link dvi ports? I thought R520 and derivatives were supposed to be all dual-link?

Pre-production models? Also, I dont trust anything from Hardware Analysis these days.

wireframe

01-Oct-2005, 03:41

anyone else wondering why those card pics from hardware analysis only have single-link dvi ports? I thought R520 and derivatives were supposed to be all dual-link?

Mind explaining to me how you can tell it's single link from those pictures?

^eMpTy^

01-Oct-2005, 03:45

Mind explaining to me how you can tell it's single link from those pictures?

/me backpedels...

I think I'm confusing the funny connector on the master cards with dual link...my bad...

Sxotty

01-Oct-2005, 04:07

I am just happy that for a launch the reported MSRP is $500 not $600

Junkstyle

01-Oct-2005, 04:25

Are you saying that personally, as in, you will buy a 512MB GTX if it comes out before the 512MB X1800XT, or are you saying that ATI is "dooooooooomed" ( :razz: ) if this happens?

I mean personally I will buy a 7800GTX 512MB if its out way before the 512MB 520 and if its relatively the same speed +/- 10%

wireframe

01-Oct-2005, 04:28

I am just happy that for a launch the reported MSRP is $500 not $600

Don't give them any ideas!

But yeah, tell me about it. Let's just hope that these are the prices we will see in stores. Last round I was initially going to buy an X800XT until the store that carried them asked for $800. I kid you not. I was ripped for $600 for the 6800 Ultra, but that felt more like $200 saved after being asked $800 for that XT..

wireframe

01-Oct-2005, 04:33

I mean personally I will buy a 7800GTX 512MB if its out way before the 512MB 520 and if its relatively the same speed +/- 10%

November 5th is about a month from now. I think it is very unlikely that Nvidia will launch anything (significantly) before then. Why not wait almost three months and have several under the christmas tree, if you're into that sort of thing.

I'm just hoping that these cards will be available to buy and at reasonable prices. The last thing I want now is to go catalog shopping and seeing "out of stock". When the GTX popped out of nowhere it really made it feel like it was forever since I bought a new video card. I guess I want to say I have waited so long that now i don't mind waiting some more, but I don't think that is true. If it begins looking like promised dates but nothing happening on the ground I will get really itchy. The worst thing that could happen now is that benchmarks reveal it to crush the GTX and none available to buy. I think I'll grow another ulcer then. :razz:

Graham

01-Oct-2005, 04:53

Sorry for being OT, but...

I'm quite seriously *amazed* at what the inq let faud get away with in his articles. When absolutly basic spelling and grammar is lacking there has got to be something seriously wrong.
Is the editor afraid to step in maybe?

Then again I havn't been very impressed with some of their other writers either, I recently emailed one guy letting him know that the vast majority of his write-up was badly misinformed... (In reguards to windows Vista, managed code, filesystems, security, etc).. The reply I got was nothing short of stunning:

".NET has been de-emphasised in Vista ... it simply isn't ready for prime
time yet:"

to which he then linked to an article on another site that was so laughably wrong it was stupid. It went as far as suggesting that Vista 'may only include .net on the CD as an optional extra'.

Back OT...

By 'render 4 back ends', is faud meaning render MRT to 4 RTs?
I know thats nothing special (well, maybe it would be on nvidia and it was fast), but the x1800 is listed as rendering 16 'back ends'... I thought directX 9 had a fixed limit of 4? what good would 16 do? (although the possibilies are interesting)

Demirug

01-Oct-2005, 04:54

Uttar, I think you are sorely wrong about this architecture being boring. :wink:

It looks to me like the first real departure on the PC platform from the traditional pipeline approach since powervr tile based rendering. :shock:

The new threaded architecture looks like it will bring performance to a new level by solving the latency issues within the graphics pipeline. :twisted:

This is the architectural design that will take us into the future and allow physics, clipping, shadowing, lighting, and manipulation of objects on the GPU. It is a huge departure from a fixed function pipeline of the past to a programmable highly threaded math engine. :grin:

It doesn't solve the memory latency problem better than the current ATI solution.

Talking about threads is a new hype started from ATI with the Xenos. Hell they need some big numbers after most of their numbers are below or equal with the nVidia solution. Anyway GPUs are allways highly "threaded" even back in the old T&L fixed function times.

After sleeping a little bit I am starting to become uncertain if this is really a new design. It is still possible that we see a R300 base design that is draw a little bit different and some new names on the function units. OK maybe I am to sceptically but to many persons have try to sell me old ideas as new in the past by useing simply some new words.

AlphaWolf

01-Oct-2005, 04:59

Demirug

01-Oct-2005, 05:01

By 'render 4 back ends', is faud meaning render MRT to 4 RTs?
I know thats nothing special (well, maybe it would be on nvidia and it was fast), but the x1800 is listed as rendering 16 'back ends'... I thought directX 9 had a fixed limit of 4? what good would 16 do? (although the possibilies are interesting)

"Back Ends" are normaly the last part of a pixel shader unit. They send the final values to the next unit. Sometimes you will find some fixed functions there like fog calculation or Z calc.

But it is obviously that he is talking about something different. IMHO he simply have mixed something up here. Because of this I tend to ignore this part.

amk

01-Oct-2005, 05:14

I'm quite seriously *amazed* at what the inq let faud get away with in his articles. When absolutly basic spelling and grammar is lacking there has got to be something seriously wrong.
Is the editor afraid to step in maybe?

Then again I havn't been very impressed with some of their other writers eitherTBH, I'm quite seriously amazed people care what the Inquirer has to say.

Junkstyle

01-Oct-2005, 05:27

TBH, I'm quite seriously amazed people care what the Inquirer has to say.

I read it, but I also know its just a rumor mill. Sometimes the rumors are spot on and sometimes they are completely off.

3dcgi

01-Oct-2005, 05:36

It doesn't solve the memory latency problem better than the current ATI solution.

Talking about threads is a new hype started from ATI with the Xenos. Hell they need some big numbers after most of their numbers are below or equal with the nVidia solution. Anyway GPUs are allways highly "threaded" even back in the old T&L fixed function times.

After sleeping a little bit I am starting to become uncertain if this is really a new design. It is still possible that we see a R300 base design that is draw a little bit different and some new names on the function units. OK maybe I am to sceptically but to many persons have try to sell me old ideas as new in the past by useing simply some new words.
I don't think fixed function units were threaded in the same fashion. I'm not sure they would have been threaded at all actually. Until shaders I don't think there was a need for threading. Just FIFOs and caches to hide latency.

Demirug

01-Oct-2005, 05:44

I don't think fixed function units were threaded in the same fashion. I'm not sure they would have been threaded at all actually. Until shaders I don't think there was a need for threading. Just FIFOs and caches to hide latency.

I personally would not call it threads, too. But if marketing think that "threads" are something cool you need to have they will call anything a thread as long as it match more than 1% to the general definition. I am sure if ATI have some success with this attempt to bring a new word in GPU business nVidia will follow very soon.

Razor1

01-Oct-2005, 05:53

It doesn't solve the memory latency problem better than the current ATI solution.

Talking about threads is a new hype started from ATI with the Xenos. Hell they need some big numbers after most of their numbers are below or equal with the nVidia solution. Anyway GPUs are allways highly "threaded" even back in the old T&L fixed function times.

After sleeping a little bit I am starting to become uncertain if this is really a new design. It is still possible that we see a R300 base design that is draw a little bit different and some new names on the function units. OK maybe I am to sceptically but to many persons have try to sell me old ideas as new in the past by useing simply some new words.

Actually you are right it does looks very similiar to the r300, r420 designs
http://graphics.tomshardware.com/graphic/20040504/images/vertex-engine.gif
http://graphics.tomshardware.com/graphic/20040504/images/pixelpipe1.gif

damn it no direct links

http://graphics.tomshardware.com/graphic/20040504/ati-x800-05.html
http://graphics.tomshardware.com/graphic/20040504/ati-x800-06.html

3dcgi

01-Oct-2005, 05:53

I personally would not call it threads, too. But if marketing think that "threads" are something cool you need to have they will call anything a thread as long as it match more than 1% to the general definition. I am sure if ATI have some success with this attempt to bring a new word in GPU business nVidia will follow very soon.
What would you call it then? A thread is an instruction stream and processors can have multiple threads in flight ready to be swapped in to hide memory latencies. This is what shader processors can do. IIRC it seem's like Sun's Niagra which supports 4 threads in this manner.

bigz

01-Oct-2005, 06:24

I saw a PDF a while ago discussing the thermal properties of R520 in comparison to R480. I'm not sure whether that is still the case now because there has been a respin, but the document said that R520 'leaks' 69W in comparison to R480's 62W.

It could be an explanation as to why the heatsink on X1800XT is so huge.

Demirug

01-Oct-2005, 06:31

What would you call it then? A thread is an instruction stream and processors can have multiple threads in flight ready to be swapped in to hide memory latencies. This is what shader processors can do. IIRC it seem's like Sun's Niagra which supports 4 threads in this manner.

In the case of R520 I am still need more infos before I will get them a name.

You say that the threads are swaped out to hide memory latencies but this is not even possible with the design wie see on this diagram. The latencies are somewere in the textureunit. To make it possible to hide them you need to split the whole texturesampling in two parts and connext each part to the unit they call the "Ultra Threading Dispatch Processor".

At least the current way to solve the memory latencies is simply a large FIFO/Cache. Because of this the NV4X/G70 Pipe ist very long and cause the problem with the low size of registerspace per pixelquad.

rwolf

01-Oct-2005, 07:00

In the case of R520 I am still need more infos before I will get them a name.

You say that the threads are swaped out to hide memory latencies but this is not even possible with the design wie see on this diagram. The latencies are somewere in the textureunit. To make it possible to hide them you need to split the whole texturesampling in two parts and connext each part to the unit they call the "Ultra Threading Dispatch Processor".

At least the current way to solve the memory latencies is simply a large FIFO/Cache. Because of this the NV4X/G70 Pipe ist very long and cause the problem with the low size of registerspace per pixelquad.

This is just a way to keep the execution units busy. Instead of building massive pipeline stages to hide the latency of a texture fetch they just put the thread on hold and switch to other threads while they wait for the texture to be fetched from memory and uncompressed.

The other issue is that they don't have to optimize the code for the execution units. The scheduler in the chip keeps the execution units busy. Normally all the execution units are utilized at the same time in rare cases and not every cycle.

rwolf

01-Oct-2005, 07:06

If true, then the efficiency of the 7800GTX is lower than I have been imagining. Or perhaps there are still big surprises in the shader pipeline?

That's a good point. But everything we are speculating here might not be based on fact so we shall have to wait and see.

Junkstyle

01-Oct-2005, 07:09

The worst thing that could happen now is that benchmarks reveal it to crush the GTX and none available to buy. I think I'll grow another ulcer then. :razz:

You mean like the ATI XT PE...coming to an Ebay near you, paper launch?

Demirug

01-Oct-2005, 07:21

This is just a way to keep the execution units busy. Instead of building massive pipeline stages to hide the latency of a texture fetch they just put the thread on hold and switch to other threads while they wait for the texture to be fetched from memory and uncompressed.

Yes I know this. But the point is that the whole R3XX/R4XX Series don't have this problem. I have run some tests with a special test that show how much a texturefetch (with AF) will stall the pipe. nVidia hardware show that it is much more sensitive as ATI hardware. On the other hand nVidia don't care much about dependet reads. ATI is more sensitive there.

The other issue is that they don't have to optimize the code for the execution units. The scheduler in the chip keeps the execution units busy. Normally all the execution units are utilized at the same time in rare cases and not every cycle.

You still need to make sure that you mix and order the code in a way that every part of the ALU/FPU is used. Even the size of the register array can force you to optimize they shader in a way to reduce the number of regs taken per pixel.

dizietsma

01-Oct-2005, 08:39

Demirug's and Uttars and others thoughts to how "new" r520 actually is are more pertinent in regards to this product than the G70. In the G70's case you did not have to worry too much how the design was new because they basically took a 4 cylinder engine and turned it into a 6 cylinder engine of bigger capacity, here hwoever they are trying to add VTEC to their current 4 cylinder engine to increase performance.

Arun

01-Oct-2005, 10:36

It looks to me like the first real departure on the PC platform from the traditional pipeline approach since powervr tile based rendering. :shock:

The new threaded architecture looks like it will bring performance to a new level by solving the latency issues within the graphics pipeline. :twisted:Wrong.
I'm really surprised that so many people are, well, surprised by this. I'll try to simplify the facts: Vertex Shading-wise, there is fundamentally nothing new in the R520. I'm unsure about how threaded the R420 is, considering it doesn't have to be able to handle any kind of "memory read latency", but the G70 has a per-VS mini-scheduler and a highly threaded VS pipeline.

If you look at the history of programmable graphics cards, generally speaking, programmability and flexibility was first given to the Vertex Shaders, then only to the Pixel Shaders. The only notable exception to this rule of course is proper texturing support. The GeForce3, for example, had a highly programmable Vertex Texturing unit, but only as extension of Register Combiners for the PS. I think some common mathematical operations such as sincos were also more accelerated in the VS, but I can't remember for sure.

At the same time, the GFFX tried to make the PS capable of what the VS could do, too. Yet the VS was still a few steps after it, with Dynamic Branching and a number of other features. Still, there clearly was an architectural influence from the NV20's VS, although at the same time it was still an extension of register combiners and had to take care of massively higher latency.

What I'm basically saying here is that this is the exact same thing ATI is doing here: influencing its PS architecture by its VS one. It's an evolutionary step, not a revolutionary one, and NVIDIA will do the same pre-unified Shaders with the G80.

Cheaper dynamic branching in the PS is nice, but if that's all this architecture brings to the table, once again don't color me impressed. Those benchmarks, if true, are quite impressive on the other hand, though.

Uttar

trinibwoy

01-Oct-2005, 10:53

ssassen

01-Oct-2005, 10:56

Just a heads up that I've added a picture with an overview of the architecture to my article and reworded the description somewhat to accurately reflect the details of the R520 architecture. As a fyi all this info is straight from the horse's mouth.

ATI's Radeon X1000 series, what to expect?
http://www.hardwareanalysis.com/content/article/1817/

Regards,

Sander Sassen
http://www.hardwareanalysis.com

Geo

01-Oct-2005, 10:59

I still have reservations about ATi's claims though - if the R520 is beating the GTX so badly you'd think they wouldn't be haggling with clock speeds so late in the game.

I could think of a few reasons. You might not care about older games, but that doesn't mean no one does. And that could include ATI when benchmarks are done. Then there's the Ultra threat, and beyond that it is hard to say how confident they are of when NV's own 90nm high-end part will show up.

Should I start the train of "it really isn't fair blah blah blah" to talk about R520 vs G70? How G70 is really a late response to X850? That we have to wait for NV's 90nm part to come to some conclusions about who is doing best at delivering features and performance on a level playing field? :cool:

trinibwoy

01-Oct-2005, 11:10

I could think of a few reasons. You might not care about older games, but that doesn't mean no one does. And that could include ATI when benchmarks are done. Then there's the Ultra threat, and beyond that it is hard to say how confident they are of when NV's own 90nm high-end part will show up.

I don't follow - how would higher clocks help with older, already CPU limited, games?

Should I start the train of "it really isn't fair blah blah blah" to talk about R520 vs G70? How G70 is really a late response to X850? That we have to wait for NV's 90nm part to come to some conclusions about who is doing best at delivering features and performance on a level playing field? :cool:

Nah, each side chose their weapon. No second chances! :grin: Besides, Nvidia's next card will probably be out in the R580 timeframe.

Demirug

01-Oct-2005, 11:16

I don't understand the cynicism. If the approach ATi took with R520 results in the kind of efficiency gains they are claiming then it is something to get very excited over, whether or not their scheduler is similar to what we've had in the VS before. Isn't the separation of the shaders and TMU's a big deal also? I don't think it's boring in the least.

As R300 already shows the behavoir of seperated ALUs/FPUs and TMUs I am not sure how new this is at all. In the past ATI was never open like his about the internal strukture of the pixelshader. And it is possible to draw the R420 pixelshader units like this without be totaly wrong. Sorry for being a little bit paranoid about this but I have see too many "creative" ways of drawing something in the past.

Geo

01-Oct-2005, 11:19

As R300 already shows the behavoir of seperated ALUs/FPUs and TMUs I am not sure how new this is at all. In the past ATI was never open like his about the internal strukture of the pixelshader. And it is possible to draw the R420 pixelshader units like this without be totaly wrong. Sorry for being a little bit paranoid about this but I have see too many "creative" ways of drawing something in the past.

Hey, it's cool. Your suspicious mind has given us benefits in the past. :wink: And if/when you come to the opposite conclusion you'll come say so. So it's all good at this point.

Bouncing Zabaglione Bros.

01-Oct-2005, 11:59

Cheaper dynamic branching in the PS is nice, but if that's all this architecture brings to the table, once again don't color me impressed. Those benchmarks, if true, are quite impressive on the other hand, though.

This reminds me of a software tender I was involved with. We were told that we had "the least advanced technology", but that we "offered the most functionality".

On the one hand you say you are not impressed by the architecture, but then go on to say that you are impressed with the results it brings. I think that qualifies you for a :roll:

Demirug

01-Oct-2005, 11:59

Just play a little bit with colors

http://img6.picsplace.to/img6/2/thumbs/r520s_2.JPG (http://img6.picsplace.to/img.php?file=img6/2/r520s_2.JPG)

N00b

01-Oct-2005, 12:19

There is something about the diagramm that puzzles me:

Texture Units --- Decompress --> Texture Cache
Texture Units <-- Compress --- Texture Cache

Shouldn't it be the other way round?

Xmas

01-Oct-2005, 12:34

It's safe to assume Fuad has no idea what he means.
That, and he's most likely talking about ROPs.

There is something about the diagramm that puzzles me:

Texture Units --- Decompress --> Texture Cache
Texture Units <-- Compress --- Texture Cache

Shouldn't it be the other way round?
:lol: What's even more weird is that the Texture Units are supposed to write to the texture cache...

Uttar,
"circles in a square".

IbaneZ

01-Oct-2005, 13:20

No more leaks? I hope ATI hasn't hunted down Fuads contact in Ibiza. :smile:
I wan't more more MORE!!

R520 looks like a beauty. I hope i can buy one next week. :wink:

Arun

01-Oct-2005, 13:58

This reminds me of a software tender I was involved with. We were told that we had "the least advanced technology", but that we "offered the most functionality".

On the one hand you say you are not impressed by the architecture, but then go on to say that you are impressed with the results it brings. I think that qualifies you for a :roll:
Please. There is more to results than just the underlying architecture. In software, that might be true, but in hardware it also depends on how much money you're ready to use to manufacture each and every chip/board. I personally cannot see how this architecture would give ATI a per-die-mm² performance advantage, assuming they used the same process as NVIDIA (which they aren't, but will in the G73/4/5 vs R580 timeframe).

The GTX was a low-noise, high-margins part. In fact I think it's been a long, long while since NVIDIA had such a high-margin part. What I'm impressed by here is that the actual performance is progressing fasther than I thought it would, thus forcing NVIDIA to play more aggressively too, and in the end that can only be good for the consumer.

Now, I'll reserve my judgement about great the architecture is, from a per-mm² point of view, until ATI unveils such details. Comparing the G72 with the RV515 will be very worthwhile here too.

All I'm saying -right now- is that this architecture isn't anywhere as innovative as I'd have hoped it to be, and that doesn't mean it's bad. Quite on the contrary, it seems relatively solid. But it's nothing more than a logical evolution of the R420, and it really isn't that different from the G70. In fact as I said earlier, the R520's ALU/texture separation might prove ot be a disadvantage when it comes to texturecache usage, unless they made the logic behind it significantly smarter.

Regarding the compress/decompress, IF that is midly accurate, it would just mean that the texture cache can only store uncompressed data. IIRC, NVIDIA couldn't do that either pre-NV40. The "compress" when reading from the texture cache makes no sense, though..

Uttar

Martin Eddy

01-Oct-2005, 14:02

No more leaks? I hope ATI hasn't hunted down Fuads contact in Ibiza. :smile:
I wan't more more MORE!!

R520 looks like a beauty. I hope i can buy one next week. :wink:

Living in Australia, I'll be lucky to be able to get one this year! :sad:

Bouncing Zabaglione Bros.

01-Oct-2005, 14:12

Please. There is more to results than just the underlying architecture. In software, that might be true, but in hardware it also depends on how much money you're ready to use to manufacture each and every chip/board. I personally cannot see how this architecture would give ATI a per-die-mm² performance advantage, assuming they used the same process as NVIDIA (which they aren't, but will in the G73/4/5 vs R580 timeframe).

But they are not on the same processes are they? What's the point in any underlying architecture but to do anything other than to deliver results?

If the results are good, what the architecture is doing behind the scenes is somewhat irrelevant. In fact, if R5x0 gives great results, regardless of how boring and "uninnovative" you think the design is, that just proves that ATI knows how to design VPUs for good results rather than marketing "look at our new checkbox" tech that would excite you.

Geo

01-Oct-2005, 15:05

There is something about the diagramm that puzzles me:

Texture Units --- Decompress --> Texture Cache
Texture Units <-- Compress --- Texture Cache

Shouldn't it be the other way round?

I must admit I had that thought as well.

Dave Baumann

01-Oct-2005, 15:34

The interest in R520 probably is because, ostensibly, it looks very similar to R420...

IbaneZ

01-Oct-2005, 15:42

The interest in R520 probably is because, ostensibly, it looks very similar to R420...

After five beers, i'm having a hard time understanding that. :???:

wireframe

01-Oct-2005, 15:50

After five beers, i'm having a hard time understanding that. :???:

The next one is on me!

ondaedg

01-Oct-2005, 16:02

obstensibly - adv : from appearances alone

Let me reword Dave's statement: "The interest in R520 probably is because, from appearances alone, it looks very similar to R420..."

Which suggests to me that Dave is basically saying "Don't let appearances fool you."

The only good side to not having five beers right now is that I can look up a word at dictionary.com

Still wish I had Ibanez's beers. :wink:

RoOoBo

01-Oct-2005, 16:06

512 threads looks more or less what I would have expected, if 1 quad = 1 thread = 1 PC. One fragment per thread looks too low.

I was very surprised this week to discover that what was limiting the performance in some random shader limited batches in Doom3 was the texture unit (and not bandwidth!). With anisotropic filtering enabled at a maximum of 8 samples (the Doom3 trace was taken with a NV35) many of the textures accesses required 4+ samples, or 4+ (8+ for trilinear) cycles per texture access (same than with current GPUs the simulator is configured with a througput of 1 bilinear/cycle). As the relation of texture instructions to arithmetic instructions was 1:2 or 1:3 at most for those shader programs it wasn't enough to fill the shader unit with arithmetic ops even when all accesses hit cache. The address calculation ALU and filter ALU were the bottleneck!! May be it was just that random zone from that random frame from a random Doom3 trace or the not that random (but likely more buggy) AF algorithm that the simulator implements but it was quite surprising. Of course that only happens for non stencil batches.

Geo

01-Oct-2005, 16:09

The interest in R520 probably is because, ostensibly, it looks very similar to R420...

Which is nice-speak for "Hopefully Demirug will feel better once he reads my review!".

:lol:

Which was a point I was just about to make when I saw this.

Geo

01-Oct-2005, 16:25

Various texture info goodness. . .

Well, it seems to me that this is one of the areas to investigate to put Demirug at ease. It seems fairly unlikely to me that they really disassociated the texture units without spending some extra lovin' on them at the same time. In other words, I would expect to see some clearly identifiable gains from texture ops that we can test and prove/disprove what they've done.

And maybe not just performance. There have been hints that "ultra quality" is one of their buzz phrases for this gen. Might we get better quality filtering as well, at least when in highest quality modes? It's a nice dream to dream, at least for the next few days. . .

Xmas

01-Oct-2005, 16:54

I was very surprised this week to discover that what was limiting the performance in some random shader limited batches in Doom3 was the texture unit (and not bandwidth!). With anisotropic filtering enabled at a maximum of 8 samples (the Doom3 trace was taken with a NV35) many of the textures accesses required 4+ samples, or 4+ (8+ for trilinear) cycles per texture access (same than with current GPUs the simulator is configured with a througput of 1 bilinear/cycle). As the relation of texture instructions to arithmetic instructions was 1:2 or 1:3 at most for those shader programs it wasn't enough to fill the shader unit with arithmetic ops even when all accesses hit cache. The address calculation ALU and filter ALU were the bottleneck!! May be it was just that random zone from that random frame from a random Doom3 trace or the not that random (but likely more buggy) AF algorithm that the simulator implements but it was quite surprising. Of course that only happens for non stencil batches.
That's not very surprising to me, it's the reason why "brilinear" and "AF sample optimizations" give a significant speedup. And this is why I'm somewhat worried about RV530 having only 4 TMUs. Sure, shaders are going to be more and more arithmetic-heavy, but if a single texture fetch can take a dozen cycles (throughput)...
Texture bandwidth is rarely a bottleneck if you use DXT1. 4 bits per texel isn't that much. It's RTT use and HDR content/tonemapping where you can't have enough bandwidth. Let's hope we'll see some FP compression scheme soon.

RoOoBo

01-Oct-2005, 17:09

Yes, benchmarks with high AF enabled in RV530 are going to be quite interesting. If UT2004 and Doom3 are representative of current games, and if you add on top of that AF, a ratio of 1:3 between shader ALUs and Texture ALUs looks a bit too aggressive. However for future games that may change. Developers for the XBox360 may find that they are better writing longer shaders with less texture accesses when until now some were still replacing ALU ops with texture accesses because it was faster (or was that just for the NV3x?).

Demirug

01-Oct-2005, 17:19

Which is nice-speak for "Hopefully Demirug will feel better once he reads my review!".

:lol:

Which was a point I was just about to make when I saw this.

One nice benefit you have from working for the press is that you don't need to wait for other reviews. But on the other hand this is the same time you have to stop write about things you know until the NDA is lifted. Anyway as long as you have no card to play with anything you believe to know could still be wrong.

digitalwanderer

01-Oct-2005, 17:23

Anyway as long as you have no card to play with anything you believe to know could still be wrong.
Truth. Dave only knows what ATi tells him, until he gets one to play with it's pretty much all just hearsay.

eSa

01-Oct-2005, 17:26

About 1800 pixelshader power, from http://www.hardwareanalysis.com/content/article/1817/

" A total of 16 pixel-shader processors are available on the X1800, organized in groups of four, hence there’s four shader cores, dubbed the quad-pixel shader cores. The pixels are handled in small 4x4 pixel groups, so the X1800 handles just 16-pixels per thread which makes the thread size small and efficient. Each of the pixel-shader processors can process per thread six different shader instructions on four pixels per clock-cycle. "

Soooo... is this 3 x (vecop + scalarop) for each pixel-shader processor ? 50 percent more than with 7800 (per unit) ?

Demirug

01-Oct-2005, 17:51

Truth. Dave only knows what ATi tells him, until he gets one to play with it's pretty much all just hearsay.

He is not the only one with this problem.

eSa, anybody who know the answer should still grounded by a NDA.

All I can give you are some R420/G70 values about this topic.

R420 claims 5 Instructions per pixel per clock.

- Full Vec3
- Helper Vec3
- Full scalar
- Helper scalar
- texture address

G70 claims 4+ instructions per pixel per clock
- 4*Full VecX (X = 0-4; Sum of X need to be 8 or lower; texture address need one Full VecX)
- + = one FP16 nrm ; helper for each full unit.

Even this makes a R420 and G70 hard to compare. There are more differences. You should not think that R520 vs. G70 will be easier to compare at this theoretical level.

cho

01-Oct-2005, 19:02

Soooo... is this 3 x (vecop + scalarop) for each pixel-shader processor ? 50 percent more than with 7800 (per unit) ?

it is not such thing .

Apple740

01-Oct-2005, 19:53

Ok, the Ibiza thing ended almost 24hours ago, and still no nada nothing leaking? Even Fuad has nothing?

I'm dying here :)
http://www.mawhyte.pwp.blueyonder.co.uk/images/woohoo.gif

Karma Police

01-Oct-2005, 20:22

It was a TRAP! The R520 was so advanced it had a mind of it's own. But it could only survive if it ate the brains of computer journalists and hacks.

Now they're all dead. DEAD, Jim!

Geo

01-Oct-2005, 20:57

Ok, the Ibiza thing ended almost 24hours ago, and still no nada nothing leaking? Even Fuad has nothing?

Sez who? I linked a French site somewhere around here that says it is a 3-day shindig (i.e. thru tomorrow).

rwolf

01-Oct-2005, 21:02

Yes I know this. But the point is that the whole R3XX/R4XX Series don't have this problem. I have run some tests with a special test that show how much a texturefetch (with AF) will stall the pipe. nVidia hardware show that it is much more sensitive as ATI hardware. On the other hand nVidia don't care much about dependet reads. ATI is more sensitive there.

You still need to make sure that you mix and order the code in a way that every part of the ALU/FPU is used. Even the size of the register array can force you to optimize they shader in a way to reduce the number of regs taken per pixel.

These points are exactly why I think this architecture is revolutionary. You don't have to manually make sure every part of the ALU/FPU are used because the scheduler does that for you. The architecture would handle stalls caused by texture fetches, loops, dependent reads.

Geo

01-Oct-2005, 21:06

Truth. Dave only knows what ATi tells him, until he gets one to play with it's pretty much all just hearsay.

You don't think Wavey wrote that post thru the miracle of the internet from Ibiza, after a full day of prodding secrets out of R520? Well, maybe you're right, I haven't gotten a postcard or anything. :lol:

And, so far as that goes, it might take the community as a whole a few months to rip most of the best secrets from behind R520's veil --tho I hope not.

kemosabe

01-Oct-2005, 21:55

Are we timidly suggesting that Wavey is learning stuff that he himself didn't expect to? Because I don't recall him ever referring to R520 being ostensibly similar to R420 before. :)

rwolf

01-Oct-2005, 22:45

Wrong.
I'm really surprised that so many people are, well, surprised by this. I'll try to simplify the facts: Vertex Shading-wise, there is fundamentally nothing new in the R520. I'm unsure about how threaded the R420 is, considering it doesn't have to be able to handle any kind of "memory read latency", but the G70 has a per-VS mini-scheduler and a highly threaded VS pipeline.

Uttar

It's not the same. The mini-scheduler is feeding pipelines and the threading is a SIMD unit. Same terminology, but with a huge difference in implementation.

rwolf

01-Oct-2005, 22:47

Are we timidly suggesting that Wavey is learning stuff that he himself didn't expect to? Because I don't recall him ever referring to R520 being ostensibly similar to R420 before. :)

Although he was the only person I know who showed us that R520 was influenced by R400/R500 yet also was derived from R420.

I bet R520 was designed to minimize driver changes while having all new internals.

SugarCoat

01-Oct-2005, 22:59

Although he was the only person I know who showed us that R520 was influenced by R400/R500 yet also was derived from R420.

I bet R520 was designed to minimize driver changes while having all new internals.

couldnt you say that about many if not all progressions of cores? seems like such a broad statement.

Geo

01-Oct-2005, 22:59

Are we timidly suggesting that Wavey is learning stuff that he himself didn't expect to? Because I don't recall him ever referring to R520 being ostensibly similar to R420 before. :)

Interesting point. I hope he sees fit in a forum post somewhere to talk a little more about his own evolving understanding. Certainly last Spring he was pushing the "just a refinement of R420 and continuation of DX9 era" theme pretty hard (while also, we should note, sticking consistently to his veiled guns on the ring bus). Whether that was because he fully believed it, or was consciously trying to beat back the high expectations driven by "32 pipes" "R400 tech" "R500 tech" etc and pulled back a little too hard on the reins, would be interesting to hear more on that story. If he can tell it. :smile:

kemosabe

01-Oct-2005, 23:15

I bet R520 was designed to minimize driver changes while having all new internals.

One could argue that whether the hardware was designed with this in mind or not, the four-month launch delay should have given the Catalyst crew more than enough time to optimize the drivers. "Driver immaturity" likely won't meet with much sympathy as an excuse if performance isn't up to par.

Ailuros

01-Oct-2005, 23:17

It's not the same. The mini-scheduler is feeding pipelines and the threading is a SIMD unit. Same terminology, but with a huge difference in implementation.

Irrelevant of IHV, everytime I hear terms like "huge" or "gigantic" when it comes to things like that, I put a larger grain of salt over it.

Xmas

01-Oct-2005, 23:36

It's not the same. The mini-scheduler is feeding pipelines and the threading is a SIMD unit. Same terminology, but with a huge difference in implementation.
What do you mean by "the threading is a SIMD unit"? Threading has no relations to the units being SIMD, at least not in this context.

Junkstyle

01-Oct-2005, 23:50

Three more days and the 520 rumors all stop. This is going to be so nice. But how many days then until the r580 rumors start making 800 posts threads?

Bouncing Zabaglione Bros.

02-Oct-2005, 00:21

Three more days and the 520 rumors all stop. This is going to be so nice. But how many days then until the r580 rumors start making 800 posts threads?

About four days I think. :wink: :lol: :shock:

CMAN

02-Oct-2005, 01:28

About four days I think. :wink: :lol: :shock:
3 days and 12 hours and 31 minutes (8:29 PM EST) to be exact :grin:

rwolf

02-Oct-2005, 01:34

One could argue that whether the hardware was designed with this in mind or not, the four-month launch delay should have given the Catalyst crew more than enough time to optimize the drivers. "Driver immaturity" likely won't meet with much sympathy as an excuse if performance isn't up to par.

ATI has probably been writing drivers for R520 for over a year.

In fact don't they usually have them pretty much ready to go when they get the first silicon back.

rwolf

02-Oct-2005, 01:45

What do you mean by "the threading is a SIMD unit"? Threading has no relations to the units being SIMD, at least not in this context.

http://www.gamedev.net/reference/articles/article1496.asp

The Vertex Shader ALU is a multi-threaded vector processor that operates on quad-float data. It consists of two functional units. The SIMD Vector Unit is responsible for the mov, mul, add, mad, dp3, dp4, dst, min, max, slt and sge instructions. The Special Function Unit is responsible for the rcp, rsq, log, exp and lit instructions. Most of these instructions take one cycle to execute, rcp and rsq take more than one cycle under specific circumstances. They take only one slot in the vertex shader, but they actually take longer then one cycle to execute, when the result is used immediately, because that leads to a register stall.

Well, I don't think I said it correctly here is a better explaination of what I was trying to say.

As you can see multi-threading in this context is not the same as what ATI is doing.

101

02-Oct-2005, 02:10

Three more days and the 520 rumors all stop.

Except for the rumors of availability of x1800xt master cards, slave cards, and crossfire motherboards. I don't have any hopes of putting a high end crossfire system together before thanksgiving, if by then.

Ailuros

02-Oct-2005, 02:19

http://www.gamedev.net/reference/articles/article1496.asp

Well, I don't think I said it correctly here is a better explaination of what I was trying to say.

As you can see multi-threading in this context is not the same as what ATI is doing.

I still don't understand what you mean. Even more since the VS units since NV3x are MIMD. That quoted paragraph refers to a NV2x.

Kanyamagufa

02-Oct-2005, 03:10

At least we can start some actual predicitons about R580s performance once R520 comes out. Right now, I think it's safe to say that unless you're under an NDA, you don't really know the future of ATI's graphics line performance-wise.

SugarCoat

02-Oct-2005, 03:56

ATI has probably been writing drivers for R520 for over a year.

In fact don't they usually have them pretty much ready to go when they get the first silicon back.

that has never been the case. This core has had the most work done to its internals since the R300. We remember how awesome that card was with the drivers that launched with it right :). Then we've had releases resembling the most recent NV40 where the card launched in summer, but it took to the holidays for the drivers to mature. That core made up alot of space in terms of performance gap between it and the X800XT/PE in a 6 month period after launch.

The lesson is that no drivers launched with a card are their best performing. Usually takes some time after launch for that.

Razor1

02-Oct-2005, 04:54

that has never been the case. This core has had the most work done to its internals since the R300. We remember how awesome that card was with the drivers that launched with it right :). Then we've had releases resembling the most recent NV40 where the card launched in summer, but it took to the holidays for the drivers to mature. That core made up alot of space in terms of performance gap between it and the X800XT/PE in a 6 month period after launch.

The lesson is that no drivers launched with a card are their best performing. Usually takes some time after launch for that.

The respins weren't ment to change the way the chip executes data so in all likelyhood the drivers are mature by now. If they still aren't mature well thats just not good for publicity.

Geo

02-Oct-2005, 05:27

Re ATI calling out texture compression on the diagrams this time, I was just reading Anand's article on Valve HDR and noticed this line:

floating point assets cannot be compressed currently using built-in hardware texture compression techniques.

Could one hope?

101

02-Oct-2005, 05:58

The lesson is that no drivers launched with a card are their best performing. Usually takes some time after launch for that.

No doubt, the true test of drivers is when thousands of monkeys start beating on them after a card is launched. Try as they might, there's no substitute for a real luggage test.

Razor1

02-Oct-2005, 05:58

Re ATI calling out texture compression on the diagrams this time, I was just reading Anand's article on Valve HDR and noticed this line:

Could one hope?

hmm isn't Open EXR a compression format?

rwolf

02-Oct-2005, 06:34

that has never been the case. This core has had the most work done to its internals since the R300. We remember how awesome that card was with the drivers that launched with it right :). Then we've had releases resembling the most recent NV40 where the card launched in summer, but it took to the holidays for the drivers to mature. That core made up alot of space in terms of performance gap between it and the X800XT/PE in a 6 month period after launch.

The lesson is that no drivers launched with a card are their best performing. Usually takes some time after launch for that.

I just don't think 4 months gives them much time to do much with the drivers.

Chalnoth

02-Oct-2005, 06:50

I just don't think 4 months gives them much time to do much with the drivers.
It does, though, because previous to launch, they can't build large numbers of test systems to do nearly as much testing as they can after launch. The quicker response time between when a driver change is implemented and when it is tested on a variety of platforms and software allows for much quicker development of the driver.

kemosabe

02-Oct-2005, 06:52

I just don't think 4 months gives them much time to do much with the drivers.

If my name was Makedon I'm not sure I'd consider that a compliment. ;)

Xmas

02-Oct-2005, 11:16

Re ATI calling out texture compression on the diagrams this time, I was just reading Anand's article on Valve HDR and noticed this line:

Could one hope?
I don't think so. Neither decompression from TMU to texture cache nor compression from texture cache to TMU make any sense. I think it's just a c&p mistake in the diagram.

http://www.gamedev.net/reference/articles/article1496.asp

Well, I don't think I said it correctly here is a better explaination of what I was trying to say.

As you can see multi-threading in this context is not the same as what ATI is doing.
That description in this paragraph doesn't relate to threading at all. The threading NVidia refers to in the VS is about having multiple vertices in flight in a single VS unit, switching between them to cover the instruction latencies.

I still don't understand what you mean. Even more since the VS units since NV3x are MIMD. That quoted paragraph refers to a NV2x.
The VS array is MIMD (meaning that each VS unit has its own instruction pointer and can branch independently). Each vector unit is SIMD.

_xxx_

02-Oct-2005, 12:40

chavvdarrr

02-Oct-2005, 12:43

I don't give a damn about the architecture itself. What I'd like to know:

- is it faster than the GTX? How much?
- does it have better IQ or some new IQ-features?
- will it cost less than a new car?

What's under the hood is pretty irrelevant to me... :smile:
one more question:
- would one need a nome-made fusion reactor to feed the beast and its cooler?
;)

Geo

02-Oct-2005, 13:40

I don't give a damn about the architecture itself. . .
What's under the hood is pretty irrelevant to me...

Anyone with nearly 3k posts at B3D who can say that either 1) Doesn't know themselves very well or 2) Their Delorean broke down here some years ago and they haven't been able to fix the flux capacitor.

mikechai

02-Oct-2005, 14:05

I don't give a damn about the architecture itself. What I'd like to know:

- is it faster than the GTX? How much?
- does it have better IQ or some new IQ-features?
- will it cost less than a new car?

What's under the hood is pretty irrelevant to me... :smile:

1. Depends.
2. Yes and no.
3. Definitely.

:smile:
________
YAMAHA YZF750 (http://www.yamaha-tech.com/wiki/Yamaha_YZF750)

IbaneZ

02-Oct-2005, 14:14

http://www.xtremesystems.org/forums/showthread.php?t=75608&page=1&pp=25

3D mark-05 competition later today. :smile:

http://www.thelab.gr/attachment.php?s=&postid=290590

Geo

02-Oct-2005, 14:18

VR-Zone on X1xxx 3dm scores. Provided for giggle value only.

http://www.vr-zone.com/?i=2793&s=1

Gotta love that "8k" guess of theirs.

Geo

02-Oct-2005, 14:20

http://www.xtremesystems.org/forums/showthread.php?t=75608&page=1&pp=25

3D mark-05 competition later today. :smile:

9700 in 3dm05 at default?

Somewhere, I believe I saw a post from someone confirming that Macci was on his way to Ibiza. . .

Edit: Btw, if you stitch together two of those pics for the promo board behind them, it seems to say "Future Technology Avaliable Now".

_xxx_

02-Oct-2005, 16:12

Anyone with nearly 3k posts at B3D who can say that either 1) Doesn't know themselves very well or 2) Their Delorean broke down here some years ago and they haven't been able to fix the flux capacitor.

You forgot "...is fed up with stupid rumors and waits for some actual facts without wearing his nerves with same speculations over and over again".

And still, what it comes down to is the user experience. In that case, I wouldn't care if there's a steam engine underneath or some alien technology. It has to deliver, that's where it lives or dies ;)

EDIT:
Btw, it would be about 5k posts if the two non-counting forums would count... :)

Sunrise

02-Oct-2005, 16:32

http://www.thelab.gr/attachment.php?s=&postid=290593

Now this is what i found most interesting.
Taking a closer look at the second line:

215RFNBKA15FG -> 1 respin and 5 metal changes ? Looks like the ASIC that HKEPC had back then was quite old and assumingly still had the famous soft-ground issue.

On the contrary, why does it indicate 1 respin, when actual common knowledge leads to at least 2 respins. Silicon change ?
Maybe i´m just reading it wrong, feel free to correct me on this.

digitalwanderer

02-Oct-2005, 16:47

On the contrary, why does it indicate 1 respin, when actual common knowledge leads to at least 2 respins. Silicon change ?
Maybe i´m just reading it wrong, feel free to correct me on this.
Eeeeenteresting, thank you!

tEd

02-Oct-2005, 17:01

i think it's 5 respins and 1 metal layer change. r300 was a13 and r420 was a11

Kombatant

02-Oct-2005, 17:29

I see people have already posted pictures and stuff from here in Ibiza. What really matters is that we're actually having a good time :cool:

Arty

02-Oct-2005, 17:35

They already reached 856/1880 ... :shock:

Kombatant

02-Oct-2005, 17:37

They already reached 856/1880 ... :shock:

Btw the 3dmark score you probably saw was with stock AthlonFX speeds; from now on the thing will get even more interesting heh

Arty

02-Oct-2005, 17:49

Btw the 3dmark score you probably saw was with stock AthlonFX speeds; from now on the thing will get even more interesting heh
Yeah, 625/1500 as default is confirmed now.

Chalnoth

02-Oct-2005, 17:57

They already reached 856/1880 ... :shock:
Sure, but people have gotten the 7800 GTX to 700MHz+, which at 24 pipes probably means higher performance, for example:
http://service.futuremark.com/compare?3dm05=1243004

Basically, ATI would have to achieve about 15% more performance per clock, per pipeline to have parity at 900MHz core to the highest-clocked GTX cores.

stevem

02-Oct-2005, 18:04

Sure, but people have gotten the 7800 GTX to 700MHz+...
Lemme know when you can do that air-cooled...

satein

02-Oct-2005, 18:07

Btw the 3dmark score you probably saw was with stock AthlonFX speeds; from now on the thing will get even more interesting heh
Aren't you at Ibiza? :shock: It would be very interesting week coming.

Chalnoth

02-Oct-2005, 18:09

Lemme know when you can do that air-cooled...
Are you saying that you think the R520 can get to 900MHz air-cooled?

Because at 900MHz it will only reach fillrate parity with a 600MHz 7800 GTX, which seems to be a common overclock for the GTX (by browsing the ORB).

Arty

02-Oct-2005, 18:13

Sure, but people have gotten the 7800 GTX to 700MHz+, which at 24 pipes probably means higher performance, for example:
http://service.futuremark.com/compare?3dm05=1243004

Basically, ATI would have to achieve about 15% more performance per clock, per pipeline to have parity at 900MHz core to the highest-clocked GTX cores.
Isnt the GTX a good overclocker? I was implying the same thing with the X1800. They are also talking about adjusting vcore settings without any hard mods .. sounds interesting.

Aren't you at Ibiza? :shock: It would be very interesting week coming.
Maybe he is at Ibiza. :twisted:

stevem

02-Oct-2005, 18:14

Are you saying that you think the R520 can get to 900MHz air-cooled?
Are you saying that you think the G70 can get to 700MHz+ air-cooled?

As per your addendum, for this to be an issue 600MHz for G70 needs to also be on air.

Chalnoth

02-Oct-2005, 18:14

Yeah, I'm just trying to put the higher clocks into perspective. Though the adjustable vcore would make overclockers happy.

John Reynolds

02-Oct-2005, 18:15

Because at 900MHz it will only reach fillrate parity with a 600MHz 7800 GTX, which seems to be a common overclock for the GTX (by browsing the ORB).

Are we forgetting the ROP limitations?

XSBagage

02-Oct-2005, 18:18

Aren't you at Ibiza? :shock: It would be very interesting week coming.
I guess they have internet there too ;)

Chalnoth

02-Oct-2005, 18:21

Are we forgetting the ROP limitations?
No. I thought this had been discussed enough. Are we going to quibble over me using the word "fillrate" instead of "number of shader pipelines times clockspeed?"

Bruce

02-Oct-2005, 18:23

They already reached 856/1880 ... :shock:

which sort of cooling if i may ask? :)

Chalnoth

02-Oct-2005, 18:24

As per your addendum, for this to be an issue 600MHz for G70 needs to also be on air.
Well, yeah. Given that I see dozens of 600MHz+ GTX's on the ORB, I would tend to guess that these are air-cooled. I can't be sure, though.

On a side note, I did find this juicy result:
http://service.futuremark.com/compare?3dm05=1002966
...810/1521MHz clocks...

Arty

02-Oct-2005, 18:26

which sort of cooling if i may ask? :)
Dont know. It looked like that was the first clocks they hit before they started pumping it up again.

digitalwanderer

02-Oct-2005, 18:43

No. I thought this had been discussed enough. Are we going to quibble over me using the word "fillrate" instead of "number of shader pipelines times clockspeed?"
You kidding? Quibblig over things like that is what we live for! ;)

Sinistar

02-Oct-2005, 18:47

Dont know. It looked like that was the first clocks they hit before they started pumping it up again.

In the picture I saw, they had LN.

satein

02-Oct-2005, 18:49

I guess they have internet there too ;)
I knew that point... but I am amazing on Kom likely play snake and peek too :grin: . Anyway, next week topic may be going to change to what NV will come out for response. Hope it would not be only 7800GTX with 512MB and a MAGIC driver update.

BRiT

02-Oct-2005, 19:10

Well, yeah. Given that I see dozens of 600MHz+ GTX's on the ORB, I would tend to guess that these are air-cooled. I can't be sure, though.

I would bet a good portion of those are all on watercooled systems, using Maze4 gpu blocks or the like. There's a large following of watercoolers at XtremeSystems that all run GTXs in their loop...

R300King!

02-Oct-2005, 19:14

I read this in xtreme Systems forums about some numbers:

FROM: hipro5:
"EDIT: ...."our" man from in there: "I just can't give you numbers yet till Wednesday, but I benched a bit with it on the 3D Mark 2005 and I must confess that with the X1800 on a 4000+, I got much more marks than the GTX on a FX57"......."

So my question is : Does anyone know what the average difference in 3DMark05 score is from a 4000+ to a FX57?

Also, he said 'much more marks', so this should give it a bit more as well. ;)

Chalnoth

02-Oct-2005, 19:19

Bah, who really cares about 3DMark05 scores? If true, all this basically means is that the R520 has a higher vertex processing rate than the G70, but I think we all expected that.

Bruce

02-Oct-2005, 19:26

So my question is : Does anyone know what the average difference in 3DMark05 score is from a 4000+ to a FX57?

there is almost no difference in default, maybe not even 100 points.

R300King!

02-Oct-2005, 19:38

there is almost no difference in default, maybe not even 100 points.

Really? Hmmm, I wonder why he'd mention it even then. :-?

9700 marks is pretty impressive though. What do you think Bowen? Is your 7800GTX on eBay yet? :D

Geo

02-Oct-2005, 19:43

Are we forgetting the ROP limitations?

Maybe we'll have enuf data points now to figure out if those even exist on 16-ROP cards this side of, oh, Q3 at 8x6! :lol:

Edit: I mean, if, as we think we know, ATI is planning to triple shader power in R580 without bumping ROPs, that pretty strongly suggests to me that ROP is just never the constraint in modern situations on 16-ROP cards.

Geo

02-Oct-2005, 19:46

dup deleted

Rys

02-Oct-2005, 19:48

Bah, who really cares about 3DMark05 scores? If true, all this basically means is that the R520 has a higher vertex processing rate than the G70, but I think we all expected that.
Ding ding ding, we have a winner! Who cares :lol:

Ibiza was fun, was good to see a bunch of faces I know from online encounters (watching Fudo and Lars do flamenco was fun, eh Kombatant? :lol: ) and these here forums.

Some of us paid attention during the presentations, too....

http://www.lungexpress.com/images/rys/wavey_pays_attention.jpg

He shouldn't be allowed to these things, he just browses these forums or sleeps :lol:

digitalwanderer

02-Oct-2005, 20:03

Ding ding ding, we have a winner! Who cares

Ibiza was fun, was good to see a bunch of faces I know from online encounters (watching Fudo and Lars do flamenco was fun, eh Kombatant?) and these here forums.

Some of us paid attention during the presentations, too....

http://www.lungexpress.com/images/rys/wavey_pays_attention.jpg

He shouldn't be allowed to these things, he just browses these forums or sleeps
http://69.93.88.162/forum/images/smiles/rofl.gif

What is "flamenco"?

Geo

02-Oct-2005, 20:22

http://69.93.88.162/forum/images/smiles/rofl.gif

What is "flamenco"?

Spanish dance. You've certainly seen it. There is much clapping, stomping, and ruffled clothing/swirling skirts involved.

ClyssaN

02-Oct-2005, 20:32

hmmm i live in Portugal - Algarve, i should have been there ... :smile: i could get in just saying "i'm a beyond3d forum member" compreendes ?!

:wink:

Geo

02-Oct-2005, 20:35

hmmm i live in Portugal - Algarve, i should have been there ... :smile: i could get in just saying "i'm a beyond3d forum member" compreendes ?!

:wink:

And they'd respond, "Sorry, only Senior Members and Regulars with Reputation of +10 or higher!".

Kidding! Kidding!!

ClyssaN

02-Oct-2005, 20:43

And they'd respond, "Sorry, only Senior Members and Regulars with Reputation of +10 or higher!".

Kidding! Kidding!!

errrrrr then i would say that despite my post number, and register date, i visit beyond3d everyday since the beginning :wink:

stevem

02-Oct-2005, 20:44

Well, yeah. Given that I see dozens of 600MHz+ GTX's on the ORB, I would tend to guess that these are air-cooled. I can't be sure, though.
That's adventurous. How about ~530MHz on air without throttling? 600MHz is likely water cooled. Active/LN + VMod + BIOS latency mods for higher clocks. (Not that I have an FM account to check).

If true, all this basically means is that the R520 has a higher vertex processing rate than the G70, but I think we all expected that.
As an aside, given G70 multiple clock domains, VS run at a higher clock than core clock (ROP/PS), IIRC. Interesting portents for future parts.

I mean, if, as we think we know, ATI is planning to triple shader power in R580 without bumping ROPs, that pretty strongly suggests to me that ROP is just never the constraint in modern situations on 16-ROP cards.
Given current architecture features/balance, I suspect 16 ROPs may not be an issue for some time. (Perhaps SLI/XFire also confirm this to a degree?)

Geo

02-Oct-2005, 20:53

Given current architecture features/balance, I suspect 16 ROPs may not be an issue for some time. (Perhaps SLI/XFire also confirm this to a degree?)

Fairly impressive mid-range cards with 4 ROPs would suggest it as well. I don't think we're going to see any 4-1 whupass opened up between top-end and mid-range . . .

Junkstyle

02-Oct-2005, 21:59

I really dont give a shit about 3dmark scores. I never have and I never will. I want to see real game benches:

BF2 benches
CoD2 benches
Elderscrolls Oblivion benches
HL2 benches
Quake4 benches
Farcry benches

I also want to see image quality comparisons in various settings and modes.

Geo

02-Oct-2005, 22:14

"Ultra-high Image Quality" it says at ati.com when you set your date to 10/5.

That's gotta be more than just AA on HDR. It's gotta. . .

"Live Tinkerbell, live!!" :lol:

AlphaWolf

02-Oct-2005, 22:16

"Ultra-high Image Quality" it says at ati.com when you set your date to 10/5.

That's gotta be more than just AA on HDR. It's gotta. . .

"Live Tinkerbell, live!!" :lol:

Don't forget avivo.

wireframe

02-Oct-2005, 22:18

"Ultra-high Image Quality" it says at ati.com when you set your date to 10/5.

That's gotta be more than just AA on HDR. It's gotta. . .

I don't read too much into that. Nvidia could have said the same when they introduced transparency AA. Whenever the word "ultra" sneaks its way into a sentence I stop thinking it is something from engineering and suspect 90% of the "product" was created by marketing. But I suppose one can always hope.

BRiT

02-Oct-2005, 22:26

Junkstyle

02-Oct-2005, 22:32

"Ultra-high Image Quality" it says at ati.com when you set your date to 10/5.

lol i just tried it:

"Enter the World of Visual Velocity"
"Ultra-high Image Quality:
"What's Next? 10/05/05"

kemosabe

02-Oct-2005, 22:52

AlphaWolf

02-Oct-2005, 23:03

PCI-SIG now showing four (http://www.pcisig.com/developers/compliance_program/integrators_list/pcie/) next-generation discrete mobile ASICs in contrast to the three on the desktop.

Any thoughts on which one doesn't have a desktop counterpart?

Not sure what you are seeing as 4 vs 3. I see a lot more than 7 products based on rv5xx, r5xx.

er nm. I was counting some of the gl parts.

02-Oct-2005, 23:04

PCI-SIG now showing four (http://www.pcisig.com/developers/compliance_program/integrators_list/pcie/) next-generation discrete mobile ASICs in contrast to the three on the desktop.

Any thoughts on which one doesn't have a desktop counterpart?

M52 could be the mobile RV505 chip? That one hasn't officially been announced. And seeing as the X1K series starts at X1300, it leaves some room for X1100/X1200.

Skrying

02-Oct-2005, 23:09

I thought one was listed as an endpoint controller and not graphics? Hmm, that list was making me sleepy and couldnt look at it for very long. Man, its only 5PM and I'm already tired....

Dave Baumann

03-Oct-2005, 01:04

I don't give a damn about the architecture itself. What I'd like to know:
What's under the hood is pretty irrelevant to me... :smile:

Seriously, what are you doing here then?

ChrisRay

03-Oct-2005, 01:13

Seriously, what are you doing here then?

Probably for your tidbits. ;)

Geo

03-Oct-2005, 01:19

Probably for your tidbits. ;)

Which are generally about architecture and "what's under the hood". :wink:

Tho I think xxx second post pretty much explained it. He's ready to end the courtship and get right to the nasty on this one. I can understand that at this point, as many of us were there in May before the coitus interruptus happened. . . :lol:

Chalnoth

03-Oct-2005, 01:41

Taken from a slightly different perspective, a whole lot of the speculation and rumors around the R520's architecture really boil down to how well it performs. From the ring bus to the clock speed to the pipeline layout, none of these affect anything but performance. So exactly how it works may be, for some, much less interesting than how it performs in the end. Especially considering how it's going to be nearly impossible to figure out how these various architectural tidbits actually affect the bottom line.

Now, there are other aspects, like Crossfire support and AA whose results can be very easy to predict based upon the base technology, and so the base technology becomes a whole lot more interesting.

Dave Baumann

03-Oct-2005, 01:44

Geo

03-Oct-2005, 01:49

Well, in the end it's a jigsaw puzzle, isn't it? And IHV's cut up the pieces in different ways, tho obviously hoping when the puzzle is put together it makes the most pleasing picture. . . Certainly the masses are only interested in the final puzzle. But around here, we like to know how the pieces fit together for various reasons, not least what it means for the future. . .

Kanyamagufa

03-Oct-2005, 01:49

I think a huge number of people were very interested how the R520 was built, what the new architecture from ATI was going to be, and how it was going to work and handle the various issues a graphics card comes up against. I know I was.

But for the most part, we know what this graphics card is. We also know how it's made and what the feature set and theoretical functunality is. And we've known all that for a while, or at least long enough to get increasingly bored with waiting. We've exhausted ourselves.

The only thing we don't know is actual performance, and being techies, we crave new information. So I also think it's understandable to be anxious and also a little impatient.

It's getting closer and closer to go time though, so let's not lose our cool.

Dave Baumann

03-Oct-2005, 01:53

As a side note, I find it odd that its been on the web for well in excess of an hour now and nobody has yet linked to it (and it partially one of the reasons why I'm a little testy right now)...

tEd

03-Oct-2005, 01:57

Skrying

03-Oct-2005, 01:58

Geo

03-Oct-2005, 01:58

igpay atinlay! Something!

Xenus

03-Oct-2005, 01:59

Where! What!

And why is it making him testy? :???:

Skrying

03-Oct-2005, 01:59

http://www.techpowerup.com/reviews/ATI/R5XX/

Has this been posted?

tEd

03-Oct-2005, 02:02

http://www.techpowerup.com/reviews/ATI/R5XX/

Has this been posted?

it has now ;)

Skrying

03-Oct-2005, 02:02

it has now ;)

OMG!

Geo

03-Oct-2005, 02:07

"New cache design/Latency improvements".

The AA performance on X1300. :shock:

tEd

03-Oct-2005, 02:09

up to 6x Alpha test AA
fp16 AA support
better AF quality

i think i fell in love :)

Skrying

03-Oct-2005, 02:09

Oh yes, AA with HDR is mentioned for this cards in those spec sheets. And if those performance numbers are real then well.......... wow.

Geo

03-Oct-2005, 02:14

"Extreme anti-aliasing and texture filtering quality." :grin:

This would also seem to be where the 3dm05 9100 score came from. . .

"Over 95% efficiency for shader processing". :grin:

6xaa twice as fast as X850 :shock:

digitalwanderer

03-Oct-2005, 02:15

Err, thats quite obvious Chalnoth, the point being that "whats under the hood" is precisely what makes B3D tick; if relative game performance is all you are interested in then this is inherantly the wrong location.
Not really, this is the best placed to get an idea of relative game performance before the cards debut. ;)

tEd

03-Oct-2005, 02:17

I think this time ati invested in the right things. I really hope reviewers can handle it(don't think so though)

Skrying

03-Oct-2005, 02:21

Wow, I cant wait for the 5th to come now. I want to see side by side comparisons for these cards. Those sheets show some simply amazing performance improvments on some new games. Its all so interesting, and I think ATi really has a winner on their hands. Features, features, and more features, and performance to go with those.

And this makes me even more exicted to see if Nvidia can top this, omg, this is why I love computers!

Geo

03-Oct-2005, 02:24

I think this time ati invested in the right things. I really hope reviewers can handle it(don't think so though)

Well, just a note of relative caution (tho I like what I've seen so far!). The X1800XT benchies in the document are probably ones where the memory speed advantage alone probably provides. . .what? 25%? Just by itself.

Skrying

03-Oct-2005, 02:26

Well, just a note of relative caution (tho I like what I've seen so far!). The X1800XT benchies in the document are probably ones where the memory speed advantage alone probably provides. . .what? 25%? Just by itself.

True, but the X1800XL has the same bandwidth as the 7800GT, which will be a very interesting match up, and will really let us dial into what the core does to gain the advantages that its looking like it has. I'm still taking these numbers in caution, but wow.

IgnorancePersonified

03-Oct-2005, 02:29

I'm at a loss as to why they would include the results for the FEAR SP demo??

Chalnoth

03-Oct-2005, 02:29

http://www.techpowerup.com/reviews/ATI/R5XX/

Has this been posted?
Looks like they're claiming that the X1300 Pro is somewhere between the 6600 and the 6600 GT, at $149 MSRP.

Since the 6600 is around $100, and the GT around $150, it seems to me that $125 street would be a competitive price.

I think this is basically what we should all have expected: it's on par, no great breakthroughs in performance for the price.

Arty

03-Oct-2005, 02:29

3 letters: O M G !

;)

digitalwanderer

03-Oct-2005, 02:29

http://www.techpowerup.com/reviews/ATI/R5XX/

Has this been posted?
Thanks, frontpaged (http://www.elitebastards.com/#news12092).

Skrying

03-Oct-2005, 02:30

I'm at a loss as to why they would include the results for the FEAR SP demo??

Because FEAR is going to be a big title game? And I'm guessing lots of people are curious, as the demo was pretty damn hard on a lot of cards out there.

Skrying

03-Oct-2005, 02:32

Thanks, frontpaged (http://www.elitebastards.com/#news12092).

Hehe, awesome, but the name is one word. Skrying, as in predicting the future through a crystal ball.

Chalnoth

03-Oct-2005, 02:33

Hrm, except the spec sheet does specifically state MSAA support for FP16 rendertargets. This was more or less expected as well, but is a definite advantage for ATI.

AlphaWolf

03-Oct-2005, 02:37

Looks like they're claiming that the X1300 Pro is somewhere between the 6600 and the 6600 GT, at $149 MSRP.

Since the 6600 is around $100, and the GT around $150, it seems to me that $125 street would be a competitive price.

I think this is basically what we should all have expected: it's on par, no great breakthroughs in performance for the price.

Well the x1600 pro 128mb version is also $149. Also a quick gander at price watch shows me I wont find a 256mb version of the 6600 gt for $150.

IgnorancePersonified

03-Oct-2005, 02:38

Yeh I get that but it's credibility. The software is demo - as in beta so practically useless for judging a buying decision on. I am buying a card in two weeks - I look at the charts and have to rescale every one of them to get rid of the obvious anomaly that is the fear results. It skews the results and, to me anyway, looks like a smoke and mirrors job. Just my opinion - a foolish result to include in official marketing material.

In 1 month time is their going to be such a deficit in performance between the two camps?

Why not include the COD2 demo resluts with a 512mb card vs any nvidia 256 mb card?

edit:cod2

SugarCoat

03-Oct-2005, 02:38

I'm at a loss as to why they would include the results for the FEAR SP demo??

FEAR is Nvida backed, think of it as their DOOM3 only they can use it against them. The 80 series drivers are suppose to help out the current Nvidia cards in that game in particuliar, but im willing to bet that card and that architecture is tailored very well for that game. It wouldnt surprise me if the performance difference was infact that huge currently, so why not market it? :)

Nothing Nvidia didnt do for Doom3 and their cards. We'll still have to wait im not overly impressed just yet.

% improvment over GTX

10% improvment in HL2 demo tested
10% improvment in Splinter demo tested
~18% improvment in painkiller demo tested
80% improvment in FEAR demo tested
32% improvment in FarCry demo tested
38% improvment in BF2 demo tested

Stock score of 9100 in 3DMARK05

So it begs a couple questions from me, are they using stock GTXs (most likely) or the overclocked ones which are more prevalant in enthusiasts computers. I think most people can answer that themselves. So i wonder where something like a 490-1300 GTX stands against it. It also makes it pretty clear that Nvidia, if they want to, can release an Ultra and take the performance crown in terms of raw FPS anytime they want, and problably will do shortly after its release. This is where im hoping ATI pulls through with a bunch of unique features. And wheres the Doom 3 benchmarks? Not that i care about the game but with Quake4 on the horizon i was hoping for some nice improvments on that engine. ;)

It also would appear ATI is at work again with some AA tricks. I do love the job they do with AA i must say.

typos ):

Skrying

03-Oct-2005, 02:40

Because the sheets were not made when the CoD2 demo was out in the public? I think its smart, a lot of people wanted to know the performance advantage, even if that advantage doesnt stay that way forever, its a good marketing move now.

Particleman

03-Oct-2005, 02:40

I am so happy angle dependant AF is biting the dust. AF quality has really taken a back seat to performance the last few years, I'm glad it is finally going to go back to the way it was during the 4600/5900 days. When angle dependant AF was first introduced it didn't bug me that much, but after spending years with it, you really begin to see it's weaknesses, that and all the other crazy optimizations ATi and nVidia have done to compromise AF quality over the years. This will no doubt cause nVidia to follow suit and dump angle dependant AF soon too (if it is not hardwired on the 6800/7800, their next gen card for sure). Here's to hoping we see the unoptimization of AF.

Geo

03-Oct-2005, 02:44

With 2x speed improvement for 6xaa, adaptive aa, and improved texturing quality, Tinkerbell has been moved from intensive care to a semi-private room in "guarded" condition.

But, in the long-honored tradition of graphics enthusiasts everywhere, I shall now blithely stick all of that in my pocket and commence to bitching. :razz:

2x the speed at 6xaa and 512mb of memory. . .why the bloody H are we still at 6xaa?? :evil:

Rur0ni

03-Oct-2005, 02:44

:grin:

Considering the X1800 has OC headroom, I think its going to be quite, quite sexy. Now i'm almost done "anticipating" r520, time for r580... ;)

Nice for the FEAR game though, thats just good.

Arty

03-Oct-2005, 02:45

FEAR is ATI backed, think of it as their DOOM3.
I stopped reading there. FEAR is faster on Nvidia cards ~20%.

http://firingsquad.com/hardware/fear_beta_performance/page4.asp

ChrisRay

03-Oct-2005, 02:47

FEAR is also title that I believe will have SLI support out of the box.

Arty

03-Oct-2005, 02:50

FEAR is also title that I believe will have SLI support out of the box.
And CoD2. (Dual VPU support)

Bruce

03-Oct-2005, 02:51

I stopped reading there. FEAR is faster on Nvidia cards ~20%.

http://firingsquad.com/hardware/fear_beta_performance/page4.asp

huh? on old hardware from ATI, yes?

IgnorancePersonified

03-Oct-2005, 02:52

Power point slides take how long to fudge up?

SugarCoat

03-Oct-2005, 02:53

I stopped reading there. FEAR is faster on Nvidia cards ~20%.

http://firingsquad.com/hardware/fear_beta_performance/page4.asp

i have no doubt its faster on a GTX when compared to ATI's current cards, i'v known that for awhile. My point is that they're using high performance drivers for the show, and since Nvidia is a sponsor of Monolith it really doesnt come as a surprise they're bosting such a huge difference comparing a GTX to an X1800. If they found that they had that, they may as well use it. It is something to be proud about considering how demanding the demos have been on hardware. I said nothing of the R4XX line of cards so i dont see why thats an issue.

Chalnoth

03-Oct-2005, 02:55

Power point slides take how long to fudge up?
True, and there is a mistake on one of the X1600 slides, but we'll all know in a couple of days.

Arty

03-Oct-2005, 02:57

i have no doubt its faster on a GTX when compared to ATI's current cards, i'v known that for awhile. My point is that they're using high performance drivers for the show, and since ATI is a sponsor of Monolith it really doesnt come as a surprise they're bosting such a huge difference comparing a GTX to an X1800. I said nothing of the R4XX line of cards so i dont see why thats an issue.
6800 Ultra is ~20% faster than a X850 XT PE, no need to even compare 7800GTX.

It looks like you are nitpicking everything, even the AA images. From the looks of it, they are PR slides so there will be something to nitpick, duh !

Headstone

03-Oct-2005, 02:58

Did I see that correctly? I thought that the x1300 series was not going to support Crossfire. But according to the sell sheets it is...

All in all it looks like we could have some winning cards here. I would have preferred to see the x1600 matched up against the 6800 vanilla rather than the 6600gt

SugarCoat

03-Oct-2005, 02:59

6800 Ultra is ~20% faster than a X850 XT PE, no need to even compare 7800GTX.

It looks like you are nitpicking everything, even the AA images. From the looks of it, they are PR slides so there will be something to nitpick, duh !

i dont understand what you're having a problem with? We're not talking about ATIs current cards...why are you bringing it here? Also im a tad bit tired and not typing too well, the game is actually a WIMTBP game, so oops to Nvidia thats kind of embarassing.

Sxotty

03-Oct-2005, 03:00

What I love
features a shader model 3 architecture done right...playable for the first time in a mainstream core...first mainstream architecture to have ...(FP32 bit processing) enabling new levels for CPU and GPU communication, and brings a new level of high detail image quality

It is funny to me anyway...

Geo

03-Oct-2005, 03:00

As a side note, I find it odd that its been on the web for well in excess of an hour now and nobody has yet linked to it (and it partially one of the reasons why I'm a little testy right now)...

Not to mention that you're up at 2am cramming for oh, something or other, in a couple days, having been handed what is probably an impossible assignment to get done by the deadlne?

Yeah, I'd be cranky too. :wink:

Richard

03-Oct-2005, 03:02

FEAR is ATI backed...

FEAR is actually a TWIMTBP game.

ChrisRay

03-Oct-2005, 03:03

He was talking about those FEAR benchmarks that he linked too. And the 6800U does appear to be faster than the X850XT PE in that game.

SugarCoat

03-Oct-2005, 03:03

He was talking about those FEAR benchmarks that he linked too. And the 6800U does appear to be faster than the X850XT PE in that game.

i know but i never questioned that...

Moloch

03-Oct-2005, 03:07

What is fear doing that makes it run better on nvidia cards?

IgnorancePersonified

03-Oct-2005, 03:09

yeh yeh I just hate having to apply the bullshit filter to so much in the technology world. Have to update that filter as much as my spam and virus filters :roll:

Acert93

03-Oct-2005, 03:09

http://www.techpowerup.com/reviews/ATI/R5XX/

Has this been posted? If that is right, then the X1300 Pro is one sick card, especially at $149.

Nice to see the X1x00 series get FP10 & FP16/10 w/ AA :grin: Toss in Avivo and that is one nice card for the price. The performance slide (http://www.techpowerup.com/reviews/ATI/R5XX/images/X1300-1.jpg) notes that 3DMark05 score of 2700, I would assume that was for the faster of the cards (X1300/X1300Pro), but seems odd considering the past leak and the fact it has more VS units than the 6600. Kind of annoying that they don't note the resolution of the game comparisons or quality levels... but what do you expect from an IHV comparison chart? Interesting that in D3 the 6600 beats out the X1300 Pro.

I kind of wonder what the street price will be. You can get a 6600 (300core/500 memory varietyies) for under $100. While the X1300 will have more features, and the Pro seems more performance, the X1300Pro at $150 is going to be going face to face with the 6600GT. It will be interesting to see how benchmarks compare (especially without AA and at lower resolutions because I find the low end cards tend to get played with minimal features most of the time because they are under powered). If they are close the FP10 and Avivo could make it a very strong product for gamers.

This is, of course, assuming that is accurate.

tEd

03-Oct-2005, 03:10

What is fear doing that makes it run better on nvidia cards?

i would say mainly the stencil shadows

Skrying

03-Oct-2005, 03:13

At $150 it'll be the X1600Pro going at the 6600GT, and if you deem the extra 128MB of memory, the X1300Pro. Weird range of cards at the price. These prices setup a lot of choosing based on memory amounts.

CMAN

03-Oct-2005, 03:15

ATI thinks the 6600GT can due Transparency AA :lol:

They must be trying to help Nvidia there. I wonder what else is false on them... :twisted:

J/k

IgnorancePersonified

03-Oct-2005, 03:15

are those slides for real? there are quite a few mistakes in grammar and they swap between chip core names and retail card names not to mention the spelling mistakes :?:

Acert93

03-Oct-2005, 03:16

Heh, of course I am blind and did not notice techpowerup.com ALSO has the X1600 and X1800 stuff up :shock:

And in the time it took me to write my comments 1.5 pages of NEW POSTS occured. And now between my last post and this... how many?

This thread is on fire...

epicstruggle

03-Oct-2005, 03:20

OMG!!

I think this time ati invested in the right things.
Lol, it looks like investing in ati while they were down a few months ago was a good idea. Now if they would only make sure their supply line is in order. No use having a badass product that is impossible to purchase.

epic

Acert93

03-Oct-2005, 03:21

I think this is basically what we should all have expected: it's on par, no great breakthroughs in performance for the price. Maybe not performance (although FP10 down the road with Xbox 360 porting is interesting) BUT the X1300's features are nice, specifically Avivo.

The X1300 seems to strike a nice balance between Performance:Features. As someone who bought NV40 based partially on some of the NV "Pure Video" hype I see the X1300 as a compelling budget gaming/media card. Especially if it hits the $125 street price you mention.

CMAN

03-Oct-2005, 03:22

OMG!!

Lol, it looks like investing in ati while they were down a few months ago was a good idea. Now if they would only make sure their supply line is in order. No use having a badass product that is impossible to purchase.

epic

All of that one marketing slides? Don't get me wrong, I hope they have a great product. However, basing one's opinion solely on marketing slides is wrong. Wait till the reviews come out...only a couple more days. (I do think my next card will be an X1800 though :grin: )