So is R580 supposed to be 48 pipes or 48 ALUs?

[MY80S]

I was just entertaining the chance that R580 might actually be clocked below 625Mhz (the shader ops figures make sense if we assume 600Mhz for R580). However, R520's numbers don't make much sense for a marketing slide, as they indicate ~5.75 ops/pixel shader unit.

As for the gflops figures, one may assume that they're also counting non-ps flops.

[Ailuros]

I had thought that the ROP clock was synced with the memory clock. Could be wrong, of course.

Not the memory but the quads. I initially thought so too, until Xmas corrected me.

[Ailuros]

Can you play with just the ROP clocks in G70? I can't remember.

It would be interesting to see the effect of a 20% cut in ROP clocks versus a 20% cut in memory clock.

Indeed you could play with all these variables for ages :lol:

Jawed

No I can't. What you're thinking is probably something in the 490/550MHz ballpark. Haven't underclocked the ram yet (does it even make sense with the latency differences?) but have a look at those. Fear again 1600/4xAA/16xAF:

430/600MHz : 24 fps
490/600MHz: 26 fps
430/685MHz: 26 fps
490/685MHz: 28fps

If I go down to 550MHz and get something very close to 22-23 fps it would tell you what exactly?

[overclocked]

the chip should be around 380million transistors if thats the case (18%). Then shrink to 80nm, and voila R600 at the same size as this bady at 500+ million transistors at 80nm. yum....

Edit in reply to Pete.

[Pete]

I was just entertaining the chance that R580 might actually be clocked below 625Mhz (the shader ops figures make sense if we assume 600Mhz for R580). However, R520's numbers don't make much sense for a marketing slide, as they indicate ~5.75 ops/pixel shader unit.Orton himself has said R580 will be clocked higher than R520, and the current rumors peg X1900 XTX at 650MHz. But I'll grant that I'm guessing, too.

I'm not sure how you arrived at ~5.75. Did you assume a 625MHz core clock? That yields 60G / (16 * 625M) = 6 ops/PS. That conforms to this pic (http://pc.watch.impress.co.jp/docs/2005/1006/kaigai03.pdf), which is (16 * 5)PS instructions + (16 * 1)texture instructions. Rearrange that for R580 into (48 * 5)PS + (16*1)TMU, and we get a 648MHz clock for R580. That lines up perfectly with current XTX rumors.

Huh, I backed into getting those figures. Understanding still eludes me.

As for the gflops figures, one may assume that they're also counting non-ps flops.Yep, I should've figured the VS units, but assuming roughly similar VS FLOPs but 3x PS FLOPS, that would mean they contribute to half of R520's overall FLOPS. Seems high to this GPU illiterate. Again, no clue about FLOP derivation. 272.5GFLOPS / 625MHz = 436FLOPs/ck.

VS: 8 * ([4 + 1] * 2) = 80
PS: 16 * (4 + 4 + 1) = 144
TMU: 16 * 1 = 16

That's just 240. I'm missing 32(.5). Halp?

[Jawed]

If I go down to 550MHz and get something very close to 22-23 fps it would tell you what exactly?
I was curious to see the effect on the performance of no-AA/bilinear - rather than 4xAA/16xAF - to see if there's any bandwidth or backbuffer-bandwidth limiting.

With the proviso, of course, that it's not also CPU-bound - doubtful, I know - but as far as I can tell care is required with FEAR to ensure it's not CPU bound as the maximum eye-candy options chew a lot of CPU power.

In other words it would be nice to "prove" that FEAR is entirely ALU bound when doing no-AA/bilinear instead of assuming it.

I wonder if Dave is going to introduce FEAR as a benchmark, since Lost Coast doesn't suit.

Jawed

[Unknown Soldier]

Saw this.. thought it interesting.

Microsoft recommends ATI hardware for Vista

WE MANAGED to see a rather interesting Microsoft document that recommends to people who want to show off Vista to use ATI over Nvidia. It says that: "ATI is preferred to Nvidia at this time due to superior driver support". It turns out that the real Vista Beta 2 is still a few weeks away and that the beta 2 we saw was the December build, a beta 2 but not the complete one.

Read More: TheInq (http://www.theinquirer.net/?article=28825)

They say it's because of the "superior Driver Support" but I have to wonder if it's not because of the unified R6xx engine.

US

[Dave Baumann]

No, nothing to do with that. ATI are just a little further on with their software suport at the moment.

[MY80S]

Orton himself has said R580 will be clocked higher than R520, and the current rumors peg X1900 XTX at 650MHz. But I'll grant that I'm guessing, too.

I'm not sure how you arrived at ~5.75. Did you assume a 625MHz core clock? That yields 60G / (16 * 625M) = 6 ops/PS. That conforms to this pic (http://pc.watch.impress.co.jp/docs/2005/1006/kaigai03.pdf), which is (16 * 5)PS instructions + (16 * 1)texture instructions. Rearrange that for R580 into (48 * 5)PS + (16*1)TMU, and we get a 648MHz clock for R580. That lines up perfectly with current XTX rumors.

Huh, I backed into getting those figures. Understanding still eludes me.

Yep, I should've figured the VS units, but assuming roughly similar VS FLOPs but 3x PS FLOPS, that would mean they contribute to half of R520's overall FLOPS. Seems high to this GPU illiterate. Again, no clue about FLOP derivation. 272.5GFLOPS / 625MHz = 436FLOPs/ck.

VS: 8 * ([4 + 1] * 2) = 80
PS: 16 * (4 + 4 + 1) = 144
TMU: 16 * 1 = 16

That's just 240. I'm missing 32(.5). Halp?


Pardon me, my calculator's been acting up lately :(

[Ailuros]

I was curious to see the effect on the performance of no-AA/bilinear - rather than 4xAA/16xAF - to see if there's any bandwidth or backbuffer-bandwidth limiting.

With the proviso, of course, that it's not also CPU-bound - doubtful, I know - but as far as I can tell care is required with FEAR to ensure it's not CPU bound as the maximum eye-candy options chew a lot of CPU power.

In other words it would be nice to "prove" that FEAR is entirely ALU bound when doing no-AA/bilinear instead of assuming it.

I wonder if Dave is going to introduce FEAR as a benchmark, since Lost Coast doesn't suit.

Jawed


Fear, 2048*1536 1xAA/1xAF bilinear:

430/600MHz: 34 fps
490/600MHz: 37 fps
430/685MHz: 35 fps
490/685MHz: 39 fps

490/550MHz: 36 fps

[Jawed]

Gee, thanks Ailuros, I wasn't actually expecting that you'd put the extra effort in - as it seems we're looking at diminishing returns on the possible conclusions.

So, the extra bandwidth, is it helping the TMUs or the ROPs?...

24.5% extra bandwidth (550 to 685) translates into 8.3% extra performance. Hmm. I suppose it should be mostly the ROPs because the blurriness of practically every texture in the scene at such a high resolution should mean that the TMUs are working mostly from cache.

Looking at the two pairs of 430/490 results (at 600 and 685MHz memory), the 14% FLOPs increase produces FPS increases of:

11.4% @ 685MHz
8.8% @ 600MHz14% extra FLOPs combined with 14% extra bandwidth generates the 11.4% performance improvement. Hmm... I suppose another way of putting it might be that at 770MHz memory the 14% extra FLOPs would produce 14% extra FPS :?:

How does the shadowing algorithm work in FEAR? It looks like stencil shadow volumes, but is it?... Is there a chance that the shadowing is bottlenecking framerates, at least partially?

It looks like both R580 and G71 will have to make do with the same amount of bandwidth as R520 and GTX-512. And from your results I guess that means that the unchanged bandwidth is going to cut performance gains at no-AA/no-AF in half compared with the theoretical FLOPs increases.

:lol: @ the rounding errors and a general feeling of ridiculousness in writing this... Well it's a silly hour and I'll prolly wake up thinking what a load of shit...

Jawed

[Ailuros]

You've confused me now...

14% extra FLOPs combined with 14% extra bandwidth generates the 11.4% performance improvement. Hmm... I suppose another way of putting it might be that at 770MHz memory the 14% extra FLOPs would produce 14% extra FPS

~14% higher clock and ram frequencies translates according to the results above into ~15% higher performance compared to default frequencies (34 vs. 39 fps).

How does the shadowing algorithm work in FEAR? It looks like stencil shadow volumes, but is it?... Is there a chance that the shadowing is bottlenecking framerates, at least partially?

Shadowing is bottlenecking AA performance even more. Less with noAA/AF but it affects performance quite a bit there too. Yes it does look like stencil shadows but I'm not 100% sure either.

And from your results I guess that means that the unchanged bandwidth is going to cut performance gains at no-AA/no-AF in half compared with the theoretical FLOPs increases.

I'd speculate rather on something in the </=40% ballpark. I'm obviously assuming something like 600(*32)/900MHz which gives a ~3:1 ratio between bandwidth and fillrate and hence the 490/550MHz experiment. Those results are of course completely unsafe to speculate on, but I obviously had spare time on my hands.

[Pete]

Pardon me, my calculator's been acting up lately :(There's no need for that. I thought you knew more than me, so I ran those calcs a few times. The upside is that I'm beginning to understand FLOPS and mere ops, so I should probably thank you for prodding my brain. :)

[IbaneZ]

11692 in 3dmark-05 with an overclocked FX-55?
http://www.pcpop.com/doc/0/125/125255_3.shtml

[Ailuros]

11692 in 3dmark-05 with an overclocked FX-55?
http://www.pcpop.com/doc/0/125/125255_3.shtml

Sounds reasonable. 3dmark05 appears extremely dated though these days.

[Jawed]

You've confused me now...
It was past my bedtime, sigh.

~14% higher clock and ram frequencies translates according to the results above into ~15% higher performance compared to default frequencies (34 vs. 39 fps).
Yes :oops: :shock: :oops: I seemed to work through every combination except that one.

So, anyway, there's little doubt that FEAR is using all the bandwidth it can and extra ALU power is wasted if the bandwidth doesn't go up. At least on NVidia 7 series. X1800XT doesn't seem to have quite enough ALU power to use up all its bandwidth - so there's a bit of breathing room.

Maybe Oblivion will provide a more ALU-limited benchmark...

I'd speculate rather on something in the </=40% ballpark. I'm obviously assuming something like 600(*32)/900MHz which gives a ~3:1 ratio between bandwidth and fillrate and hence the 490/550MHz experiment.
You'll need to explain in longhand what you're doing there. 32 ROPs at 600 MHz * 4 bytes of colour per pixel?

Those results are of course completely unsafe to speculate on, but I obviously had spare time on my hands.
I feel reasonably happy, now, concluding that FEAR isn't ALU-bound to the extent that'll make it the flag bearer for R580.

Jawed

[Jawed]

11692 in 3dmark-05 with an overclocked FX-55?
http://www.pcpop.com/doc/0/125/125255_3.shtml
ATI's R520 marketing material, based as heavily on 3DMk scores as it was, was so misleading that I think it even mislead ATI into thinking R520 was great.

Jawed

[Ailuros]

You'll need to explain in longhand what you're doing there. 32 ROPs at 600 MHz * 4 bytes of colour per pixel?

I estimated 600MHz core (8 quads) and 900MHz GDDR3 for a hypothetical G71. That much bandwidth for that much fillrate and thus the 490MHz/6quad, 550MHz (ram) experiment.

Where do you get the 32 ROPs from anyway? I have severe doubts that G71 will have more than 16.

ATI's R520 marketing material, based as heavily on 3DMk scores as it was, was so misleading that I think it even mislead ATI into thinking R520 was great.

Uhmmm the PS/VS results are still interesting and verify IMHO that it actually has 10 VS units. Other than that it indicates 16% higher performance in the PS test compared to a GTX 512, one of the reasons why I think 3dmark05 is dated for those kind of GPUs.

[_xxx_]

OT:
geo, your sig found its way to MSI, it seems ;)

http://www.hardwarezone.com/articles/view.php?cid=1&id=1792

EDIT: Just started a new topic on this, I think it might be worth a deeper look

[Jawed]

I estimated 600MHz core (8 quads) and 900MHz GDDR3 for a hypothetical G71. That much bandwidth for that much fillrate and thus the 490MHz/6quad, 550MHz (ram) experiment.

Where do you get the 32 ROPs from anyway? I have severe doubts that G71 will have more than 16.
Ah, when you said "fillrate" I thought you meant ROP-based fillrate. Not shader rate.

That makes more sense now.

The 490/550 result seems badly bandwidth limited, so G71 is going to suffer the same fate - even before AA/AF are turned on. SLI seems like the better solution for FEAR.

Jawed

[Moloch]

ATI's R520 marketing material, based as heavily on 3DMk scores as it was, was so misleading that I think it even mislead ATI into thinking R520 was great.

Jawed
You're a funny one :lol:

[no-X]

11692 in 3dmark-05 with an overclocked FX-55?
http://www.pcpop.com/doc/0/125/125255_3.shtml


http://img364.imageshack.us/img364/3292/r5805gy.png

note the strange VS result...

based on:
R580
http://www.pcpop.com/doc/0/125/125255_3.shtml

R520
http://www.xbitlabs.com/articles/video/display/geforce7800gtx512_29.html
http://www.xbitlabs.com/articles/video/display/geforce7800gtx512_30.html
http://www.xbitlabs.com/articles/video/display/radeon-x1000_22.html
http://www.xbitlabs.com/articles/video/display/radeon-x1000_23.html

[Karma Police]

That complex Vertex Shaders get a nominal performance boost while simple Vertex Shaders get a very large performance boost?

[ERK]

So, anyway, there's little doubt that FEAR is using all the bandwidth it can and extra ALU power is wasted if the bandwidth doesn't go up. At least on NVidia 7 series. X1800XT doesn't seem to have quite enough ALU power to use up all its bandwidth - so there's a bit of breathing room.
...
I feel reasonably happy, now, concluding that FEAR isn't ALU-bound to the extent that'll make it the flag bearer for R580.
Ummm... Didn't you just prove the opposite?
I mean, increasing the core frequency by 14% (430->490) increased the FPS by 9-11%, while increasing the mem clock 14% (600->685) only increased the FPS by 3-5%. Should I not then conclude that FEAR is 2-3 times more ALU limited than it is memory limited (no AF/AA), since it benefits more from core clock increase?

By the same method, it seems that with AF/AA the memory clock and core clock proportional changes make equal improvements (both 24->26 FPS), and thus represent a 'balanced' limitation? (Seems like a good design, from this POV to me.)

Or were you just talking about the AA/AF results?

[Jawed]

All Ailuros and I have been doing is looking at no-AA/AF results to try to gauge the degree of ALU limitation in FEAR.

I agree that FEAR is generally more ALU limited than bandwidth limited - but it also seems that X1800XT and 7800GTX are operating close to their bandwidth ceiling.

Jawed

[ERK]

I hope so or they are inefficent designs.

[Geo]

Hexus on a rumor-mongering roll today: http://www.hexus.net/content/item.php?item=4346

I'm still not believing that XTX clock. I refuse to believe until reviews are posted. Nyah nyah!

:wink:

[tEd]

1 week

[_xxx_]

I'm still not believing that XTX clock.

Why, what do you expect? It's already very high IMHO.

[AlphaWolf]

Why, what do you expect? It's already very high IMHO.

The 3% XTX clock boost over the XT doesn't seem to warrant a new SKU.

[_xxx_]

The 3% XTX clock boost over the XT doesn't seem to warrant a new SKU.

Any additional X or PE or whatever extra letters seem to be enough nowadays, for both vendors ;)

[N00b]

ATi RADEON X1900 XTX 512MB (650MHz VPU / 1550MHz GDDR-3) (AMD A64 FX-57 2.8GHz)
11,149 3DMarks

ATi RADEON X1900 XT 512MB (625MHz VPU / 1450MHz GDDR-3) (AMD A64 FX-57 2.8GHz)
~10,000 3DMarks1100 3DMarks for a 25Mhz VPU / 100MHz GDDR-3 increase? That's rather unlikely.

[AlphaWolf]

1100 3DMarks for a 25Mhz VPU / 100MHz GDDR-3 increase? That's rather unlikely.

Well I'd guess that the 10,000 is low for the XT. Thats ~10% over the x1800xt, I'd expect a bit more of a gap in 3dmark05, but I don't know.

[N00b]

Well I'd guess that the 10,000 is low for the XT. Thats ~10% over the x1800xt, I'd expect a bit more of a gap in 3dmark05, but I don't know.I lack data here, so I won't speculate.
Does anybody know how much more a 7800 scores over a 6800 in 3DMark05 @ same clockspeeds / pipes? That could give an indication what to expect.

[_xxx_]

I lack data here, so I won't speculate.
Does anybody know how much more a 7800 scores over a 6800 in 3DMark05 @ same clockspeeds / pipes? That could give an indication what to expect.

I don't think so since the architectures are different.

[N00b]

I don't think so since the architectures are different.I disagree. Since the primary improvement of the G70 over the NV40 is the added ALU, this might give an indication how 3DMark05 responds to more ALU power. I'm not saying that performance difference between NV40 and G70 can be applied to R520 and R580, but it could give a hint and make an educated guess possible.

[_xxx_]

Dunno. NV40 had "1.5" ALU's per pipe, G70 has two. R580 has 3 with unknown capabilities etc., hard to make any parallels there IMHO.

[Dave Baumann]

R580 has 3 with unknown capabilities etc.
Unknown? If you can't figure it it then you're not very good at this game!

[Ailuros]

Dunno. NV40 had "1.5" ALU's per pipe, G70 has two. R580 has 3 with unknown capabilities etc., hard to make any parallels there IMHO.

I notice the brackets, but there's no half of anything unit. Let's just say that all recent GPUs have one ALU per SIMD channel that's split up into two main subunits, with such and such capabilities; if that should help to make things easier.

[N00b]

Dunno. NV40 had "1.5" ALU's per pipe, G70 has two. R580 has 3 with unknown capabilities etc., hard to make any parallels there IMHO.That's where the "educated" part of educated guess comes in. ;-)

I was thinking along the following lines:
- I know how 1800XT scores in 3DMark05. (a)
- I know how 7800GTX 512 scores in 3DMark05. (b)
- I believe that 1900XT = ~ 1800XT + 2 * ALU + some other improvements (c)
- I assume that 3DMark05(1900XT) >= 1.1 * 3DMark05(7800GTX512) (d)
- x = (3DMark05(G70) - 3DMark05(NV40)) / 3DMark05(NV40) (@ same clockspeeds / pipes)

So I assume that
3DMark05(1900XT) = y * 3DMark05(1800XT)
and
y = 1 + z + eg * x
with
z = gain from architectural improvements other than added ALUs
eg = educated guess factor (between 1.0 and 2.0)

(a), (b) and (d) make me assume: z + eg * x >= 0.2

Now let's speculate:
- I speculate that z is in the range 0.0 - 0.1
- I speculate that eg is in the range 0.1 - 0.25

If I had x I could verify some of my assumptions and speculations.

[_xxx_]

I know there's no half of anything, but that's the best approximation I could think of :)

EDIT:
N00b, you just gave me headache :lol:

[N00b]

N00b, you just gave me headache :lol:Sorry! :oops:

[chavvdarrr]

http://www.pcpop.com/doc/App/125255/000219387.html
10600 in 3dmark05
http://www.pcpop.com/doc/App/125255/000219571.html
X1900XTX - 650/750
7900GTX - 750/950 , 8 quads , 24 rops?

[Ailuros]

Why would the G71 have 24 ROPs?

***edit: not that 8 quads@750MHz makes that much more sense either :roll:

[N00b]

http://www.pcpop.com/doc/App/125255/000219387.html
10600 in 3dmark05
http://www.pcpop.com/doc/App/125255/000219571.html
X1900XTX - 650/750
7900GTX - 750/950 , 8 quads , 24 rops?
10600 3DMarks05 @ 1280x1024 (http://www.pcpop.com/doc/App/125255/000219387.html)
11600 3DMarks05 @ 1024x768 (http://www.pcpop.com/doc/App/125255/000219388.html)

[Geo]

1 week

So you're planting a flag for the 17th instead of the 24th?

[chavvdarrr]

Why would the G71 have 24 ROPs?

***edit: not that 8 quads@750MHz makes that much more sense either :roll:
dunno, in that table there is something thats 16 for X1900 and 24 for 7900 .... my wild guess is thats ROPs, could be anything else... like VS .... :D

[Fodder]

dunno, in that table there is something thats 16 for X1900 and 24 for 7900 .... my wild guess is thats ROPs, could be anything else... like VS .... :D1. PS/TMU, 2. ROPs or ???, 3. VS, 4. ROPs or ???.

[fellix]

Could be this (http://img313.imageshack.us/img313/5908/r5801mt.jpg) the R580 general block layout, as if it's a direct RV530 structure scale-up?!

[Sunday]

dunno, in that table there is something thats 16 for X1900 and 24 for 7900 .... my wild guess is thats ROPs, could be anything else... like VS .... :D
G71 will only be a speed bum up over G70! Thanx to the 90nm tech, this GPU will go around 700 MHz, and in that case, 24 Pixel Pipes (6 quads) will be sufficient! There is no need to go for 32 pipes and larger, more expensive, and ultimately slower core!!

[Blacklash]

Any guess on how the X1900XL will compare to the X1800XT? About the same in performance?

[Dave Baumann]

Could be this (http://img300.imageshack.us/img300/4562/r5803ar.jpg) the R580 general block layout, as if it's a direct RV530 structure scale-up?!
RV530's Shader "pipes" are procesing parallel pixels - i.e. the depth of each "pipe" is unchanged, but the parallelism is increased.

[tEd]

So you're planting a flag for the 17th instead of the 24th?

Depends where you live. Maybe a day earlier. Just rumors

[chavvdarrr]

G71 will only be a speed bum up over G70! Thanx to the 90nm tech, this GPU will go around 700 MHz, and in that case, 24 Pixel Pipes (6 quads) will be sufficient! There is no need to go for 32 pipes and larger, more expensive, and ultimately slower core!!
maybe. but in the link at hkepc it clearly states 8 quads & 32 pipes
And there is something ELSE which is 24...

I don't know chinese ... but maybe they mean 32 PS pipes, 24 TMUs, 16 Rops ?!

[Pressure]

Why would the G71 have 24 ROPs?

***edit: not that 8 quads@750MHz makes that much more sense either :roll:

From a resolution point of view, there is really no reason to have more than 16 Raster Operation Units at this time, if I recall correctly.

[fellix]

Anyone got a hint, whatever the G71 ROPs would support more than single loop for MSAA sampling?!

[_xxx_]

Unknown? If you can't figure it it then you're not very good at this game!

Nah, I just left the door open for some potential surprises, that's all :)

[pjbliverpool]

Apologies if this has already been posted:

http://img225.imageshack.us/img225/8917/super9in.jpg

It was posted on a nother forum so I don't know the source or if its real or not however if ti is real the numbers would seem to suggest that an R520 pixel shader does not equal an R580 pixel shader.

EDIT: Actually I think I was adding up wrong, the difference in pixel shader ops would come from the fact that R580 only has 16 texture address ALU's. Running at 650Mhz would then give the correct figure of around 166 ops/sec. Still not sure how those GFLOP counts work though.

[Unknown Soldier]

A question. Since ATI announced that their chips can do GPUPhysics, when will we be seeing the benefits of this? Also, will the G71 also be doing any Physics when it gets released?

US

[Chalnoth]

A question. Since ATI announced that their chips can do GPUPhysics, when will we be seeing the benefits of this? Also, will the G71 also be doing any Physics when it gets released?
Well, any DX9 GPU should be able to accelerate physics. The question is only how efficient the GPU's are at doing it. Physics calculations, however, are going to be very reliant upon FP32 inputs and outputs, as well as dynamic branching. So I would tend to suspect that ATI hardware would do a better job at it, for the time being.

I don't expect the G71 to offer anything new to the table over the G70 except higher performance. I do, however, expect nVidia to have a brand new architecture ready for debut in concert with Windows Vista.

[Pete]

Apologies if this has already been posted:

http://img225.imageshack.us/img225/8917/super9in.jpgMY80S and I (http://www.beyond3d.com/forum/showpost.php?p=668259&postcount=505) quibbled over it on pages 20-21. I had the same trouble with the GFLOPS # as you.

[Pete]

A question. Since ATI announced that their chips can do GPUPhysics, when will we be seeing the benefits of this?
Hmmm, the Oblivion interview posted today notes that Bethesda opted for HDR via FP buffers rather than pixel shaders b/c the pixel shaders were already saturated. Doesn't seem to leave much room for physics, but dual-core CPUs seem ripe for exploration.

[Fodder]

Doesn't seem to leave much room for physicsI believe this where you slap an X1600 next to your X1900.

[Apple740]

Any guess on how the X1900XL will compare to the X1800XT? About the same in performance?

Afaik there won't be a X1900XL. X1900XT/XTX will be Ati's ultra hi-end, current X1800XL/XT will be continued.
X1600 tough dissapears and will be replaced by X1700.

[Dave Baumann]

X1600 won't disappear even after the release of RV560 since RV560 is in a different bracket. More likely to happen is X1600 will be silently refreshed with with an 80nm version and drop down in price.

[Apple740]

X1600 won't disappear even after the release of RV560 since RV560 is in a different bracket. More likely to happen is X1600 will be silently refreshed with with an 80nm version and drop down in price.

Ah well, i was basing my info from what i've read from Gibbo's (oc.uk) mouth. :)

http://forums.overclockers.co.uk/showpost.php?p=6225648&postcount=13

[Fodder]

The only difference between Fudo and Gibbo is Fudo doesn't realise when he's making things up.

[Dave Baumann]

Ah well, i was basing my info from what i've read from Gibbo's (oc.uk) mouth. :)
Yeah, well...

[Geo]

X1600 won't disappear even after the release of RV560 since RV560 is in a different bracket. More likely to happen is X1600 will be silently refreshed with with an 80nm version and drop down in price.

And then we'll see it under eleventy different names for the next several years? :lol:

Really, that'd be a pretty good bump up of the baseline.

[kemosabe]

RV530 has started racking up some OEM design wins in both desktop and mobile, so it would indeed seem to be around for awhile. With that tiny die (which will shrink further at 80nm), this looks to be ATI's next 9550 for the high-volume mainstream segment. But really that off-the-bat, next-generation midrange retail killer has been elusive for years as far as ATI is concerned - it's time they got it right with RV560.

[BrynS]

AnandTech mention (http://www.anandtech.com/tradeshows/showdoc.aspx?i=2669&p=9) that Sapphire had R580 boards at CES and suggest product availability at launch. If I were ATI's new logistics tsar, this would be the product that I made sure all the trains ran on time for, as geo is wont to say! :grin:

Presumably other AIB's had R580 boards at the show, although I have yet see any other reporting on this. Anyone hear anything on whether G71 samples made a veiled appeareance at CES?

[Mariner]

The only difference between Fudo and Gibbo is Fudo doesn't realise when he's making things up.

And to be fair to Fudo, he isn't trying to sell you anything. :wink:

[Geo]

AnandTech mention (http://www.anandtech.com/tradeshows/showdoc.aspx?i=2669&p=9) that Sapphire had R580 boards at CES and suggest product availability at launch. If I were ATI's new logistics tsar, this would be the product that I made sure all the trains ran on time for, as geo is wont to say! :grin:



It's a start! Being first out of the gate for a change, is prolly a pleasant thing. Having a chance to sell at high-margins until G71 is ready to go must have them licking their lips.

[_xxx_]

until G71 is ready to go

And that's still not quite clear when that will be.

[Geo]

Depends where you live. Maybe a day earlier. Just rumors

I'm thinking not.

http://www.hexus.net/content/item.php?item=4346


Chris said that we’d have to wait until later this week when ATi officials came over to London.


That would be very short lead, tho obviously the weekend would help a touch. That would be back in R520 territory for the screams of pain it would cause from our reviewer friends. And, y'know, it's a new year, and a new launch "done right" (at least that seems to be the subtext this time), so let's hope they don't cut the fuses that short. I mean, we're still minus quite a few X1300/X1600 reviews <kaff, kaff> because of the way they rushed the last one.

[Jawed]

Perhaps ATI's going for one of those launches where the cards appear on shelves before the reviews are out and the NDA's up?

Jawed

[Jawed]

As for X1300/X1600 reviews, perhaps Dave was waiting until he knew they were real, rather than X700XT-like phantoms...

Jawed

[Geo]

The other thing that occurred to me, re clocks and that otherwise inexplicable (to me at least) 625/650 thing. . .is maybe that 650 is about as indicative of what you'll actually be buying as G70's 430. . .

Not formally making that prediction, but it did occur to me. . .have we seen indications of AIB clocks?

[Sunday]

The other thing that occurred to me, re clocks and that otherwise inexplicable (to me at least) 625/650 thing. . .is maybe that 650 is about as indicative of what you'll actually be buying as G70's 430. . .

Not formally making that prediction, but it did occur to me. . .have we seen indications of AIB clocks?
AIB clocks WILL be same!!

[Geo]

Okay, so one vote for "undecided" it is. :lol:

[trinibwoy]

The other thing that occurred to me, re clocks and that otherwise inexplicable (to me at least) 625/650 thing. . .is maybe that 650 is about as indicative of what you'll actually be buying as G70's 430. . .

Not formally making that prediction, but it did occur to me. . .have we seen indications of AIB clocks?

geo, why are you so adamant that R580 should/will clock higher than 650? It looks like G71 is a smaller chip and most put it at around the 700 mark. Same process to boot.

[Geo]

geo, why are you so adamant that R580 should/will clock higher than 650? It looks like G71 is a smaller chip and most put it at around the 700 mark. Same process to boot.

Wrong context. I'm not adamant that XTX should clock higher than 650. I'm adamant that it should clock more than 4% higher than the "next down" part, XT. It's the delta between XT and XTX that I'm pointing at as cognitive dissonance. You think there is a signficant yield difference there?

[trinibwoy]

Wrong context. I'm not adamant that XTX should clock higher than 650. I'm adamant that it should clock more than 3% higher than the "next down" part, XT. It's the delta between XT and XTX that I'm pointing at as cognitive dissonance. You think there is a signficant yield difference there?

Ah ok, I was going off of your 650XTX ~ 430GTX comment above. I agree with you that 650 on the XTX is strange. Maybe it's really the 695 number that's been going around as well.

[Jawed]

X800XTPE was only 4% higher-clocked than the X800XT.

Jawed

[karlotta]

X800XTPE was only 4% higher-clocked than the X800XT.

Jawed But the x800xt came out a year later. Or do you mean the x800pro that was released at the same time ~ as the x800xtpe, Or am i just a confused buyer? Pro 475hz, PE 526hz , XT 500hx?
i changed the pro speed to 475 from 425. Opps i guess it was a few months later for the XT..?

[ANova]

I would bet ATI has something just in case the G71 proves to be faster.

[AlphaWolf]

I would bet ATI has something just in case the G71 proves to be faster.

X1900UTESE (unlikely to ever ship edition)

[ANova]

X1900UTESE (unlikely to ever ship edition)
Yes, sadly. XTX PE at 700 anyone?

[R300King!]

Don't forget the 7800GTX512PSE (publicity stunt edition) ;-)

[Moloch]

Don't forget the 7800GTX512PSE (publicity stunt edition) ;-)
:lol:
They should rename the 512MB GTX that.

[Fodder]

But the x800xt came out a year later. Or do you mean the x800pro that was released at the same time ~ as the x800xtpe, Or am i just a confused buyer?The original R420/423-based XT launched within a month or two of the Pro and XTPE.

[N00b]

:lol:
They should rename the 512MB GTX that.No. I don't think so. Unlike the PEs you could buy a 512 GTX at launch time. Plus it was available in europe.
I know because I own one. And no I did not preorder. I actually ordered mine two weeks after launch.
LE (= Limited Edition) would be more fitting.

[R300King!]

Yes they should, nahner, nahner. :-P

But seriously, if a card's not going to be available in any kind of quantity, it's a plubicity stunt plain and simple. Nvidia even said it's a limited card. That tells me it was thrown out there as a stunt to just barely top the X1800XT for the e-pen|s rights, nothing more.

[Unknown Soldier]

What I want to know is why the X1900XT doesn't have faster RAM. I mean the 7900GTX is said to have 900Mhz ram whereas the X1900XT only has 800Mhz.

US

[Unknown Soldier]

I would bet ATI has something just in case the G71 proves to be faster.

DDR4 ;)

[Apple740]

What I want to know is why the X1900XT doesn't have faster RAM. I mean the 7900GTX is said to have 900Mhz ram whereas the X1900XT only has 800Mhz.

US

Availability?

Maybe the 7900GTX will be as hard, or even harder to get than the 7800GTX/512, due to
the used mem. Maybe later on when 1.0ns mem is starting to ship in good quantity's we'll see
X1900XTXPE's :) with that.

[kemosabe]

Well since R580 will apparently go unchallenged for a month or two, it makes sense for ATI to try to reap the highest margins by keeping costs under control. It also makes sense for them to want to keep some ammunition to pair with those cherry-picked R580 cores when G71 hits the shelves. So will it just be a 700MHz R580 with 900MHz memory come CeBit, or is there something different to expect from those R590 rumours that were floating around? Anyone know if R590 is still on the roadmap and whether it's expected to be anything other than a 80nm die shrink?

[Pete]

I dunno, given the GTX 512's current unavailability, seems to me NV named it as such to raise the original GTX's profile (in addition to kicking ATI's ass in benchmarks).

The RAM speed question is interesting, b/c there are so many options. Are they using tighter timings than NV? Are they aiming for a lower power envelope? Or, worst of all, are they having trouble clocking the ring bus that high? The last doesn't make too much sense given that they've touted it as being forward-looking (which I and others have interpreted to mean it can scale with DDR4, as well as improving performance and die heat locality), but who knows how crafty their marketing department is.* ;)


* Like Google's motto being "Don't be evil." Is that grad student idealism or just the best capitalist hand-waving (look over there! /is evil while ppl are distracted) motto ever? :lol: (Yes, extreme cynicism on my part is the other, more obvious and likely plausible answer.)

[CJ]

Anyone know if R590 is still on the roadmap and whether it's expected to be anything other than a 80nm die shrink?

I believe I read somewhere that all process nodes after 90nm are using low-k. So if 80nm has low-k, why should R590 (if it's still on the roadmap...) just be 'anything other than a 80nm die shrink'? So could it be that it's not like R430 which was 'just a 110nm die shrink of R423 without low-k hence lower clockspeeds'? Especially if they put some effort into it (which I doubt they did with R430 on 110nm), it should be faster (just look at what nV did with 110nm and 300M trannies). And it could explain why it was listed as 'optional' on the roadmap I saw, as if they still had something extra up their sleeve.

[dizietsma]

The only difference between Fudo and Gibbo is Fudo doesn't realise when he's making things up.

I'm not sure who's more damned with that sentence.

I can't stop laughing though.

[Fodder]

And it could explain why it was listed as 'optional' on the roadmap I saw, as if they still had something extra up their sleeve.I read the 'optional' as 'if we have time before R600'.
I can't stop laughing though.:grin:

[_xxx_]

I would bet ATI has something just in case the G71 proves to be faster.

Well if we're about to speculate into the blue, I'd say no, they don't. Rather the opposite, nV might have the next step ready if so needed, since they had much time on their hands to do something.

[trinibwoy]

Well if we're about to speculate into the blue, I'd say no, they don't. Rather the opposite, nV might have the next step ready if so needed, since they had much time on their hands to do something.

This silence from Nvidia (and by extension Sony w/ RSX) is beginning to loom very ominously. Either they are keeping some very juicy secrets or they're having mucho problems with the 90nm transition.

[satein]

This silence from Nvidia (and by extension Sony w/ RSX) is beginning to loom very ominously. Either they are keeping some very juicy secrets or they're having mucho problems with the 90nm transition.
I seem to agree on that... at least I hope NV would be lucky on this 90nm process than that of ATi on R5xx part :wink:.

[Ailuros]

This silence from Nvidia (and by extension Sony w/ RSX) is beginning to loom very ominously. Either they are keeping some very juicy secrets or they're having mucho problems with the 90nm transition.

Maybe or maybe not; I don't see anything moving with inofficial driver revisions lately, which could maybe indicate something.

[Geo]

http://www.theinquirer.net/?article=28953

50% in FEAR? Will that lead dissappear after a few days like last time?

[BrynS]

Seems (http://forums.overclockers.co.uk/showpost.php?p=6264510&postcount=52) as though some e-tailors already have cards or will have very soon.

[ERK]

http://www.theinquirer.net/?article=28953

50% in FEAR? Will that lead dissappear after a few days like last time?
Most likely this is with no AA/AF.

I think the lead will hold. Keep in mind how the X1600XT is about twice as fast as an X1300Pro in that circumstance.
ERK
edit: link. (http://www.anandtech.com/video/showdoc.aspx?i=2575&p=4)

[Jawed]

X1300Pro has much slower memory and it doesn't have the fancy memory controller, nor the double-rate ROPs.

X1600XT is also about twice as fast with 4xAA/8xAF.

Jawed

[kemosabe]

So has the R580 Netherlands tour begun? Let's get the hype show on the road. :razz:

Oh, and why would RD580 and other newfangled Canadian stuff (http://www.pcisig.com/developers/compliance_program/integrators_list/pcie/) show up without R580?

[Pete]

This silence from Nvidia (and by extension Sony w/ RSX) is beginning to loom very ominously. Either they are keeping some very juicy secrets or they're having mucho problems with the 90nm transition.
Everything's always looming with you, isn't it? The End is looming, bad weather is looming, a vegemite sandwich is looming. Well, I'm tired of it!

:D

Oh, and some (looming) distributor pricing (http://forums.anandtech.com/messageview.aspx?catid=31&threadid=1778900&enterthread=y) to whet your whistles. Not too shabby.

[Moloch]

We must be getting close...

[volt]

...you'll be surprised how it's priced, especially the high-end boards :D

[Love_In_Rio]

Less than 30 fps X1900-dual crossfire in new 3dmark 2006 ?

http://www.bit-tech.net/news/2006/01/12/futuremark_3dmark_06/

It seems a very HDR intensive test.

[CJ]

So has the R580 Netherlands tour begun? Let's get the hype show on the road. :razz:


Yep, the R580 Tour hit Amsterdam yesterday. But several reviewers and other people who were invited didn't go, because they already had been able to testdrive the X1900XT for a week and the R580 Tour would only be a marketingpresentation of things that they already knew. The people who went there though are now officially under NDA so you won't get much out of them. ;)

[trinibwoy]

Everything's always looming with you, isn't it? The End is looming, bad weather is looming, a vegemite sandwich is looming. Well, I'm tired of it! :D

Oh, and some (looming) distributor pricing (http://forums.anandtech.com/messageview.aspx?catid=31&threadid=1778900&enterthread=y) to whet your whistles. Not too shabby.

Well if it's looming it's looming! I can't help that ;) Those prices are sweet, wonder if Nvidia will still bother with GTX512. It'll have to sit ~ $550 to be competitive now.

[BrynS]

Less than 30 fps X1900-dual crossfire in new 3dmark 2006 ?

http://www.bit-tech.net/news/2006/01/12/futuremark_3dmark_06/

It seems a very HDR intensive test. Too late -- the benchmark guardians got there first! :razz:

VR-Zone apparently (http://forums.vr-zone.com.sg/showthread.php?t=53164) have a 1900XT, which they've run through 3DMark05. The scores seem to be inline with the Hexus info and they managed to overclock their card from 625/725MHz to 715/855Mhz. The admin suggests 13K is attainable on a fully air cooled setup.

[_xxx_]

I seem to agree on that... at least I hope NV would be lucky on this 90nm process than that of ATi on R5xx part :wink:.

ATI's problem was not the process, but the failure in one of the libraries they used for designing the chip as I got it. So I don't really think nV will encounter such problems, but of course who knows for sure?

[satein]

ATI's problem was not the process, but the failure in one of the libraries they used for designing the chip as I got it. So I don't really think nV will encounter such problems, but of course who knows for sure?
Thank you for correct me :oops:, actually, I ment the 90nm process here not only the manufacturing process but also a transition, design or whatever they have to do in order to get products on this 90nm.

[R300King!]

Too late -- the benchmark guardians got there first! :razz:

VR-Zone apparently (http://forums.vr-zone.com.sg/showthread.php?t=53164) have a 1900XT, which they've run through 3DMark05. The scores seem to be inline with the Hexus info and they managed to overclock their card from 625/725MHz to 715/855Mhz. The admin suggests 13K is attainable on a fully air cooled setup.

So if the X1900XT can be clocked to these speeds pretty easily then what's the difference between the XT and XTX? Not much I'd say. People would just buy the XT only and OC. This doesn't make much sense on ATI's part to put 2 cards out at very close clock speeds. Maybe the XTX will have the 1.0 ram, who knows? But still seems the clocks are closer together than even the X1800XT and XL.

I did hear the X1900s will use the same HSF as the partners use but will run cooler than the X1800s.


They are all built by ATI so all partners are identical except for bundles! They are still using the same blower fans as X1800's so loud on startup but pretty quiet when running.
However the cards run very cool, no longer hot at all.

http://forums.overclockers.co.uk/showthread.php?t=17520845&page=3&pp=30

[no-X]

...but will run cooler than the X1800s

X1800XL - 1.1V core, 1.9V RAM
X1800XT - 1.35V core, 2.1V RAM
X1900XT - 1.2V core, 2.1V RAM (http://resources.vr-zone.com/newspics/Jan06/12/3DMark05-625-725-FX55.gif)

[trinibwoy]

X1800XT - 1.35V core, 2.1V RAM
X1900XT - 1.2V core, 2.1V RAM (http://resources.vr-zone.com/newspics/Jan06/12/3DMark05-625-725-FX55.gif)

Wonder what they did to chop .15v off vcore.

[Jawed]

A year playing with TSMC's 90nm?

Jawed

[Geo]

So it went to 715mhz on the core on air cooling, and no volt mods? Is that second right? Because if R520 was 1.35 and R580 is 1.20 and OCing to 715mhs, then I'm scratching my head even harder over just what the thinking is here on XT at 625 and XTX at 650. There's got to be another shoe to drop, of some fashion.

[ERK]

Low power is good for it's own sake, of course, but I wonder if this clock/volts decision has anything to do with the OEM market?

[no-X]

A year playing with TSMC's 90nm?

Jawed
And possibly the new silicon revision (hkepc showed A22 or A23 revision of R580 (http://img189.imageshack.us/img189/7199/r5801lgn20umqtoil1iu.jpg))

[SugarCoat]

Can someone explain to me why they're being stingy on the ram clocks? Wouldnt they still want to maximize raw speed for the pixel benchies and games and populate todays reviews and gamers libraries? As it stands i dont see this being a dominating force over either the R520 clocked at 700 or the GTX512 in what people play today save a very limited amount of titles. Its a mid-year update alright, just seems like a gamble that it will hold them over until the R600. Not exactly what i expected.

[The_Wolf_Who_Cried_Boy]

Can someone explain to me why they're being stingy on the ram clocks? Wouldnt they still want to maximize raw speed for the pixel benchies and games and populate todays reviews and gamers libraries? As it stands i dont see this being a dominating force over either the R520 clocked at 700 or the GTX512 in what people play today save a very limited amount of titles. Its a mid-year update alright, just seems like a gamble that it will hold them over until the R600. Not exactly what i expected.

Possibly faster ram isn't currently available in the quantites they want to ship in, and ATI always has been a bit conservative with memory clocks.

[tEd]

A year playing with TSMC's 90nm?

Jawed

Yes probably. I remember how richard huddy mentioned that even though they moved to a smaller process(90nm) they needed to keep the volts high because of leakage. It seems they have the problem better under control now

[ANova]

I have to question ATI's strategy myself. An X1900 XT at 625 with 750 MHz ram, $500-550 and an XTX or PE at 700 MHz with 800-900 MHz ram, $600 would seem like a better decision.

[Moloch]

I have to question ATI's strategy myself. An X1900 XT at 625 with 750 MHz ram, $500-550 and an XTX or PE at 700 MHz with 800-900 MHz ram, $600 would seem like a better decision.
Ati knows no wrong.. they don't make mistakes... they are the alpha and the omega!
When it looks like they make mistakes it is in fact people who are in the wrong, but ati humbleing takes the blame:wink:

[N00b]

Ati knows no wrong.. they don't make mistakes... they are the alpha and the omega!
When it looks like they make mistakes it is in fact people who are in the wrong, but ati humbleing takes the blame:wink:
Yes, your're absolutely right... wait... do I smell sarcasm here? Heretic! Someone call the B3D Inquisition. Let's drown him in bad rep! :wink:

[Unknown Soldier]

:) I was expecting the X1900XT out 16th January .. but guess now that's wrong. So the 25th it is.

Oh Well. :(

US

[Moloch]

Yes, your're absolutely right... wait... do I smell sarcasm here? Heretic! Someone call the B3D Inquisition. Let's drown him in bad rep! :wink:
:lol:

[eSa]

Nicked from the other thread;

http://www.jr.com/JRProductPage.process?Product_Id=4077619&JRSource=google.datafeed.ATI+100435805


After endless speculation, what is so great about the GPU ?

Apart from a lot more shaderpower what is new ? I dont recognice any new features. A more like an expensive refresh nothing more. Plus at least in Europe price is simply pathetic. 700+ euros for gfxcard ?!

[Joe DeFuria]

After endless speculation, what is so great about the GPU ?

Um..."endless speculation?" We've pretty much "known" exactly what this GPU was ever since the R5xx launch. (Everything but clock-speed).

Apart from a lot more shaderpower what is new ?

Approx 3x the shader power is nothing to sneeze at.

There may be some other new tid-bits in there, but no one is expecting any significantly new features, nor has anyone expected any. I don't think anyone expects any significantly new features from either ATI or nVidia on any new GPUs until DX10.

[Moloch]

Ya but how will that translate to high res (1280x960+) performance with 4x fsaa(and TAA) and 16XAF?

[Joe DeFuria]

Ya but how will that translate to high res (1280x960+) performance with 4x fsaa(and TAA) and 16XAF?

Probably about the same as a similarly clocked (core and memory) R520.

Is anyone expecting anything different?

[Moloch]

Probably about the same as a similarly clocked (core and memory) R520.

Is anyone expecting anything different?
Well then what's to get excited about then?
Only thing such a part would be good for is 3dmark.. for now.

[Joe DeFuria]

...Only thing such a part would be good for is 3dmark.. for now.

Christ...Crossfire it if the raw pixel power is not "enough" for you... ;) What you're looking for is goverened more by memory bandwidth than GPU power.

Prepare to be "disappointed" for the forseeble future...or at least at the whim of the memory manufacturers.

[Moloch]

Christ...Crossfire it if the raw pixel power is not "enough" for you... ;) What you're looking for is goverened more by memory bandwidth than GPU power.

Prepare to be "disappointed" for the forseeble future...or at least at the whim of the memory manufacturers.
So I have to spend 1000 bucks to get the high res power I want?
Is it really that hard to understand that when it comes down to it high resolution benchmarks with fsaa and AF are really more important than simply having a card that may be much faster at low res where the bottleneck is more on the ALUs?

[Joe DeFuria]

So I have to spend 1000 bucks to get the high res power I want?

To play the modern games at the res/power you want? Yup.

Is it really that hard to understand that when it comes down to it high resolution benchmarks with fsaa and AF are really more important than simply having a card that may be much faster at low res where the bottleneck is more on the ALUs?

Is it really hard to understand that it is memory bandwidth that is the constraint you are dealing with? You can have 3x the number of ROPS/TMUs, etc, on a GPU but it won't do jack-all for your personal needs unless there is more bandwidth available to feed them.

So rather than let a huge number of transistors go to WASTE on a GPU that is starved for bandwidth, they are put to use toward shading power, which can be taken advantage of in ways that are not as dependent on bandwidth.

So as I said. Prepare to be disappointed, or be prepared to shell out the bucks.

[Moloch]

To play the modern games at the res/power you want? Yup.



Is it really hard to understand that it is memory bandwidth that is the constraint you are dealing with? You can have 3x the number of ROPS/TMUs, etc, on a GPU but it won't do jack-all for your personal needs unless there is more bandwidth available to feed them.

So rather than let a huge number of transistors go to WASTE on a GPU that is starved for bandwidth, they are put to use toward shading power, which can be taken advantage of in ways that are not as dependent on bandwidth.

So as I said. Prepare to be disappointed, or be prepared to shell out the bucks.
Then how come ati isn't clocking up the memory?
Atleast it would help them be more competetive in current games.
I guess it really can be thought of a X1600...

[Ailuros]

Then how come ati isn't clocking up the memory?
Atleast it would help them be more competetive in current games.
I guess it really can be thought of a X1600...

When IHVs reach decisions like that they do not only take their internal test results under account, but memory availability projections and pricing for a given timeframe.

[Moloch]

When IHVs reach decisions like that they do not only take their internal test results under account, but memory availability projections and pricing for a given timeframe.
...But I want my high res gaming:evil:
:grin:

[Martin Eddy]

...But I want my high res gaming:evil:
:grin:

At what cost? :)

[Moloch]

At what cost? :)
Preferably ati's:wink:
I'd rather them do yet another R300 spin and have them make say a 24FP part with.. 1800 mhz ram:D then just a 16FP part with 48 ALUs.
But I guess just it's cheaper for them just to upgrade the shader units :D
I just prefer like high res gaming so I wouldn't mind say missing out on some shader effects because of a lack of shading power.
To that end I prefered to the R4XX to the NV4X.

[Martin Eddy]

Thank god everyone doesn't think like you or we would still be back in the days before hardware T&L. :wink:

[Moloch]

Thank god everyone doesn't think like you or we would still be back in the days before hardware T&L. :wink:
You don't have me figured out I'm afraid :razz:
Whats the point of all that shader power if at the settings I play it it's say 10% faster than a standard X1800XT?
Untill we have games that are taking advantage of the additional shader power it's a waste for consumers.

[no-X]

Untill we have games that are taking advantage of the additional shader power it's a waste for consumers.
Some people don't have enough money to buy a new video card each three months...

[Moloch]

Some people don't have enough money to buy a new video card each three months...
You'd have to be a pretty patient person to buy a R580 based on future performance since it will perform along the same line as the R520 in in todays games.
The only game I know of thats coming out soon that I assume will show some benefits from the addition alus is elder scrolls oblivion.
Well atleast the R580 should bring nice numbers in for HDR rendering... though I have yet to see a game that makes good use of it.

[Chalnoth]

Radeonic, the extra shader power isn't going to be that under-utilized except in games where either card will perform very highly anyway.

[no-X]

You'd have to be a pretty patient person to buy a R580 based on future performance since it will perform along the same line as the R520 in in todays games.
The only game I know of thats coming out soon that I assume will show some benefits from the addition alus is elder scrolls oblivion.
Well atleast the R580 should bring nice numbers in for HDR rendering... though I have yet to see a game that makes good use of it.
Don't expect, that the difference will be so low. I expect at least 20% increase (X1800XT/X1900XT) in D3D games (BattleField2, FarCry, SC, COD2...) and a bit more in FEAR. If X1900XT won't be more expensive than X1800XT-PE, it will probably have better price/performance ratio and thats the goal (for today). And if the performance difference will be more significant in the course of time, it will be very good for ATi as an advertisement and for customers too.

[Tim]

You'd have to be a pretty patient person to buy a R580 based on future performance since it will perform along the same line as the R520 in in todays games.

The R580 will have a big lead over the r520 in recent games, especially at high resolutions. I do not consider the slim advantage in DX8-class games to be a big problem, the performance in these games a good enough already with current cards.

[Joe DeFuria]

Some people don't have enough money to buy a new video card each three months...

And, for those who really WANT such high-res settings...you can still get it today. You just have to pay for it. (SLI / Crossfire.)

You want to play at settings that is an extremely niche market? You'll have to play niche market prices.

[Joe DeFuria]

Preferably ati's:wink:
I'd rather them do yet another R300 spin and have them make say a 24FP part with.. 1800 mhz ram:D then just a 16FP part with 48 ALUs.

Doens't nVidia already have a similar a part? Why don't you buy one of those...oh...wait a minute...you can't actually find any. I wonder if bandwidth price / availability (that speed ram) has anything to do with it...

[Tahir2]

I expect the X1900XT to be a lot more than 10% faster in most situations than the X1800XT. Having had a chance to play with the X1600XT it is obvious that even some of the current games which use only certain shader effects are benefitting like NFS: Most Wanted.

[Moloch]

I expect the X1900XT to be a lot more than 10% faster in most situations than the X1800XT. Having had a chance to play with the X1600XT it is obvious that even some of the current games which use only certain shader effects are benefitting like NFS: Most Wanted.
The problem with the X1600 is the inconsistant performance.
And joe.. please.. see those big smiles?
I was just saying what I would like to see, not a serious this it was ati should do.
If the R580 is similarly clocked to the R520 why should it maintain a large lead over the R520 at high res where it should be bandwidth limited?

[Dave Baumann]

The problem with the X1600 is the inconsistant performance.
Can you explain how it is inconsitent? If you are comparing it to a competetive part then how can you say which is consistent and which isn't?

[Joe DeFuria]

I was just saying what I would like to see, not a serious this it was ati should do.

Hey, I would like to see octal SLI performance in a single chip selling for $10....

[Tahir2]

Hmm thinking about that, why not get a second hand 9700 Pro for $10 which has 8 pipelines - even better than SLI, it is internal SLI!!! Sorry I'll be quiet now.

[neliz]

I expect the X1900XT to be a lot more than 10% faster in most situations than the X1800XT. Having had a chance to play with the X1600XT it is obvious that even some of the current games which use only certain shader effects are benefitting like NFS: Most Wanted.

that'd be 50% faster in FEAR, right?

[no-X]

If the R580 is similarly clocked to the R520 why should it maintain a large lead over the R520 at high res where it should be bandwidth limited?
High resolutions are fill-rate limited, not bandwidth limited.

[Moloch]

Hey, I would like to see octal SLI performance in a single chip selling for $10....
Now that's just being silly Joe:wink:
Can you explain how it is inconsitent? If you are comparing it to a competetive part then how can you say which is consistent and which isn't?
Because it's a free country :razz:
The 6800gs performs well across a number of games.. old and new, a 1600 performs well in newer games which are more shader heavy.
It's simple really.
I dont know why you are trying to argue it isn't.

High resolutions are fill-rate limited, not bandwidth limited.
And where is it getting this uber pixel pushing power from?
It's a 16 "pipe" card with 48 alus right?
So where's it going to get the additional fillrate from to maintain the lead at high res?

[Chalnoth]

Because it's a free country :razz:
The 6800gs performs well across a number of games.. old and new, a 1600 performs well in newer games which are more shader heavy.
It's simple really.
I dont know why you are trying to argue it isn't.
Well, the X1600 is also a part which has a lower transistor count, and is on a smaller process, and as such should be quite a bit cheaper to manufacture than the 6800GS.

And where is it getting this uber pixel pushing power from?
It's a 16 "pipe" card with 48 alus right?
So where's it going to get the additional fillrate from to maintain the lead at high res?
Because the vast majority of the time, GPU's these days aren't outputting pixels. So it doesn't really matter how many pixels it can output per clock. What matters the most is how much processing power it has in relation to what kind of processing the game needs.

What the R580 is going to bring to the table is an increase in one particular aspect of processing: pixel shader math. This won't improve performance in all games, but there will be dramatic improvements in some. Just due to the way technology is moving forward, is it highly likely that nVidia will go a similar route with its next-generation architecture (more math power in relation to texture power).

This really is the way of the future. It's pointless to worry about slight relative decreases in performance for older games (given same transistor count, clock speed, yada yada yada), because performance is going to be high anyway.

That all said, while I think that many things about the R580 architecture are really, really good, that doesn't mean that I believe it to be necessarily better than the upcoming G71 from nVidia. I'm sure that the R580 will be significantly better, however, than the R520, and that it will manage to beat the G71 in some games as well (Even if the G71 ships at the rumored 700MHz with 32 pipelines).

[Dave Baumann]

The 6800gs performs well across a number of games.. old and new, a 1600 performs well in newer games which are more shader heavy.
It's simple really.
I dont know why you are trying to argue it isn't.
But you aren't making a valid comparison. X1600, as a chip, is designed to meet a price range of $130-$200, 6800 GS is different product, with a different composition, a 256-bit bus with a rock bottom price of $200 - the direction that titles will go is the direction that X1600 points to, and this is proved by the fact that it outperforms the chip its replacing on virtually any title, even down to the the texture heavy UT2004.

[Geo]

6800gs is a great value that folks looking for that price point should hop on while they still can. But I should be extraordinarily suprised if it is around much past March.

Tho it will be very interesting to bench 7600 against it too.

[Moloch]

Well, the X1600 is also a part which has a lower transistor count, and is on a smaller process, and as such should be quite a bit cheaper to manufacture than the 6800GS.


Because the vast majority of the time, GPU's these days aren't outputting pixels. So it doesn't really matter how many pixels it can output per clock. What matters the most is how much processing power it has in relation to what kind of processing the game needs.

What the R580 is going to bring to the table is an increase in one particular aspect of processing: pixel shader math. This won't improve performance in all games, but there will be dramatic improvements in some. Just due to the way technology is moving forward, is it highly likely that nVidia will go a similar route with its next-generation architecture (more math power in relation to texture power).

This really is the way of the future. It's pointless to worry about slight relative decreases in performance for older games (given same transistor count, clock speed, yada yada yada), because performance is going to be high anyway.

That all said, while I think that many things about the R580 architecture are really, really good, that doesn't mean that I believe it to be necessarily better than the upcoming G71 from nVidia. I'm sure that the R580 will be significantly better, however, than the R520, and that it will manage to beat the G71 in some games as well (Even if the G71 ships at the rumored 700MHz with 32 pipelines).
I agree with almost everything you just said.
But you aren't making a valid comparison. X1600, as a chip, is designed to meet a price range of $130-$200, 6800 GS is different product, with a different composition, a 256-bit bus with a rock bottom price of $200 - the direction that titles will go is the direction that X1600 points to, and this is proved by the fact that it outperforms the chip its replacing on virtually any title, even down to the the texture heavy UT2004.
Outperforming the chip it's replacing was ati's argument :roll:
Ati originally targeted the 250 doller range.. mind explaining why they wanted to sell it for 250 before nvidia raped them with the 6800gs?
Nvidia forced them to drop about 100 dollars off the price tag, so ati was originally thinking it was more of a mid-high end part then a low-midrange chip.
For the people who can't spring for the 6800gs, it's too bad since it's a high end part with a midrange price tag.. you can quote me on that:wink:

[Dave Baumann]

Radeonic - fundamental priciple of semis is to sell a chip for what you can, but try never to sell it for less than its worth. The die size of RV530 is congruent with the die sizes of the parts that existed in the price bracket that it does now, and it probably wouldn't be on the market if it couldn't exist at these prices. Its obvious that ATI priced it to where the competition was at the time of announcement, but did so rather stupidly given that there was a 2 month delay between announcement and introduction.

As for outperforming the chip its replacing - I don't recall ATI making references to X700 in relation to X1600, however I have just spent a fair bit of time benchmarking them in relation to one another. The point I'm making is that despite a lower colour and texture fillrate the X1600 XT manages to outperform the X700 PRO in 100% of the gaming scenarios I threw at it.

[Moloch]

Radeonic - fundamental priciple of semis is to sell a chip for what you can, but try never to sell it for less than its worth. The die size of RV530 is congruent with the die sizes of the parts that existed in the price bracket that it does now, and it probably wouldn't be on the market if it couldn't exist at these prices. Its obvious that ATI priced it to where the competition was at the time of announcement, but did so rather stupidly given that there was a 2 month delay between announcement and introduction.

As for outperforming the chip its replacing - I don't recall ATI making references to X700 in relation to X1600, however I have just spent a fair bit of time benchmarking them in relation to one another. The point I'm making is that despite a lower colour and texture fillrate the X1600 XT manages to outperform the X700 PRO in 100% of the gaming scenarios I threw at it.
Yes, the X1600 does perform well compared to the part it replaces depite what looks like on paper to be an easy win for the x700 pro.

In these very forums Opengl_guy comments about it.
http://www.beyond3d.com/forum/showpost.php?p=620267&postcount=17
http://www.beyond3d.com/forum/showpost.php?p=620267&postcount=20
Also see the X1800XL and X800XL bit.
http://www.beyond3d.com/forum/showpost.php?p=620303&postcount=26

So it's clear their strategy was to outperform themselves.

[Dave Baumann]

Regardless of whether one guy from ATI say it or not, its not a message they are pushing in relation to X1600 - but even then, testing can show it to be true. But of course newer boards are going to offer greater performance than the previous generation at similar price ranges (Ummmm, duh!), thats not the point.

The point of this is that despite lower texture and fillrates X1600 always outperforms X700 even on older titles that don't rely on sahders as much - it can equally be argued that X1600 may be more balanced in many resects than older generation boards if they are outperforming them (depsite lower theoretical rates that were considered to be important on older titles) on older titles let alone newer ones!

[Moloch]

Regardless of whether one guy from ATI say it or not, its not a message they are pushing in relation to X1600 - but even then, testing can show it to be true. But of course newer boards are going to offer greater performance than the previous generation at similar price ranges (Ummmm, duh!), thats not the point.

The point of this is that despite lower texture and fillrates X1600 always outperforms X700 even on older titles that don't rely on sahders as much - it can equally be argued that X1600 may be more balanced in many resects than older generation boards if they are outperforming them (depsite lower theoretical rates that were considered to be important on older titles) on older titles let alone newer ones!
And what can you contribute the wins to?
I might mention the geforce 3 also outperformed the geforce 2 ultra depite lower theoretical performance.
The theoretical differences between them though wasn't that great however.

[Dave Baumann]

Sorry, "contribute the wins to" or "attribute the wins to"? If you're asking what do I attrubite the wins to then its going to be a mix of things - X1600 has higher Z rate (useful for Doom/Quake) a higher vertex rate, higher bandwidth, but also a considerably higher Pixel Shader rate which will be a huge contributing factor in most cases on current titles.

[Moloch]

Sorry, "contribute the wins to" or "attribute the wins to"? If you're asking what do I attrubite the wins to then its going to be a mix of things - X1600 has higher Z rate (useful for Doom/Quake) a higher vertex rate, higher bandwidth, but also a considerably higher Pixel Shader rate which will be a huge contributing factor in most cases on current titles.
Attribute I mean :oops:
Now you say it outperforms the X700 pro in 100% of your testing which would include UT2004 which is texture heavy?

[Chalnoth]

Attribute I mean :oops:
Now you say it outperforms the X700 pro in 100% of your testing which would include UT2004 which is texture heavy?
He did say that :)

I would be willing to bet that the X1600 also gains over the X700 from the new memory controller.

[Moloch]

He did say that :)

I would be willing to bet that the X1600 also gains over the X700 from the new memory controller.
So a card with a 2300mpixel fillrate and a better memory controllor can outperform a card with a 3800 mpixel and worse (slightly?) MC in a game that only uses pixelshaders to speed up the game?

[Pete]

I would be willing to bet that the X1600 also gains over the X700 from the new memory controller.Among (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai12.jpg) other (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai13.jpg) things (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai14.jpg), apparently (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai216.htm).

Is NV sacrificing the margins they tout (per Uttar's conference call summaries) by selling the 6800GS for so "little," or is that GPU and its components just very common/cheap by now?

[Ailuros]

So a card with a 2300mpixel fillrate and a better memory controllor can outperform a card with a 3800 mpixel and worse (slightly?) MC in a game that only uses pixelshaders to speed up the game?

http://www.xbitlabs.com/articles/video/display/radeon-x1600_11.html

http://www.xbitlabs.com/articles/video/display/radeon-x1600_3.html

[Dave Baumann]

I think you'll probably find that the performance differences not just lie in the RV5xx architectural advantages, but also because there are cases where the X700 PRO's memory bandwidth just isn't going to support 8 textures in general usage (and especially not 8 pixels). X700 is probably the more imbalanced of the two architectures here - its primary reason for existing was because ATI wanted an 8 fragment pipeline part, but R4xx didn't have the architectural advances to scale textures and ROPs differently from fragment pipes, so it was a convenience just to go 8:8:8 throughout.

[Moloch]

http://www.xbitlabs.com/articles/video/display/radeon-x1600_11.html

http://www.xbitlabs.com/articles/video/display/radeon-x1600_3.html
I read daves article:wink:

[Chalnoth]

Among (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai12.jpg) other (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai13.jpg) things (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai14.jpg), apparently (http://pc.watch.impress.co.jp/docs/2005/1007/kaigai216.htm).
That's, um, exactly what I was talking about in relation to the memory controller. Nice to see that I was right, though :)

[Tim]

And where is it getting this uber pixel pushing power from?
It's a 16 "pipe" card with 48 alus right?
So where's it going to get the additional fillrate from to maintain the lead at high res?

The R520 has enough fixed function fillrate to do 5000FPS at 1600x1200 - why do you need more? To increase the programable (pixel shader) fillrate, like they do with the r580, seems very logical to me.

[Mabru]

384 million of transistors.:wink:
http://img29.imageshack.us/img29/1350/x1900xtx9zu.th.jpg (http://img29.imageshack.us/my.php?image=x1900xtx9zu.jpg)

[Joe DeFuria]

384 million of transistors.:wink:
http://img29.imageshack.us/img29/1350/x1900xtx9zu.th.jpg (http://img29.imageshack.us/my.php?image=x1900xtx9zu.jpg)

Dynamic voltage and clockspeed control.

Is that new? (Dynamic voltage?)

[Mabru]

Is that new? (Dynamic voltage?)
I think R520 has it as well.

[overclocked]

384M transistors hmmm. That makes one pixelprocessor come in at slightly less than 2 million transistors.

[Geo]

384 million of transistors.:wink:
http://img29.imageshack.us/img29/1350/x1900xtx9zu.th.jpg (http://img29.imageshack.us/my.php?image=x1900xtx9zu.jpg)

Nice! Rep love left in appreciation. . .:grin:

[Chalnoth]

384M transistors hmmm. That makes one pixelprocessor come in at slightly less than 2 million transistors.
It's unlikely that the additional pixel processors is the only difference between the designs.

[EX||illuminati]

I think R520 has it as well.
With my X1800 XL I can change the voltage and the clock speed with something like ati tool. However I would not call this dynamic. Dynamic to me means the card automatically ups the clock speed and voltage with increase load. It sounds like it shares some of the laptop power and heat saving features. Or maybe I am just reading to much into it.

[Joe DeFuria]

Also, I believe ATI does change clockspeed (and voltage?) when switching between 2D to 3D mode. That could be considered "dynamic", though it's impossible to tell from a bullet point if that's what is meant.

[Moloch]

The R520 has enough fixed function fillrate to do 5000FPS at 1600x1200 - why do you need more? To increase the programable (pixel shader) fillrate, like they do with the r580, seems very logical to me.
ya ok.. 5000fps.. I'll get back to you on that.
Fillrate (and bandwidth) is king:wink:

[Geo]

Also, I believe ATI does change clockspeed (and voltage?) when switching between 2D to 3D mode. That could be considered "dynamic", though it's impossible to tell from a bullet point if that's what is meant.

Well, you can tell it to when OCing anyway. My X1800xl AIW has a checkbox in the Overdrive tab for telling it to only use the OC settings when in 3D mode. The clocks do seem to take that (I've had it open when gaming and could see them change). But the default clocks, so far as I can see, remain the same for 2d/3d --in other words, there is no UI exposed to let you underclock the advertised clocks for 2d. Given they've proved they have the capability, you'd think they could program it to have lower default clocks in 2d, standard "advertised" clocks in 3d, and let you just overclock the latter. Clearly the capability is there, it is just a matter of how they implement it.

[Moloch]

I have a evga 7800GT and it's always clocked at 445mhz.
whats up with that?

[EX||illuminati]

Well, you can tell it to when OCing anyway. My X1800xl AIW has a checkbox in the Overdrive tab for telling it to only use the OC settings when in 3D mode. The clocks do seem to take that (I've had it open when gaming and could see them change). But the default clocks, so far as I can see, remain the same for 2d/3d --in other words, there is no UI exposed to let you underclock the advertised clocks for 2d. Given they've proved they have the capability, you'd think they could program it to have lower default clocks in 2d, standard "advertised" clocks in 3d, and let you just overclock the latter. Clearly the capability is there, it is just a matter of how they implement it.

Yeah my card is the same, that’s one reason I like using ati tool so much I can set it to lower the voltage, clock speed, and fan speed for 2d making the card run silently. It would be nice if the R580 had something like AMD's cool and quiet built in.

[Geo]

Yeah my card is the same, that’s one reason I like using ati tool so much I can set it to lower the voltage, clock speed, and fan speed for 2d making the card run silently. It would be nice if the R580 had something like AMD's cool and quiet built in.

I've been meaning to give ATI Tool a try, thanks for reminding me. :smile: Third party tools are great, don't get me wrong; but it'd be nice if ATI beefed up their own included interface in that direction, with whatever suitable safeguards they feel are necessary.

[SugarCoat]

Yeah my card is the same, that’s one reason I like using ati tool so much I can set it to lower the voltage, clock speed, and fan speed for 2d making the card run silently. It would be nice if the R580 had something like AMD's cool and quiet built in.


It does, im kind of surprised nobody knows about it from the responces so far.

The R520 core dynamically clocks memory and core speed as well as voltage depending on the operation.

My X1800XT for example runs at 594 Core 693 Memory @ 1.3V in 2D
in 3D the card switches to 695 Core 792 Memory @ 1.48V

thats under PE bios, yes stock bios does the same thing, 594 goes to ~625 and 693 goes to ~750 while the voltage goes from 1.3 to 1.4

I'm sure the X1800XL does the same. Anyone can download Rivatuner 2.0 and run the hardware monitor in the background and go play a game and see this for themselves. Thats what the R580 will do as well. As i have already stated my gripe with this dynamic clocking is the drivers dont distinguish from windowed 3D apps and 2D mode so they dont increase clocks when you game in a window.

[EX||illuminati]

I did not know that, I will have to give it a try when I get home.

It does, im kind of surprised nobody knows about it from the responces so far.

The R520 core dynamically clocks memory and core speed as well as voltage depending on the operation.

My X1800XT for example runs at 594 Core 693 Memory @ 1.3V in 2D
in 3D the card switches to 695 Core 792 Memory @ 1.48V

thats under PE bios, yes stock bios does the same thing, 594 goes to ~625 and 693 goes to ~750 while the voltage goes from 1.3 to 1.4

I'm sure the X1800XL does the same. Anyone can download Rivatuner 2.0 and run the hardware monitor in the background and go play a game and see this for themselves. Thats what the R580 will do as well. As i have already stated my gripe with this dynamic clocking is the drivers dont distinguish from windowed 3D apps and 2D mode so they dont increase clocks when you game in a window.

[Mintmaster]

384 million of transistors.:wink:
http://img29.imageshack.us/img29/1350/x1900xtx9zu.th.jpg (http://img29.imageshack.us/my.php?image=x1900xtx9zu.jpg)
Makes you wonder why ATI even bothered with R520. 20% more transistors, maybe translating into 10% more cost, and they'll probably get 20-100% increase in performance.

Myself, I can't wait until R600.

[Pete]

Heh, I like how they're advertising the transistor count and manufacturing process above all else.

Mint, I thought we'd all agreed that ATI got R520 out the door to meet their Oct 1st conference call and that BBQ flavor really is better than Sour Cream 'n Onion. In fact, I'm receiving the commemorative plaques tomorrow.

;^)

[Geo]

Mint, I thought we'd all agreed that ATI got R520 out the door to meet their Oct 1st conference call and that BBQ flavor really is better than Sour Cream 'n Onion. In fact, I'm receiving the commemorative plaques tomorrow.

;^)

F@nboy LIES!! Don't make me start a poll. Sour Cream 'n Onion will CRUSH BBQ!

[Ailuros]

ya ok.. 5000fps.. I'll get back to you on that.
Fillrate (and bandwidth) is king:wink:

He said fixed function fillrate (despite the obvious exaggeration).

Fillrate might be "king" as you put it, yet nowadays it's rather internal shader fill-rate or how many floating point operations a GPU can perform per clock. In that department R580 ups the ante by 3x times compared to R520 and in that consensus "that" fillrate is truly "king".

[Pete]

That's, um, exactly what I was talking about in relation to the memory controller. Nice to see that I was right, though :)Sorry, was my ignorance showing? :oops: For some reason I considered the memory controller distinct from the cache talked about in those slides.

[dizietsma]

From the specs of the X1900 then the only major change seems to be the x3 fragment rate and so it will be interesting to see how this approach compares to nvidia's g71 which I guess is going down the historical route of adding a bit of everything.

Fragment rate, I love it, just need to find someone now to bamboozle with it :D

[Moloch]

He said fixed function fillrate (despite the obvious exaggeration).

Fillrate might be "king" as you put it, yet nowadays it's rather internal shader fill-rate or how many floating point operations a GPU can perform per clock. In that department R580 ups the ante by 3x times compared to R520 and in that consensus "that" fillrate is truly "king".
Ya I know for shading the R580 rocks the socks off anything and everything but I'm curious how it will perform in todays games.

[neliz]

Ya I know for shading the R580 rocks the socks off anything and everything but I'm curious how it will perform in todays games.

depends what you think of todays games? fear yes ...quake4? nah..

[Martin Eddy]

The only thing that concerns me about R580 is 8 vertex processors, I was hoping for 10. Maybe 8 is plenty for now though.

[Moloch]

depends what you think of todays games? fear yes ...quake4? nah..
what about serious sam 2.. been playing that lately:smile:

[Tim]

The only thing that concerns me about R580 is 8 vertex processors, I was hoping for 10. Maybe 8 is plenty for now though.

I was also expecting at least 10-12 - 8 seems a bit unbalanced. Ati has most likely done a lot of analyzing on how the number of vertex shaders impacts actual game performance, so they most likely know what they are doing.

Anyway they might end up significant behind in vertex shader performance when the G71 is released.

[Bouncing Zabaglione Bros.]

I was also expecting at least 10-12 - 8 seems a bit unbalanced. Ati has most likely done a lot of analyzing on how the number of vertex shaders impacts actual game performance, so they most likely know what they are doing.

Anyway they might end up significant behind in vertex shader performance when the G71 is released.

The question you have to ask is will the games coming up over the next 12-18 months be pushing more vertices or more shaders? If it's the latter, then ATI made the right choice.

[Martin Eddy]

Time will tell.

[Razor1]

The question you have to ask is will the games coming up over the next 12-18 months be pushing more vertices or more shaders? If it's the latter, then ATI made the right choice.


Both, U3 is doubling or tripling in some cases vertex counts per scene.

But most of the bottlenecks will still be on the pixel shader side.

[Ailuros]

Ya I know for shading the R580 rocks the socks off anything and everything but I'm curious how it will perform in todays games.

Given how the architecture is being layed out I have severe doubts that it'll lack in pure multitexturing fillrate for one.

As for the rest I don't think it'll be too long until the first reviews appear.

[Jawed]

Given how the architecture is being layed out I have severe doubts that it'll lack in pure multitexturing fillrate for one.

http://www.beyond3d.com/reviews/ati/rv5xx/index.php?p=07

This shows the X1600XT's texture fillrate taking a big hit when multitexturing. Much less "graceful" than X700Pro.

Of course X700Pro also has the old cache and memory interface architecture, so it's a wobbly comparison.

X1800XL is somewhat smoother, though the single-texture fillrate looks like it could be (relatively, more so than X1600XT) bandwidth limited:

http://www.beyond3d.com/reviews/ati/r520/index.php?p=15#fill

So I'm going to guess that the 3:1 ALU:TEX ratio "unnaturally" favours single-texturing... Perhaps that's not a bad thing though.

Jawed

[Dave Baumann]

There's a different ratio of texture sampling ability to bus width at the low end.

[Jawed]

Which plays into X1600XT's favour, as far as I can tell. 4 texture pipes fed by 256-bits versus 16 pipes fed by 512 bits.

Additionally, without L2 texture cache, R520/580 waste internal memory bus bandwidth sending duplicates of texture data to each of the four texture caches, one per TMU - that is presuming that the textures concerned repeat across the screen.

Jawed

[Dave Baumann]

Which plays into X1600XT's favour, as far as I can tell. 4 texture pipes fed by 256-bits versus 16 pipes fed by 512 bits.

Err, more like there is 4x the texture sampling capability at the high end, but only 2x the bus width - ergo the high end is likely to be bandwidth limited with single texturing, but less so multitexturing. The low end is less bandwidth limited single texturing, hence relatively more texture fillrate limited with multitexturing.

Note the X700 PRO is bandwidth limited with single texturing, but less so multitexturing, which is why it drops less dramatically.

Additionally, without L2 texture cache, R520/580 waste internal memory bus bandwidth sending duplicates of texture data to each of the four texture caches, one per TMU - that is presuming that the textures concerned repeat across the screen.
Relatively there is no difference between a single and multiple quad ATI design here since the single quad is still working on tiles, so it too will be needing to resample that texture when it moves from one tile to the next.

[Chalnoth]

To tell you the truth, I doubt that memory bandwidth has much of anything to do with texturing these days, except perhaps in the case of filtered FP16 textures, which the X1x00 architecture doesn't do anyway.

[Jawed]

Err, more like there is 4x the texture sampling capability at the high end, but only 2x the bus width - ergo the high end is likely to be bandwidth limited with single texturing, but less so multitexturing. The low end is less bandwidth limited single texturing, hence relatively more texture fillrate limited with multitexturing.
Yep, that's why I originally said X1800XL's single-texturing rate is relatively more bandwidth limited than X1600XT's - hence it's multitexturing rate won't show such a strong fall-off.

Relatively there is no difference between a single and multiple quad ATI design here since the single quad is still working on tiles, so it too will be needing to resample that texture when it moves from one tile to the next.
I can't agree. All four TMUs in R520/R580 need the same texture data simultaneously, which puts a relatively higher peak workload on the memory bus - whereas the same pattern of texturing in X1600XT will not suffer such exaggerated peaks, because the same number of texture requests is spread out in time (4x, in effect).

Now, whether the memory controller in R5xx understands that texels are often shared by TMUs and so gathers-together the distinct TMUs' texture requests and services them with a single read from memory, followed by a multi-way write to TMU cache - that's another question.

Jawed

[Jawed]

To tell you the truth, I doubt that memory bandwidth has much of anything to do with texturing these days, except perhaps in the case of filtered FP16 textures, which the X1x00 architecture doesn't do anyway.
Well I still think that L1/L2 patent by ATI should become relevant - I don't think it was just for Xenos.

Jawed

[Dave Baumann]

To tell you the truth, I doubt that memory bandwidth has much of anything to do with texturing these days, except perhaps in the case of filtered FP16 textures, which the X1x00 architecture doesn't do anyway.
Texturing is the second largest consumer of bandwidth. Filtering in the TMU as opposed to filtering in the shader will still require the same number of samples, hence bahdwidth used.

I can't agree. All four TMUs in R520/R580 need the same texture data simultaneously, which puts a relatively higher peak workload on the memory bus - whereas the same pattern of texturing in X1600XT will not suffer such exaggerated peaks, because the same number of texture requests is spread out in time (4x, in effect).
I assume you mean 4 groups of four texture units? However, that is not the case. Each of the 4 thread units are working on separate tiles and separate separate localities, and there is not necessarily any reason these have to be concurrent tiles (dependant on fast one quad has finished its previous tiles in relation to the other quads) - because of the tiled nature of the different quads the localities of the pixels are much further appart (even if they were working on adjacent tiles) hence an L2 cache is not likely to produce much hits in the first place. If there is bandwidth contention between different tiles then its the job of the dispatch procesor to scheddule the insructions to work around that as much as possible for each tile; if worst comes to worse one or two tiles will just use up al the bandwidth, but finish that much faster than the others. Overall the bandwidth usage for textures stays the same from 1 to 4 quad designs (or even 8 quad with Crossfire, which is why Supertiling isn't loosing any efficiency, whereas such a method would on a current NVIDIA part).

[Edit] Do a historical search on "Sireric" and tiling, he's talked along these lines before.

[RoOoBo]

Texturing is the second largest consumer of bandwidth. Filtering in the TMU as opposed to filtering in the shader will still require the same number of samples, hence bahdwidth used.

Texture is the largest consumer of bandwidth when using relatively 'long' shaders (longer than two or four instructions and with a couple or more texture accesses), not using depth/stencil passes and not using AA. When you start to output a color value to the framebuffer every 5 to 10 cycles color bandwidth becomes meaningless. And z bandwidth is even less due to compression. With blending that may move to 10 rather than 5 but it won't make a difference when you shader is 20-30 instructions long and fetches 8 textures.

[Chalnoth]

Texturing is the second largest consumer of bandwidth. Filtering in the TMU as opposed to filtering in the shader will still require the same number of samples, hence bahdwidth used.
Yes, but filtering in the shader will use up more shader math, putting less strain on bandwidth against the amount of ALU power needed.

[Chalnoth]

Texture is the largest consumer of bandwidth when using relatively 'long' shaders (longer than two or four instructions and with a couple or more texture accesses), not using depth/stencil passes and not using AA. When you start to output a color value to the framebuffer every 5 to 10 cycles color bandwidth becomes meaningless. And z bandwidth is even less due to compression. With blending that may move to 10 rather than 5 but it won't make a difference when you shader is 20-30 instructions long and fetches 8 textures.
Sure, but when this is the case, the bandwidth required is small regardless. And in the case when it's framebuffer bandwidth that is the primary limitation, then you're much more likely to be hitting the badnwidth limits of the GPU, but texture memory bandwidth is a relatively small portion of the total usage.

[Jawed]

I assume you mean 4 groups of four texture units?
Yeah - I always feel foolish writing out the long-hand descriptions of these things and prefer "quad" to be implicit.

However, that is not the case. Each of the 4 thread units are working on separate tiles and separate separate localities, and there is not necessarily any reason these have to be concurrent tiles (dependant on fast one quad has finished its previous tiles in relation to the other quads)
I was trying to explain the behaviour of the single-/multi-texturing fillrate benchmark, primarily - though I think it's partly relevant to games. I dunno how the benchmark is configured though, so I dunno if the texturing repeats and if so, how regularly. But obviously my comments only relate to textures that do repeat.

As regards locality, well I think that depends on the graphics engine - e.g whether it generates a terrain as one big mesh with multiple (differently-scaled) repeating textures across it.

Having said all that, clearly the freely-schedulable 128 threads in each shader unit generate a lot of leeway - but only if the code's texturing-intensity isn't too high (taking account of the variety of shaders in those 128 threads). Those 128 threads are the equivalent of (assuming 16x16 tiles):

8 screen tiles in R520
24 screen tiles in RV530 and R580.While a 1024x768 screen consists of 3072 tiles.

Jawed

[eSa]

http://www.firingsquad.com/hardware/3dmark_06/page8.asp

AMD Athlon 64 FX-60
NVIDIA GeForce 7800 GTX 512MB

3DMark06 1280x1024x32 SM 2.0 = 2162
3DMark06 1280x1024x32 HDR/SM 3.0 = 2203


http://forums.vr-zone.com.sg/showthread.php?t=53164&page=6

AMD Athlon 64 FX-60 @2616 MHz
ATI Radeon X1900XT

3DMark06 1280x1024x32 SM 2.0 = 2081
3DMark06 1280x1024x32 HDR/SM 3.0 = 2279



So, where is the superior shader performance ?

[Chalnoth]

As always, don't bother to use 3DMark as a representative benchmark of performance. Look at actual games.

[Mariner]

So, where is the superior shader performance ?

Bear in mind that due to the way Futuremark wrote the shaders, ATI cards are doing more work than NV cards to obtain their score. It's not an 'apples to apples' comparison.

Check out this (http://www.beyond3d.com/forum/showthread.php?t=26232) thread for details. Somewhere around page 5 or 6 I think!

[Chalnoth]

Bear in mind that due to the way Futuremark wrote the shaders, ATI cards are doing more work than NV cards to obtain their score. It's not an 'apples to apples' comparison.

Check out this (http://www.beyond3d.com/forum/showthread.php?t=26232) thread for details. Somewhere around page 5 or 6 I think!
Okay, I read the entire thread, and I didn't see any indication that ATI hardware needs to do more work than nVidia hardware.

[AlStrong]

Okay, I read the entire thread, and I didn't see any indication that ATI hardware needs to do more work than nVidia hardware.

Check page 9 if you have the forums set to 20 posts per page.

[Chalnoth]

Check page 9 if you have the forums set to 20 posts per page.
Still nothing.

If you're talking about the PCF vs. Fetch4, it's a non-issue (in this context). Look on page 10, Neeyik's post.

[R300King!]

We asked Futuremark about its decision to use the 24-bit format, and got the following response:

"In 3DMark06 we require 24-bit z-buffers at minimum to ensure that the quality and accuracy of the shadows are top-notch in all surroundings throughout the benchmark (during the benchmark, the demo and the game). During the development of 3DMark05 we made research into using 16 bit z-buffers, but noticed that the accuracy was not sufficient for the shadow technique we used. In 3DMark06 we use a new shadow technique, which is much better and more robust than the one in 3DMark05, but there are still cases where 24-bit z-buffers are barely enough. There are of course some cases and places where 16-bit z-buffers might be suffice, but to make sure that we do not get different rendering output with different hardware we require 24-bit z buffers for all hardware. 24-bit z-buffer is much more robust, and it ensures that the quality of the shadows is at the level we require."

I'm not sure if z-buffer is exactly the thing we're talking about—depth stencil textures and z-buffers aren't precisely the same thing, and pretty much all modern 3D hardware supports 24-bit z-buffers. Regardless, the decision to use the 24X8 texture format for the depth stencil texture favors current Nvidia cards heavily with a significant reduction in required bandwidth.

here (http://www.extremetech.com/article2/0,1697,1912117,00.asp)

[Razor1]

We asked Futuremark about its decision to use the 24-bit format, and got the following response:

"In 3DMark06 we require 24-bit z-buffers at minimum to ensure that the quality and accuracy of the shadows are top-notch in all surroundings throughout the benchmark (during the benchmark, the demo and the game). During the development of 3DMark05 we made research into using 16 bit z-buffers, but noticed that the accuracy was not sufficient for the shadow technique we used. In 3DMark06 we use a new shadow technique, which is much better and more robust than the one in 3DMark05, but there are still cases where 24-bit z-buffers are barely enough. There are of course some cases and places where 16-bit z-buffers might be suffice, but to make sure that we do not get different rendering output with different hardware we require 24-bit z buffers for all hardware. 24-bit z-buffer is much more robust, and it ensures that the quality of the shadows is at the level we require."

I'm not sure if z-buffer is exactly the thing we're talking about—depth stencil textures and z-buffers aren't precisely the same thing, and pretty much all modern 3D hardware supports 24-bit z-buffers. Regardless, the decision to use the 24X8 texture format for the depth stencil texture favors current Nvidia cards heavily with a significant reduction in required bandwidth.

here (http://www.extremetech.com/article2/0,1697,1912117,00.asp)

Hmm 24 bit stencil are supported on the x1600 I think, and a few other ATi cards.

Edit:
x1600 and x1300 support it also they use Fetch4

3DMark06 does support FETCH4, uses Dynamic Flow Control, uses 24 bit "DST" (prefer to use wording "Hardware Shadow Mapping") for any hardware that has hardware "DST" support.

I don't see it as one sided.

http://www.beyond3d.com/forum/showthread.php?t=26232&page=8


Couple posts up

24 bit depth stencil textures and Fetch4 are supported on X1300 and X1600 parts.
__________________

[Chalnoth]

I'm not sure if z-buffer is exactly the thing we're talking about—depth stencil textures and z-buffers aren't precisely the same thing, and pretty much all modern 3D hardware supports 24-bit z-buffers. Regardless, the decision to use the 24X8 texture format for the depth stencil texture favors current Nvidia cards heavily with a significant reduction in required bandwidth.
I'm not sure why. In these situations, the Radeons are using the DF24 format with Fetch4 for retrieving the depth information. I really don't see how this translates to less bandwidth used by nVidia hardware.

[Unknown Soldier]

I think people are reffering to the R520(which doesn't have Fetch4), but that's passed over your heads.

US

[Dave Baumann]

XT in stock:

http://www.digitallyunique.com/viewproduct.htm?category_id=419522&productId=4169611&srccode=cii_14110944&cpncode=10-16007872-2

http://www.computers4sure.com/search.asp?Operator=ALL&CONTEXT=SITE&Keyword=x1900&x=0&y=0

[boltneck]

WOW. :shock:

[Karma Police]

XT in stock:

http://www.digitallyunique.com/viewproduct.htm?category_id=419522&productId=4169611&srccode=cii_14110944&cpncode=10-16007872-2

http://www.computers4sure.com/search.asp?Operator=ALL&CONTEXT=SITE&Keyword=x1900&x=0&y=0

I haven't seen a card physically available for purchase before it's online reviews were released in a lonnnnnnnnng time. It's a strange feeling.........but good? :D

[Rur0ni]

Expensive too.

[Joe DeFuria]

I haven't seen a card physically available for purchase before it's online reviews were released in a lonnnnnnnnng time. It's a strange feeling.........but good? :D

Now we'll wait and see if reviewers bitch and moan in every video card review at nVidia for not following ATI's lead of having cards on the market days before they are reviewed...

[satein]

Now we'll wait and see if reviewers bitch and moan in every video card review at nVidia for not following ATI's lead of having cards on the market days before they are reviewed...
Techreport did one today on 7300GS ANOUNCEMENT (http://techreport.com/onearticle.x/9291) :wink: . Here is what they posted up on the front page today...
This product launch stands out from other recent introductions from the green team because it's not accompanied by hardware—neither for us press types to review nor for retailers in North America or Europe to sell. NVIDIA's official explanation is that the launch was pulled forward in time for the holiday buying season in the Asia Pacific region, where the first GeForce 7300 GS cards are being shipped. That's nice, but after giving ATI bucketfuls of grief over product availability at launch, NVIDIA is out on a limb here.

And the limb is creaking.

I am waiting to see what will [H] comment on this :razz: .

[MuFu]

I haven't seen a card physically available for purchase before it's online reviews were released in a lonnnnnnnnng time. It's a strange feeling.........but good? :D

7800GTX 512MB was available before it was reviewed.

[Joe DeFuria]

7800GTX 512MB was available before it was reviewed.

Yeah...just not after. ;)

[Moloch]

7800GTX 512MB was available before it was reviewed.
Define available :wink:

[Chalnoth]

How about, you can order one today, and get it shipped tomorrow?
http://www.newegg.com/Product/Product.asp?Item=N82E16814143044

Regardless, there were many stores who had GTX 512's in stock for the first day or two of availability. They sold out quickly, of course, and have been continually behind on stocking the parts by a few days.

Edit:
For posterity, because the link I have above may change, please note that at the time of writing, Newegg lists the BFG GeForce 7800GTX 512MB as being out of stock with an ETA of 1/20/2006.

[ERK]

Didn't stay in stock long.

Edit: Fine, go ahead and edit and make me look dumb. ;)

[ZioniX]

Some game benchmark figures. (GeForce 7800GTX 256MB vs. Radeon X1900XT 512MB)

http://www.dailytech.com/article.aspx?newsid=310

[Jawed]

Did you just buy their last one?

Jawed

[IbaneZ]

Some game benchmark figures.

http://www.dailytech.com/article.aspx?newsid=310

Cool.

Cards already in shops and benchmarks everywhere. Weird lauch. :)

[trinibwoy]

XT in stock:

http://www.digitallyunique.com/viewproduct.htm?category_id=419522&productId=4169611&srccode=cii_14110944&cpncode=10-16007872-2

http://www.computers4sure.com/search.asp?Operator=ALL&CONTEXT=SITE&Keyword=x1900&x=0&y=0

$560 :shock: Impressive!

[Moloch]

those are some mighty fine fear numbers.. however I'd like to see 512MB gtx scores.