GPU speculation

[DemoCoder]

It seems reasonable that Nvidia's G70 will have 24 pipes @ ~450Mhz, which depending on how you count, allows it to execute up to 42 billion shader ops/s (dual issue vector, dual issue scalar * 24 * 450Mhz) vs XB360's 48 billion. (except that the R500's alu ops are independent, and GF6's are dependent on shader code) However, it will have much superior fillrate which made aid global illumination and shadowing algorithms. e.g. if you need to render shadow buffers, stencils, cube map RT's, being able to fill fast is important, and long fast shaders won't allow you to do such lighting.

However, it would seem strange for Sony to just put a G70 which is a refreshed GF6 into the PS3. Since CELL is so great at processing vertices, it would make more sense to rip out the vertex shaders and video decoding circuitry, and use the die space for eDRAM or more pipes. Moreover, I would expect a PS3 GPU to be on .09um not .11m.

That means to me, there is no way in hell the PS3 can be launched in 2005. NVidia needs all year to get the GPU ready. I think we would have heard about a tape out of another non-G70 sample by now.

Somehow I expect current PS3 E3 demos will be on GF6 Ultra or G70 prototype, but that the final PS3 will have a completely different GPU.

Finally, could Sony do something really insane and put SLI'ed GPUs in their console?

 

[JF_Aidan_Pryde]

Suppose 4 SPEs are devoted to vertex processing. Is that comparable or better than the six or more vertex engines in next gen GPUs?

 

[DemoCoder]

Six vertex engines @ 500 Mhz = 3 billion vertex ops/sec or 6 billion if you count scalars.

4 SPEs @ 4.8 Ghz each = 19.2 billion ops/s assuming you can dispatch 1 per clock. Moreover, the SPEs are ALOT more flexible than the vertex shaders, being even more powerful than the Geometry Shaders of Longhorn's WGF2.0.

I'd say it makes sense to remove the VS from the GPU, since the most interesting geometry stuff, especially decompression/tesselation of geometry, skinning, shadow volumes, et al, will be done by the CPU anyway.

 

[ninelven]

erased

 

[marconelly!]

Finally, could Sony do something really insane and put SLI'ed GPUs in their console?

Well, they apparently do have two SLI-ed 6800s in their current devkits. I'm not sure if I'd look too much into that for making any conclusions, though. PS3's GPU will likely be G80 or whatever it's going to be called. It should be finished, by Nvidia's words, at the end of this year.

Btw, how comes that Nvidia's GPU with only 24 pipes and lower clock can execute almost as many shader ops as R500 and (which sounds even more weird) have higher fillrate?

 

[Megadrive1988]

since when is PS3 GPU the G70 or based on the G70?

I thought PS3 GPU was the totally new architecture that Nvidia has for PC GPUs for next year.


G70 = NV4x (NV47 or NV48?)

G80 could be NV4x or NV5x (NV48 or NV50)



PS3 would at least be getting a derivative of G80, but only if G80 is indeed a next gen architecture beyond NV4X.

 

[JF_Aidan_Pryde]

Six vertex engines @ 500 Mhz = 3 billion vertex ops/sec or 6 billion if you count scalars.

4 SPEs @ 4.8 Ghz each = 19.2 billion ops/s assuming you can dispatch 1 per clock. Moreover, the SPEs are ALOT more flexible than the vertex shaders, being even more powerful than the Geometry Shaders of Longhorn's WGF2.0.

Are there any fixed function geometry operations which are performed for 'free' on the GPU that would need some extra work on Cell? It seems oldd that two SPEs can handsomely beat six vertex engines. Surely the comparison is not that black and white.

Edit: Acutally, thinking about the die size of two SPEs on 90nm and its frequency, it seems reasonable that it can do the work of about six vertex engines.

 

[DemoCoder]

Btw, how comes that Nvidia's GPU with only 24 pipes and lower clock can execute almost as many shader ops as R500 and (which sounds even more weird) have higher fillrate?

Its # of ALUs x clockrate. The R500 has 48 ALUs. A 24-pipe GF6 architecture would have 24*2=48 ALUs. Although Nvdia's ALU's are less flexible (they can each only do a subset of ops, but 2 of them together are as general as one of XB360's). However, as far as peak rate, we're just assume that shaders include a "friendly" mix of instructions that work good on NVidia's split-ALU architecture.

In anycase, we have 48 ALUs * 500Mhz vs 48 ALUs * 450Mhz. Of course, I didn't count the vertex shaders, which is an addition 6 ALUs * 450Mhz. But the XB360 has the 3 VMX cores @ 3.2 Ghz, which is like having 18 vertex shaders.

 

[Wunderchu]

PS3 Nvidia GPU will deploy more transistors than ALL GPUs + CPUs of current consoles

http://phx.corporate-ir.net/phoenix...eventID=1060901

Nvidia CEO Jen-Hsun Huang said in the recent conference call, that Playstation3's GPU will deploy (contain) more transistors than all the GPUs and all the CPUs in all three of the current consoles.

hmmm!!!

PS2 EE CPU: 13 million
PS2 GS 43 million

Gamecube CPU: 21 million?
Gamecube GPU: 51 million

Xbox CPU: 9 million ?
Xbox GPU: 63 million

200 million transistors for all the console CPUs and GPUs


so Nvidia PS3 GPU will have more. well that isnt saying much since GeForce 6800 is 222 million anyways.

[source: http://www.ga-forum.com/showthread.php?t=47415 ]

 

[one]

direct link to the nvidia conf call

 

[Titanio]

I heard it too, more transistors than all cpus + gpus in all 3 current systems combined. So judging by the above numbers, that'd be > 200m.

I wonder if the Q+A is worth listening to? It's long..

 

[Pugger]

With regard to the 360 GPU what is the general take on what ATI has delivered. I don't fully understnad all the technicalities but it seems from the comments I have read that ATI have given MS a GPU that does things slightly differently from previous. Can someone give a quick lamens guide, as all I'm hearing it GPU is a let down.

 

[nAo]

It seems reasonable that Nvidia's G70 will have 24 pipes @ ~450Mhz, which depending on how you count, allows it to execute up to 42 billion shader ops/s (dual issue vector, dual issue scalar * 24 * 450Mhz) vs XB360's 48 billion.

Shader ops is a concept that shouldn't be used, imho.
Let's talk about floating ops per second:
NV40: 2 ALUs per pipeline, first ALU is capable of a vec4 fmadd per clock, the other one is capable of a vec4 mul per clock, so we have 12 ops per pipeline x 16 pipeline x 425 Mhz = 81.6 Gigaflop/s (+ 20 Gigaflop/s if we factor in vertex shaders too)
R500: 48 ALUs, each ALU is capable of a vec4 and a scalar fmadd, so we have 10 ops per pipeline x 16 x 500 Mhz = 240 GigaFlop/s

. Since CELL is so great at processing vertices, it would make more sense to rip out the vertex shaders and video decoding circuitry, and use the die space for eDRAM or more pipes.

This is what we said a lof times before, but it seems this is not the case...

Moreover, I would expect a PS3 GPU to be on .09um not .11m.

It has been reported by some japanese magazine the GPU is built around a 90 nm process.

Finally, could Sony do something really insane and put SLI'ed GPUs in their console?

Cost? heat? how many memory chips? SLIed GPUs would be cool but it seems a bit off target for a console.

 

[Wunderchu]

With regard to the 360 GPU what is the general take on what ATI has delivered. I don't fully understnad all the technicalities but it seems from the comments I have read that ATI have given MS a GPU that does things slightly differently from previous. Can someone give a quick lamens guide, as all I'm hearing it GPU is a let down.

as I posted in another thread:

The spec seems pretty clear.

- 48 ALUs, each capable of one scalar and one vector operation per clock.
- 8 pixel pipelines, each capable of writing one 32 bit color, and one 32 bit Z per clock, along with Z compare and multi-sample expansion.

Current ATI graphic processors have one ALU per pipeline (one scalar & one vector operation per clock, not two ALUs per pipeline). So a 360 pipeline is just the ending part of what we now call a pixel pipeline. A configuration a lot like how pipelines worked before pixel shaders.

- Old and busted: Add pixel shaders to the pipelines.
- The new hotness: Leave the pipelines alone. Add pixel shading as it's own unit (and lump them together with the vertex shaders).

The 360 graphics processor would make for a fairly mediocre PC graphics card. All the old games (with no pixel shaders) would run the same as an 8 pipeline ATI card. Games with longer pixel shaders would run like a 48 pipeline (if one existed) ATI card. Sounds like a good trade off to me.

One thing I worry about is Z pre-calculation passes, and shadowing. Really any render pass that is fill rate heavy, and light on shader calculations. I wonder if Doom3 would run very well on this chip. I don't think so without a major re-write.

One clarification: the X360 does not have pipes like you speak of them. It just has limits to the tech surrounding the ALUs. That's why it has a theoretical limit of 500 million polys/second: it can only setup 1 tri per cycle. If you devoted every ALU to a simple vertex shader, you could get WAY more polys than that. And the chip is also limited to 8 pixels output per clock, but that's not an indication of the number of ALUs processing pixels.



Quote:
Originally Posted by xpfeil
One thing I worry about is Z pre-calculation passes, and shadowing. Really any render pass that is fill rate heavy, and light on shader calculations. I wonder if Doom3 would run very well on this chip. I don't think so without a major re-write.

Wrong, actually. Writing to the color buffer is turned off for Z passes and stencil operations. If the leak is accurate, the chip can do 16 pixels per cycle in a Z/Stencil only pass. Expand that with anti-aliasing and you have more than enough fillrate.

[source: http://forum.teamxbox.com/showthread.php?t=345091&page=4&pp=15 ]

 

[Apoc]

What is the fillrate of the x360 gpu?

 

[nAo]

8 pixels per clock at 500 Mhz = 4 Gigapixel/s
If we factor in 4x MSAA on compressed tiles it should reach 16 Gigapixel/s

 

[Apoc]

8 pixels per clock at 500 Mhz = 4 Gigapixel/s
If we factor in 4x MSAA on compressed tiles it should reach 16 Gigapixel/s

Only 8 pipes? Having to render at 720p I thought it'd have at least 12 pipes.

 

[Guden Oden]

Only 8 pipes? Having to render at 720p I thought it'd have at least 12 pipes.

Yeah, because everybody knows 4300 screen fills per second is wholly inadequate. I mean, how can we POSSIBLY live with only 72 render passes per frame?! We'll be back to space invaders-level graphics! *weeps pitifully*

 

[Apoc]

Only 8 pipes? Having to render at 720p I thought it'd have at least 12 pipes.

Yeah, because everybody knows 4300 screen fills per second is wholly inadequate. I mean, how can we POSSIBLY live with only 72 render passes per frame?! We'll be back to space invaders-level graphics! *weeps pitifully*

It's not inadequate, but seeing 6800u, knowing that G70 and r520 are gonna have 24 or 32 pipes, and after watching people saying that it 'd be a lot more powerful that pc graphic cards I thought it'd have at least 12 pipes.

 

[Fafalada]

Only 8 pipes? Having to render at 720p I thought it'd have at least 12 pipes.

On what grounds?
720P is only 3x higher then 480P and chips with ~1GPix have done the job quite well at the latter, even though the rasterizers were rather primitive, only capable of 1-2operations per pixel at peak rate, and not extremely efficient at reaching peak rates either.

Console chips don't have the necesity to waste transistors on accelerating legacy software like PC ones do.