Cell outperforms 7800GT in rendering by 5 - 6x!

[Carfax]

Whoa, who'd have thought that a CPU would outperform a GPU in rendering!

Barry Minor has a very interesting blog benchmarking a 3.2 GHz CELL against a 7800GT OC. The test was a Julia fractal raytracer originally developed for GPUs using CG (see here). This code was ported to CELL using the SDK with no specific optimizations and it was found that the 3.2 GHz DD3.1 CELL was about 30% faster than the 7800 GT OC. Further improvements yielded much higher performance:


To our surprise, well not really, we found that using only 7 SPEs for rendering a 3.2 GHz Cell chip could out run an Nvidia 7800 GT OC card at this task by about 30%. We reserved one SPE for the image compression and delivery task. Furthermore the way CG structures it SIMD computation is inefficient as it causes large percentages of the code to execute in scalar mode. This is due to the way they structure their vector data, AOS vs SOA. By converting this CG shader from AOS to SOA form, SIMD utilization was much higher which resulted in Cell out performing the Nvidia 7800 by a factor of 5 - 6x using only 7 SPEs for rendering.

The full specifications of the test machine can be found in a response posted in the blog comments:

I used a UP Cell bringup system. 3.2 GHz DD3.1 Cell processor with 8 good SPEs, 512 MB XDR Memory, 100 Mb network. Pixels were 128 bit, 32 bit float per color channel. All rendering parameters were the defaults set by the Cg program with the exception of the window size which was increased to 1024×1024.

To read the actual blog, go HERE

This is amazing! Man, Cell will make a wicked gaming CPU if it has this much horse power.

PS3 might be more powerful than we expect. If Cell really takes off, we may see them in some PCs hopefully

[ahmad]

This is just testing pure computational power... A cell could easily annihilate many many systems in terms of pure number crunching applications. I don't mean to sound bitter, but there are a lot of things that need to be accounted for when claiming that the cell is faster than the 7800GT.

Let me ask you this: why would Sony spend so much money on a GPU (Nvidia RSX) if the CPU is so much more powerful than any GPU nvidia could provide?

I, myself, do think I partly know the answer to that question. You have to understand that programming/writing games is not all about number crunching.

EDIT: Why the fu.. I mean heck is he adding vertex + pixel shading power to compute number of flops for the 7800, he should be comparing that to the VS rating which is 34 (and about 5-6x slower than Sony's 200Gflop rating)

[Carfax]

Let me ask you this: why would Sony spend so much money on a GPU (Nvidia RSX) if the CPU is so much more powerful than any GPU nvidia could provide?

Thats a good question. I'd agree with you that gaming is more than just pure number crunching/pixel pushing power, and the Cell cannot replace a GPU in every respect.

I've also heard that the GPU is there mostly for compatability/features, and not for performance reasons.

However, I don't think the guy who did the benchmark set out to prove that.

The benchmark shows that the Cell processor can outperform a GPU in heavy number crunching, whilst still retaining the advantage of a general purpose processor in that it can be made to do anything.

No general purpose processor could claim that, until the Cell came.

That to me is whats amazing!

[Starscream]

It proves that Cell is great at number crunching so for certains erver application it would b a great CPU.
Wich ofcourse will mean that If the Cell CPu hits it of in this market Sony+IBm will have an extra market so Extra CPus wil b made etc etc.
so to b short the Cell being a hit in the server market could lower the manufacturing price ornot? wich should lower the manufacturing price on the PS3.

I thought i read that the RSX GPu in the PS3 is allowed to use one of the Cell cores in the PS3 to help it out.
in certain games this could give the GPU a huge performence boost.

[koosjebram]

Now if they only write programs for it!!!!!!
We even haven't got proper dualcore optimisations.
All this potential but there is not much that can be done with it in real life.

[MaxxxRacer]

From what I gathered the rendering that was done was infact a single frame that was created in a program akin to 3dsmax or maya. the way that scenes are rendered in 3dsmax, maya and other 3d animation programs is completely different than the way games are rendered. Its not even a remote comparison.

More recently 3dsmax and maya have live rendering for the scence that is being worked on, but the rendering is of much lower quality than the final product and requires an incredibly powerful gpu to get good visual quality at a usable FPS.

[[XC] leviathan18]

we sony are going to use a 200gigaflops CPU with 8 cores XDR and we dont need GPU cuz we have the best of the best

2 years later well nvidia is going to help us with the gpu

1 year after that well aegis is going to help us with the PPU

so now you have 8 wasted cores inside a console...

[MaxxxRacer]

the conclusion to what lev said is simple.

Sony is full of S&!T and always have been. PS2 was supposed to be the most powerful thing the world has ever seen... in actuality the performance was underwhelming.

[[XC] leviathan18]

the conclusion to what lev said is simple.

Sony is full of S&!T and always have been. PS2 was supposed to be the most powerful thing the world has ever seen... in actuality the performance was underwhelming.

LOL indeed they said ps2 was going to have 10000 times more power than any pc even the xbox with a celeron 700 was more powerful in any aspect

instead MS went the easy road a console so similar to a pc that is a charm to program in it, as always the PS is the more difficult thing to program

[MaxxxRacer]

ya..

as much as we all despise MS, they really do know what they are doing regarding consoles. I've talked to a few game programmers who have worked for ps2 and xbox game developers and NONE of them prefer ps2 development becuase sony treats game developers like garbage while MS is much more kind to them.

BUT I wouldnt call the X360 a charm to code for. It will be more difficult to code due to the change of cpu from a PC / Xbox. BUT MS provides incredibly good developers tools that make it WAY easier to code for than it would be without them.

[[XC] leviathan18]

well still is power pc so it isnt that different all the developers tool were in a g5

also MS use direct x libraries so is WAY more easy to program for xbox than ps2


on topic this news is what i call misleading we dont want to see crunch power i want to see how it renders a complete game i dont think it would be near the 7800gt

[trakslacker]

its an impressive figure, but thats really all it is. AFAIK it has little bearing on real-world performance. Heck, I love my PS2, and plan on getting a PS3 after the price drops a bit from launch. And while the PS3 certainly looks promising, I'll withhold final judgement until a see some real-world gaming performance.

[Mr. Tinker]

benchmark the Cell vs the 7800gt in a DirectX or OpenGL render.

[angra]

Interesting post. Since the blog doesn't seem to post the specific Cg code that he ported, I can only make some educated guesses. My line of work is closely related to this question so I have some background.

First off, to the poster who said that "he" should have been using the vertex shader FLOPS count - I believe that that is not correct. The fragment processor (or pixel shader) is the part of the GPU that is typically used for GPGPU operations. Some people are using the vertex processor, but it's just not as commonly done.

The current crop of nVidia cards have a bizarre issue that hurts their top end sustainable flop-count. When using a pixel format that represents 4 single-precision floating point values per pixel, pixel fill rates drop dramatically. I believe that this is the manifestation of a hardware "bug". This is toally irrelevant for games, so I don't think it's been a high priority to correct. ATI cards do not have this problem, and the most recent generation of nvidia chips mitigated this problem somewhat, so I kind of expect this problem to be totally fixed on the next chip. This issue along could account for poorer than expected performance for the GPU, as they mention SIMD frequently in the post. (multiple data elements per pixel is one axis of SIMD that is exploited in GPGPU).

Some of the comments in the blog post indicated that maybe the Cg code had dynamic branching, which would disproportionately hurt the current generation of GPUs, especially the nvidia ones.

Maybe I missed it, but I sure which there was a link in the post to at least the GPU portion of the test code.

The Cell is a great architecture in its own right, as are the nvidia and ATI chips. As a technologist and scientist, I find all of this sniping between their marketing departments really annoying. They all have plenty going for themselves without needing to cherry pick functional kernels that make the other guys look bad.

At this year's High Performance Embedded Computing Workshop, there was a pretty hefty amount of (relatively polite, compared to stuff like this) sniping between the Cell contingent and the GPGPU contingent. The Cell guys showed a slick demo of a real-time terrain rendering engine, and the GPGPU guys stood up in front of the everyone and said, more or less, "is that all you can do, punk?" Well, ok, I guess that was pretty OK. Academic style conferences have too little competition anyway.

[[XC] leviathan18]

we need more post like yours

[Carfax]

Yeah great post Angra. It's good that we have industry professionals here that can make sense of all this

So Angra, based on what you know of Cell, how do you think it would perform using an OpenGL or Direct X render?

[Mikesta]

Everybody Loves Hypno-Toa......Aghh Angra.

On a serious note nice post. On Cell vs GPU, even though the tests are most probably concocted to make the Cell look better then it is, it is nice to know that powerful rendering is now possible on something other then a GPU. Look at 3DMark0X and compare the CPU test's vs a proper GPU. CPU's suck hard and would still suck hard even if they took full advantage of Dual or even Quad core x86 Processors.

However with the Cell in the PS3, it's job isn't to render anyway. It has the Nvidia chip to do it. What I like about the Cell is its raw number crunching potential. My reason for getting a PS3 is in the hope that you'll be able to run Linux on it and F@H with it.