Cell benchmarked

[Titanio]

Please. Stop pretending that most people can correctly read a benchmark spreadsheet. Most will gravitate to the speed up factor column and conclude that Cell is 20x-30x better than what they have. And that's exactly what IBM wants, which should cast a revealing light on the paper.

Well more fool them. That 30x is a real speedup - in that application.

Think also as to whom these docs are aimed at. They're aimed at the community, a technical audience. They're not really aimed at gamers or the like.

 

[Jawed]

Looking at those benchmarks CELL clearly is ideal for algorithms that require two things over a standard processor, and that's high bandwidth + high computation, or in otherwords, algorithms that have need a lot of work done on large linear data sets (large matrices).

Yes, that's why it's called the Broadband Engine.

CELL is perfectly suited as a game console processor, much more so than a traditional general purpose processor.

Nowhere in that article is there evidence of Cell's suitability for a game console processor.

The Cell BE is primarily targeted for game applications some of which demand a high video processing capability. Consequently, significant effort was devoted to optimizing MPEG-2 video decoding.

That would appear to be justification for hours of FMV in PS3 games. It would seem that certain E3 videos were more prescient than we thought: that PS3 games will be short bursts of interactive action in-between long periods of FMV

Jawed

 

[hongcho]

Did anyone notice that in the article the Cell BE they used has 8 SPUs?

Hong.

 

[Shifty Geezer]

Did anyone notice that in the article the Cell BE they used has 8 SPUs?.

Yes, because that's the format of the current hardware. PS3's Cell is also a 1:8 config, but one of those won't be being used either because its broke, or to preserve uniformity between broke and working processors.

 

[Edge]

Did anyone notice that in the article the Cell BE they used has 8 SPUs?

You have not read the thread, as that has already been discussed.

 

[Edge]

Yes, that's why it's called the Broadband Engine.

Nowhere in that article is there evidence of Cell's suitability for a game console processor.

So a Broadband Engine is not good for a game console?

 

[Inane_Dork]

IBM released the information as is, and if someone reads it and jumps to the wrong conclusion because they don't understand most of the information in the article, then that is hardly IBM's fault.

Right. Misunderstandings are ALWAYS the listener's fault.

I claim that IBM's goal is to get people interested in Cell primarily and to put forth truth secondarily (if it was on the list at all). From there it's obvious which will pull rank when it comes to phrasing, wording, editing, graphing, etc.

Well more fool them. That 30x is a real speedup - in that application.

I guess I should exempt all Terrain Render Engine companies from what I said.

Think also as to whom these docs are aimed at. They're aimed at the community, a technical audience. They're not really aimed at gamers or the like.

I'm not talking about gamers or individuals. I'm talking about companies where the decision making process is incredibly often started or bottlenecked at people who can't truly read that article. Just because it's a technical audience doesn't make the audience insusceptible to PR. I bet up to half of MS couldn't understand that article more than the concluding chart.

 

[Bobbler]

So a Broadband Engine is not good for a game console?

I think thats the million dollar question at the moment. We know it will work well, obviously (from the little we've seen it shows it should work at very least acceptably enough that it shouldn't be a detriment)... but what we don't know is if the Cell's actual power can be leveraged while avoiding the things its not good at, in a game. Can dev's get the kind of 10x increase over a typical processor using the Cell for games? That's the real question, and I don't think anyone here has the answer (at least none of those that can talk).

I'm willing to bet that it's likely not the case (where Cell will be 10x vs a comparable CPU)... there are certain things the Cell should be fantastic at that games should be doing, but there are also things mixed in that the Cell won't be so hot at. Game's don't consist entirely of doing thousands of 1m by 1m matrix operations or video decoding/encoding, etc, etc -- the more developers are able to convert their algorithms and ideas into a form that the Cell will rock at the better, but there is some stuff it won't rock at that games include that these benchmarks don't show. I think we'll be surprised in some cases though.

 

[Titanio]

I'm willing to bet that it's likely not the case (where Cell will be 10x vs a comparable CPU)... there are certain things the Cell should be fantastic at that games should be doing, but there are also things mixed in that the Cell won't be so hot at. Game's don't consist entirely of doing thousands of 1m by 1m matrix operations or video decoding/encoding, etc, etc -- the more developers are able to convert their algorithms and ideas into a form that the Cell will rock at the better, but there is some stuff it won't rock at that games include that these benchmarks don't show. I think we'll be surprised in some cases though.

First off, I don't think anyone realistically expects a general 10x speedup in games over competing processors. Second, a related question - not all tasks are created equally, so if you're optimising for a game, what do you optimise for? Does it matter if the things you are "not so hot at" are only taking a small percentage of CPU time (and can, for example, be done pretty trivially on the PPE)? (That's not to suggest that everything Cell isn't good at is trivial, of course..although there was a slight whiff of that in some dev comments sofar).

I guess I'm saying, we can't just consider the numbers of tasks Cell is good or bad at, we need to consider their relative importance. If you had 10 tasks to do per frame in a game, and I told you I had a chip that was really brilliant at one of them, and not so good at all at the rest, would that be telling you enough to say if that chip was good or bad? No. If that task it was really good at was taking 90% of the frametime, then it would be good, but if it were taking 1%, it wouldn't be so good.

I think we should also move away from the notion that Cell is either going to really outperform, or underperform, competing chips. There is a range there, there'll be things Cell is a really excellent at, very good at, average at, poorer at etc. I dunno, I just get the notion that a lot of people are thinking in extremes all the time.

 

[Bobbler]

Oh, I agree with you Titanio -- I wasn't really trying to answer the question, but just point out that these Benchmarks really don't relate to game performance.

I mostly didn't want someone answering him with a "yes" or a "no", because its simply a question we don't have an answer to outside of "it works." And, frankly, I'm happy with "it works," but I know a lot of others aren't.

 

[mckmas8808]

Just because it's a technical audience doesn't make the audience insusceptible to PR. I bet up to half of MS couldn't understand that article more than the concluding chart.

Why do you see this as PR? If the numbers went so hot would it be PR then? And if half of MS couldn't understand that article then I guess that $250,000 per programmer is going to waste. Must be why my OS keeps crashing.

 

[Edge]

Oh, I agree with you Titanio -- I wasn't really trying to answer the question, but just point out that these Benchmarks really don't relate to game performance.

Nonsense! Maybe you should try reading the section "Optimization of Triangle Transform and Lighting".

The other benchmarks clearly show a processor that allows HUGE performance advantages in processing large matrices over a traditional CPU like the XeCPU, and your saying that is not gaming related?

 

[Jawed]

Games stopped using the CPU for vertex shading ages ago. GPUs do a far better job. Of course, that's from the perspective of a console which has a clear-cut CPU-GPU split.

Jawed

 

[Titanio]

Games stopped using the CPU for vertex shading ages ago. GPUs do a far better job.

Well Cell ain't bad at it at all, and if you happened to be bound by your vertex shading it might make sense to leverage the CPU if possible.

 

[macabre]

If there are any open questions about how a single cell does the Terrain demo:

Barry_Minor

I was the lead developer on a team of three that wrote the IBM Cell terrain ray-caster that you mentioned. While Daniel Cohen did develop ray-casters on IBM PVS HW back in the 90s our effort was completely independent from his work.

Our work has shown that one 3.2 GHz Cell processor can ray-cast and compress 1280x720 super-sampled images at frame rates better than 30 frames/sec without the assist of a GPU. We start with a 16 bit (5/6/5) texture and a 16 bit digital elevation map and render into a 128 bit (32 bit float/channel) frame buffer. The atmosphere is procedurally generated on the fly with multiple octaves of Perlin noise and the finished images are JPEG "like" compressed and delivered by Cell all in real-time.

http://www-128.ibm.com/developerwor...reeDisplayType=threadmode1&forum=739#13762062

 

[Jawed]

Well Cell ain't bad at it at all.

Compared to what? A GPU? I'd be interested in comparative data for a GPU if you have some.

Jawed

 

[ralexand]

Games stopped using the CPU for vertex shading ages ago. GPUs do a far better job. Of course, that's from the perspective of a console which has a clear-cut CPU-GPU split.

Jawed

The question I have though is can cell and gpu be used in conjunction with each other that will lead to a synergistic increase in performance.

 

[czekon]

so it looks like cell is a real monster. Does xeCPU have any chance to compete with Cell??

 

[Titanio]

Compared to what? A GPU? I'd be interested in comparative data for a GPU if you have some.

Well do we know the paper peak transform rate of a 7800GTX, for example? Or better, do we have a benchmark? The benched peak of one SPU is >200m verts/sec.

I'm not saying Cell would make a good replacement for vertex shading on the GPU, but you would think it would speed it up to do some in parallel with the GPU if possible. If you happened to be in a position where your vertex shading was the bottleneck, any help would have an impact. It's not about favouring the CPU to do this work instead of the GPU, doing it on both. Cell wouldn't have to be better than a GPU for that to be of benefit.

 

[Shifty Geezer]

The question I have though is can cell and gpu be used in conjunction with each other that will lead to a synergistic increase in performance.

That in theory is what the Cell<>RSX interconnect is for, though I don't know how it'll be used. As has been said, why transform verts on Cell if the GPU can do it? Though I guess like Xenos has a prepass, perhaps you could use Cell to calculate some vertex parameters as a precursor to filter what's sent to the GPU, or say to calculate shadows in tandem with the GPU calculating pixel shaders and composite them over the top?