Cell benchmarked

[Shifty Geezer]

PC-Engine likes to brag that the Cell doesn't have a double precision floating point capabilities, or that it's diminished in some way, yet the article points out a test using Linpack for the DP floating point benchmark. So the Cell is capable of DP-FP work?

The current SPE's perform DP float work at 1/10th their SP performance (which also takes shortcuts to be faster, so isn't as precise as other SP float computers).

This is also quite an atagonistic question the way you've worded it. Better just to ask about single and float performance, and if you've got arguments to amke in that other thread, leave them there.

 

[drpepper]

Cell's DP power is inferior to it's SP power, that much is true. But it's DP power is still quite competetive with other contemporary processors. Quite good actually.

Yes, it looks like 156 vs 9.67 GFlops for 8 SPEs (Table 13). Are they comparing it to an intel chip for that comparison? "3.6GHz IA32/SSE3"

 

[scooby_dooby]

So what's IBM waiting for? Why don't they release CELL into the wild and let 3rd parties benchmark this thing? It just seems for like a year now, they've just released one extremely good benchmark after another, when are we gonna see real benchmarks from non-biased sources?

Just wondering...

 

[Titanio]

So what's IBM waiting for? Why don't they release CELL into the wild and let 3rd parties benchmark this thing? It just seems for like a year now, they've just released one extremely good benchmark after another, when are we gonna see real benchmarks from non-biased sources?

If you'd like to call up Mercury Systems and buy one, we'd be glad to hear your results

There's nothing "not real" about these benchmarks. They're real benches. Whether they're selectively chosen or whatever is another matter.

You could also code your own apps on the simulator and bench performance that way (it's supposedly cycle-accurate - but take into account the margin of error discussed in this paper too).

 

[scooby_dooby]

I think you know what I'm saying.

 

[London Geezer]

I think you know what I'm saying.

I don't... really.

 

[scooby_dooby]

I'm saying enough with the controlled IBM benchmarks, if they are so confident in this CPU then releases to 3rd parties to to their own independant benchmarking, no strings attached.

The answer is probably because we would see some not so great results in some areas.

I've just never seen such a stream of information to make a case for a piece of hardware before, it's been almost a year since ISSCC and still more benchmarks from IBM.

It seems a little funny to me, and annoying to say the least. I would like to see some more balanced or 'real' banckmarking out of interests sake, but hey that's just me, I don't know about you guys....

 

[London Geezer]

Heh, you would think that. You'll get your benchies, obviously now it's too early to release the whole boards to third parties.

Did we ever see benchies for Xenos or Xenon? At least IBM are showing something this time. Eventually we'll get more info.

 

[Titanio]

I'm saying enough with the controlled IBM benchmarks, if they are so confident in this CPU then releases to 3rd parties to to their own independant benchmarking, no strings attached.

The answer is probably because we would see some not so great results in some areas.

I've just never seen such a stream of information to make a case for a piece of hardware before, it's been almost a year since ISSCC and still more benchmarks from IBM.

It seems a little funny to me, and annoying to say the least. I would like to see some more balanced or 'real' banckmarking out of interests sake, but hey that's just me, I don't know about you guys....

Would you rather have no data, as is usually the case? I don't think we've ever had so much insight into a custom chip for a games console before, like this.

This data is fine as it stands. Unless you feel people are taking it that this applies generally across any benchmark/application - and I don't think they are - I'm not sure what you're worried about.

 

[Carl B]

I think you know what I'm saying.

I know what you're saying Scooby, but I don't think we'll get a lot of independent benchmarks until PS3 becomes prolific and Linux is released on it. Until then the fact is only people with Cell will be able to properly bench it, and right now not many people have Cell...

I'm also looking forward to more benchmarks, and by people other than IBM. But until then, this is a great step forward in info from where we were before.

 

[Shifty Geezer]

Cell is in the wild, if you want to buy a system. Nothing stopping anyone buying a Cell blade and writing their own software. It's not like IBM are deliberately (TTBOMK) withholding Cell's from customers on the grounds they only want good benchmarks. And those that do release 'benchmarks', like Mercury or Alias's cloth dynamic, are going to show a performance increase because that's why their using Cell. Mercury won't buy Cells and code for them if there wasn't to be an advantage. Hence you're no likely to see benchmarks of office software running on SPU's until Cell's are common as muck and people write such benchmarks for fun.

I'll add too that I'm pleased we've got this level of noise. We have absolutely nothing on XeCPU to compare bar those leaked Alpha-kit comparisons. It'd be nice to see how XeCPU's VMX units are at being fed data, perhaps running an optimised terrain demo.

 

[DarkRage]

And those that do release 'benchmarks', like Mercury or Alias's cloth dynamic, are going to show a performance increase because that's why their using Cell. Mercury won't buy Cells and code for them if there wasn't to be an advantage.

I agree and disagree with it.

I agree Mercury wouldn't go with Cell if they were not getting any advantage. Cell has to be great for the tasks they are using.

However, I disagree that Mercury saying "wow, 100 times faster, we are the best" is indicative of the real performance of Cell.

 

[Shifty Geezer]

Well what else is that 100x? How can it be unreal, unless Mercury are faking their results? I'm happier with the notion what they are doing is optimal for Cell's architecture and hence the large improvement, rather than they are artificial inflating Cell's performance or misrepresenting the performance of their existing products.

 

[Edge]

But I am not that confident with those benchmarks of x40 times faster (or x4 with a single SPU) with no code to compare with and no information regarding optimizations in the other platform.

That could be argued both ways. You can optimize your algorithms till the end of time. Who says you can't squeeze even more performance out of CELL, by optimizing some more. In end, we most likely will not know how well things were optimized on either side.

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

This result are not so surprising, as the specs have pointed this out quite clearly, and has been advocated by quite a number of people on this forum for a long time. Thus also proving that specs along with an understanding of the architecture are far from meaningless. I say that, because many people on this forum have downplayed CELL's specs for a long time now.

 

[Asher]

Seeing how both the G5 and Cell are made by IBM, why would they make the test "unfair" with their own hardware? It's not like the 2 chips are competing, one is for consoles.

You just answered your own question. They don't compete for the same markets, so why not show all usecases where you can show good performance compared to "other" processors if it won't hurt the sale of your other processor?

 

[Asher]

cryptography, protein folding, weather mapping/prediction, financial modeling...

IMO their double precision performance needs to be improved before they're used in most of the above situations. It's fine for a game console, but when you start doing scientific research you probably don't want 32-bit floats used.

Not to mention that Cell's FP output isn't IEEE-compliant...

 

[ralexand]

Since this is the console forum can we talk about what these benches mean for the chances of cell enhancing games ie., can developers create higher poly games due to the TnL performance, the ability to map and playback mpeg on texture surfaces etc.

 

[ADEX]

Quote:

cryptography, protein folding, weather mapping/prediction, financial modeling...

IMO their double precision performance needs to be improved before they're used in most of the above situations. It's fine for a game console, but when you start doing scientific research you probably don't want 32-bit floats used.

It depends, research can make use of all sorts of data types and not that much needs full 64 bit FP. One of the top 10 supercomputers spends it's time processing 8 bit integer data.

Not to mention that Cell's FP output isn't IEEE-compliant...

The 32bit stuff has limited rounding modes, 64bit is fully IEEE.

--

BTW I knew about the terrain demo numbers months ago. IBM were not expecting that level of speed up.

The 970's memory bandwidth can't be compared directly since part of the bandwidth (10% ?) is taken up by signalling and it's evenly divided half read half write. Even at peak it'll only be getting about 1/5 of the Cell. I am guessing the data set didn't fit in the cache and it kept getting flushed, this is a lot less likely to happen in a local store. The Cell can also do streaming which reduces memory access dramatically (I don't know if it was done in this case though).

As for how optimised it is, aside from the multicore aspect, programming VMX and SPEs is very, very similar. They probably just ported the VMX code to the SPE. The SPEs do allow extra optimisation though (extra loop unrolling) so they would have used that.

These benchmarks are certainly fair, if you look at the encryption stuff several show the SPEs running *slower* than whatever it's being compared against.

 

[Inane_Dork]

Would you rather have no data, as is usually the case? I don't think we've ever had so much insight into a custom chip for a games console before, like this.

I cannot speak for him, of course, but I would rather have no data. I prefer it that way for the same reason that your second sentence misses the point: this is as much information as it is disinformation. What's the "insight"? According to MS studies, XP is more stable and secure than any other OS. Does that qualify as "insight" now? IBM didn't even bother with a sponsored third party. They did it all themselves.

No data right now is preferable. We'll get practical information leaked as time goes on, and that information is more worthwhile (though not above question).

This data is fine as it stands. Unless you feel people are taking it that this applies generally across any benchmark/application - and I don't think they are - I'm not sure what you're worried about.

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.

 

[Mordecaii]

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.