Mercury Cell presentation (new performance comparison with FFTs, image filters)

[chachi]

Of course they would try to show CELL in the best light possible, otherwise why would you buy their shiny new CELL machine (that they obviously spent money/time/effort developing) instead of some dull old Intel/AMD/PowerPC box that you can get from anyone?

Exactly, this is their new baby and they're trying to sell it. We don't know how the new system compares to alternatives for the given task (tracking targets or something in a C&C vehicle wasn't it?) or what the comparable costs are for the systems, but the bamboozle factor on the slides is high. It could be that they really are 108 times better than the competition or whatever but I don't believe 100% of anything published by non-independant sources.

And FWIW, I'm a Cell advocate - I think for a lot of things it'll be better than Xenon or PX or XCPU or whatever we're calling it this week. I just think these slides should be taken with a grain of salt. Unrealistic expectations leading to backlash when independent benchmarks come out with not so rosy results are dangerous to the success of Cell (not necessarily the PS3, but Cell has a chance to be bigger than that).

 

[randycat99]

For the skeptical ones, do the numbers for the competitor platforms really look "low"? How much higher would you expect them to go, accounting for a "lifting away" of the conspiratorial layer that Mercury might have on this?

 

[LunchBox]

What does FFT stand for?

 

[rounin]

FFT ?

Anyways I doubt they would fake the results. Isn't FFT (and very much optimized for P4s / AMD cpus) done in Prime95? If they were false/tweaked down then they would be called out for shens in no time...

 

[Bohdy]

Not surprising. We have already seen that Cell is well suited to FFT.

 

[LunchBox]

FFT ?

Anyways I doubt they would fake the results. Isn't FFT (and very much optimized for P4s / AMD cpus) done in Prime95? If they were false/tweaked down then they would be called out for shens in no time...

Thank you for answering my query.

 

[aaaaa00]

FFT ?

Anyways I doubt they would fake the results. Isn't FFT (and very much optimized for P4s / AMD cpus) done in Prime95? If they were false/tweaked down then they would be called out for shens in no time...

I'm sure the results are tweaked to show CELL in the best light possible, but they're probably not outright fakes.

The question is, does a benchmark result running FFTs accurately predict performance running other code?

 

[DarkRage]

Exciting. FFT running on Cell. Again.

Pity I don't want to run just FFTs. Games are not using loads of FFTs, AFAIK.

 

[Rodéric]

Pity I don't want to run just FFTs. Games are not using loads of FFTs, AFAIK.

Correct, but what if they aren't using them because the CPUs can't handle them ?
...

 

[DarkRage]

Correct, but what if they aren't using them because the CPUs can't handle them ?
...

Yeah, maybe.

What use can they have in games?

Some picture filtering on what possibly can be done much cheaper with similar results, and audio processing capabilities which are great nowaday with very low CPU usage?

Any games developers planning to rely heavily on FFTs?

I see the point for picture improvement. PS3 could be great at picture quality when working with films. I wonder if Sony is going to allow to have such quality in a cheap console, killing their top-of-line BR players.

PS3 has to be amazing running SETI@home. Maybe that was what Sony was suggesting on E3 with the outter space slides ;-)

 

[Rolf N]

Correct, but what if they aren't using them because the CPUs can't handle them ?
...

I think a better reason for not using them in games is because games tend to make snapshots of dynamic environments to form a simulation. An FFT is useful for pretty much the opposite: to deconstruct/analyse a serious of snapshots.

As an analogy, compressing sound into MP3 format most certainly asks for an FFT. Playing the thing back doesn't.

In all likelihood there are some fancy sound effect algorithms that rely on FFTs, but you'll never get that to be even close to the majority of processing load in a game IMO.

 

[Shifty Geezer]

The question is, does a benchmark result running FFTs accurately predict performance running other code?

Why would Mercury care about that when their processes use FFTs? This whole topic is about Cell performance in activities it was designed to do well at. First you say the numbers can't be trusted (what about the medical imaging example a while back? Did they slow down their current PPC based imaging system that their customers currently use so they can show Cell outperforming it such a huge degree?) and then you start asking how this reflects Cell's performance in other code, which isn't the point of the thread. No-one is saying Cell is 10x faster than every processor at every activity. Mercury are saying that they have built medical imaging systems for years, and switching to a Cell based platform that's ideally suited for these tasks in a way conventional processors aren't, the improvements are pronounced. If the improvements in Cell performance aren't that pronounced, if they're in the realm of 2-3x, do you think Mercury would rewrite and slow down their existing code to show Cell being 10x faster, when the market their aiming for would happily buy Cell at 3x the speed of current solutions anyway? And as has been mentioned the figures, for Intel processors at least, come from Intel.

Exactly, this is their new baby and they're trying to sell it. We don't know how the new system compares to alternatives for the given task (tracking targets or something in a C&C vehicle wasn't it?)

Except according to Titanio the info on the Intel processors comes from Intel, so unless the Cell performance figures are completely made up (the military will be very pleased if a company has outright lied to them about performance...) they can be trusted.

I don't believe 100% of anything published by non-independant sources.

What constitutes independent though? Almost all reports and articles have an agenda or bias from the part of the author or editorial team. Mercury is not in the pocket of STI, so are independent in that respect. And if they want to fiddle numbers to sell more stuff and make more money, if performance is in reality 3x the alternatives that alone is enough to warrant upgrades. They don't need to fake it to 10x the performance to get the sales and piss off the military who, I would have thought, are likely to run their own benchmarks too. There's some very educated technicians on the forces.

 

[Shifty Geezer]

Yeah, maybe.
What use can they have in games?

They should be useful for EyeToy. Dunno about anywhere else.

 

[Titanio]

Audio, maybe? Perhaps audio-in (voice) also..

I remember from Devstation, one of the presentations was on a FFT-based water simulation.

There was a thread here in General Discussion about fourier transforms and their power/applications (perhaps for physics in particular, which would tie in with the Devstation water simulation - although the suggestion in this thread seems to be that the power required would be very large):

http://www.beyond3d.com/forum/showthread.php?t=28299

 

[aaaaa00]

n/m

 

[randycat99]

So what point are we trying to convey here?...Mercury shouldn't be too concerned about FFT performance on Cell because FFT's are not on the forefront of videogame kinds of functions? I don't get it- is Mercury also signed on for some super PS3 game I haven't heard of? I thought their primary concern was for medical imaging and military uses? Are there people here who are unable to conjure the importance of professional caliber FFT and related mathematical tasks for the applications Mercury is targeting? The task focus of their benchmarks seem pretty reasonable to me. Videogames aren't the center of the world are they?

For those that really need a justification for something outside of "FFT", consider that FFT work is generally considered a sophisticated and hardware intensive task. Something that can run that and run it well is pretty much in good company in the great continuum of CPU prowess. Lesser and simpler tasks will likely run just as well or even better (by virtue of be more simple than a macro-esque FFT task). Videogame functions should fit there quite nicely. So rest assured, Cell's got your back.

 

[stepz]

Other processor numbers look ok, you can check the results for a 3.6GHz P4 @ http://www.fftw.org/speed/Pentium4-3.60GHz-icc/
Look for single-precision complex 1d transform, it's 8GF @ 1k and 6GF @ 64k just as the graphs say.

I wonder why they didn't have x86 results for filter kernels. Maybe an off-the-shelf CPU scoring so close to their DSP on steroids times eight would have been embarassing, on their on turf nonetheless.

 

[Shifty Geezer]

I wonder why they didn't have x86 results for filter kernels. Maybe an off-the-shelf CPU scoring so close to their DSP on steroids times eight would have been embarassing, on their on turf nonetheless.

Reason could be that no-one has an optimised filter kernal algorithm available for x86, whereas Mercury have that for PPC as they developed it themselves . Performance bottlenecks will mostly be RAM BW and internal space. Cell should have the adavantage on both. For a 9x9 16 bit matrix you're needing 81 registers or else you're tied to L1, which will be slower than SPE with 128 registers. External BW should be higher on Cell too, plus LS provides larger, nearer datasets than image data held in L2. x86 should be around PPC performance (faster for higher clock of course) and I'll be very surprised if not.

 

[Guden Oden]

Great posts, Shifty. I really enjoyed reading them, but can't currently "green" you so you'll have to do without. Thanks!

 

[stepz]

Reason could be that no-one has an optimised filter kernal algorithm available for x86,

Intel IPP should be pretty optimised.

x86 should be around PPC performance (faster for higher clock of course) and I'll be very surprised if not.

I agree that performance per clock should be close to PPC. That would put the P4 about a factor of 4 away from Cell. (for the 15x15 case) Make that a dual-core processor and we're within a factor of 2. With Conroe coming out this summer with greatly improved SIMD capability I wouldn't be surprised if Cell's performance is nearly matched by it. And that's in a highly parallelisable non branchy unrollable predictable workload that should be the Cell's forté.

Not trying to downplay Cell here, but if these numbers are anywhere near to correct, Cell's usefulness outside of really niche markets seems pretty questionable.