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

Titanio

Legend
Mercury Systems had a webcast presentation today titled "Deploying the Cell Broadband Engine Processor in Military Ground Vehicles", talks about their Cell-based PowerBlock 200. I didn't have time to listen to the whole presentation, but I skipped to the bits about performance, and thought some here might find that part of the presentation interesting. Here are their slides:

33ue.jpg


51th.jpg


68jg.jpg


And finally a wrap-up with some image filter comparisons thrown in..

78mn.jpg


We've seen FFT benchmarks before, but this compares with a broader selection of chips. The whole presentation is here:

http://www.mc.com/mediacenter/events/webinar.cfm
 
Asher said:
"simulated"?

Done on the simulator. My guess is they don't have 3Ghz Cells lying around (though they mention with the 64K FFT it was implemented on the hardware itself, on 2.4Ghz hardware, with performance scaled to 3Ghz). The product itself isn't going to be available till early next year, so they've plenty of time to get those 3Ghz Cells ;)
 
if they really just upscaled the results, then they should have included the results of the 2.4 ghz version
 
rounin said:
Very impressive but not totally unexpected by now :LOL:
Mercury sells Cell systems (say that 3 times fast, or something) so take this with a grain of salt. I'd be surprised if Cell didn't have some advantages but this seems like a PPT presentation that might be used to bamboozle tech-unsavvy military procurement agents. "108 times faster?! Cut the check!"
 
People what's the doubt about? FFTs are like Cell's 'thing.' Or do we all forget last year's Barcelona conference so quickly? On the contrary - what I take away from this isn't a sense of being impressed - it's more like, show me something other than FFTs! :)
 
chachi said:
Mercury sells Cell systems (say that 3 times fast, or something) so take this with a grain of salt. I'd be surprised if Cell didn't have some advantages but this seems like a PPT presentation that might be used to bamboozle tech-unsavvy military procurement agents. "108 times faster?! Cut the check!"

And what do you think the reason is that they are selling it for?

"C'mon guys, lets fool the military!"

:rolleyes:
 
chachi said:
Mercury sells Cell systems (say that 3 times fast, or something)
Mercury sell Computer systems. They are not an STI company, and not affiliated to Cell in any way, other than they benchmarked the processor and found it was better for their uses than the alternatives.

Though I guess there is the possibility that as a new line of processors they can sell new systems, convincing owners of current systems that they'd want to upgrade. However, given the real demos we've seen I see no reason to doubt these figures. It's not like they're over-the-top figures. Cell isn't reported as 100x faster in every way. If you want to pull the wool over people's eyes, may as well do a good job of it!
 
The BE is still four PPE's isn't it? I think that's what the patent showed. Does that also mean that each PPE has 8(?) SPE's with it? No wonder the numbers are so bloody high!
 
slider said:
The BE is still four PPE's isn't it? I think that's what the patent showed. Does that also mean that each PPE has 8(?) SPE's with it? No wonder the numbers are so bloody high!

broadband engine is one cell.

failed to see one "single SPE benchmarks" ?.... or the GFLOPS figure on the left? :) 100Gflops is half of what one Cell is "announced" to be capable of
 
xbdestroya said:
People what's the doubt about? FFTs are like Cell's 'thing.' Or do we all forget last year's Barcelona conference so quickly? On the contrary - what I take away from this isn't a sense of being impressed - it's more like, show me something other than FFTs! :)

...and I am always very sceptical about the quality of the code other than the "target platform", as you'd need an assembly expert for each of these platforms...
 
dskneo said:
100Gflops is half of what one Cell is "announced" to be capable of

Don't start this again..the last time people started comparing realised performance with the paper specs of Cell and other platforms, things got very stupid very quickly :LOL: Besides, they reached near enough to 100% efficiency with one of their tests, if you go read the whole presentation.
 
slider said:
The BE is still four PPE's isn't it? I think that's what the patent showed. Does that also mean that each PPE has 8(?) SPE's with it? No wonder the numbers are so bloody high!
The Broadband Engine seems to be a term to cover any number of PPEs and SPEs. In the original papers the hope was for four of what we'd call a Cell processor, but a BE/BBE is now a 1:8 configuration, or typical Cell processor. The numbers are high because these are jobs Cell is good at doing. It's like looking at a GPU's graphics performance versus a CPU. The GPU is designed to run these tasks faster so scores orders of magnitudes faster. Anyone who has been following Cell shouldn't be surprised by these results.

maven said:
...and I am always very sceptical about the quality of the code other than the "target platform", as you'd need an assembly expert for each of these platforms...
The guys that wrote these benchmarks wrote the code for the other platforms to do the same job. Or do you think Mercury crippled their existing software solutions so when Cell comes it'd score artifically highly? I can understand people questioning the validity of benchmarks regards IBM figures, or other companies promoting their hardware, but this is an independent company free to choose any hardware platform they like. Why think they aren't going to evaluate each effectively? Why think they'll be better at writing optimized code for Cell, which is new and different, than writing optimized code for conventional processors they've had access to and used in their own products for years?
 
I couldn't include every slide in the original post (4 picture maximum), but many of these questions are answered in the presentation.

The performance on the 970, and I think also the Freescale, were using Mercury's own counterpart libraries on those chips, which would be their best effort on those chips also (they've used these chips in other products).

The intel numbers come from that website on FFT performance, using intel's own optimised libraries.
 
BTW, the PowerBlock 200 has a total of 5GB memory

XDR 1GB - CELL - south bridge - DDR2 4GB

so as far as the 5GB link between the south bridge and Cell is enough to buffer the data that can't fit in the XDR DRAM it seems you can build a render farm with Cell systems.
 
Shifty Geezer said:
The guys that wrote these benchmarks wrote the code for the other platforms to do the same job. Or do you think Mercury crippled their existing software solutions so when Cell comes it'd score artifically highly? I can understand people questioning the validity of benchmarks regards IBM figures, or other companies promoting their hardware, but this is an independent company free to choose any hardware platform they like. Why think they aren't going to evaluate each effectively? Why think they'll be better at writing optimized code for Cell, which is new and different, than writing optimized code for conventional processors they've had access to and used in their own products for years?

It is more of a general point of these benchmark comparisons, especially when source-code is not easily available.
I would not exactly call a company that is trying to sell these systems impartial, although they would be foolish to make their old systems look bad / worse than they are.

In any case, I am not looking for a fight, these results are nice, so let's look forward to better radar imaging in our game consoles... ;)
 
Titanio said:
Don't start this again..the last time people started comparing realised performance with the paper specs of Cell and other platforms, things got very stupid very quickly :LOL: Besides, they reached near enough to 100% efficiency with one of their tests, if you go read the whole presentation.

eh, true

i just stated that to make a point that a total of 4 PPe and 32SPE wouldn't just score "100glfops"...
 
[maven] said:
I would not exactly call a company that is trying to sell these systems impartial, although they would be foolish to make their old systems look bad / worse than they are.

Or they're just highlighting how super amazing the new system is and why you must buy it. If the old systems didn't look bad why update them? Welcome to business.
 
Shifty Geezer said:
Or do you think Mercury crippled their existing software solutions so when Cell comes it'd score artifically highly?

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? ;)
 
Back
Top