Intel Xeon Phi battles GPUs, defends x86 in supercomputers

Intel's Xeon Phi might have started out with the goal of creating an x86-based graphics engine, but it ended up defending the x86 architecture's hegemony in high-performance computing against the onslaught of GPU coprocessors from Nvidia and AMD. However, it ends up being a battle among GPUs anyway – but in a different market …

[K]

Ummm

Where is the DVI output?

[Dani Eder]

Re: Ummm

This is a math co-processor to assist the CPU to crunch numbers faster. It is not a graphics card, so it does not have monitor outputs. If you want a workstation that can do both heavy duty number crunching and real time graphics, you would need two cards - this one plus a graphics card.

[Matt Bucknall]

Re: Ummm

Didn't spot the 'Joke Alert' icon huh?

[VeganVegan]

Re: Ummm

not a joke:

http://www.theregister.co.uk/2012/11/12/amd_firepro_s10000_gpu_card

<icon>heat produced by these cards</icon>

[Anonymous Coward]

More important questions

How long will it take the Phi to break into my neighbor's WIFI? :P

...but all joking aside, browsing the Top 500 site I noticed Stampede did not publish any power usage stats. I guess it's not too much of an issue when you have a 1.1 megawatt nuclear reactor next door, but inquiring minds will want to know how efficient the Phi boards are compared to the new Nvidia K20s. Just going off the specs sheet:

K20, 225W, 1.17 TFDP, $3199 estimated -- 5.2 GFDP/Watt, 0.37 GFDP/$ estimated

K20X, 235W, 1.31 TFDP, >$3199 estimated -- 5.57 GFDP/Watt, <0.41 GFDP/$ estimated

5110P, 225W, 1.011 TFDP, $2,649 -- 4.49 GFDP/Watt, 0.38 GFDP/$

3120A, 300W, 1.003 TFDP, ~$2,000 -- 3.34 GFDP/Watt, 0.50 GFDP/$

Negotiated volume discounts aside, Nvidia seems to have the edge with the power efficiency (16% for the K20 and 24% for the K20X vs. the 5110P) while staying somewhat competitive (for the K20 at least - no telling for the K20X) in terms of capital cost unless you just want performance as cheap as possible (efficiency be damned!) in which case the 3120A is your ticket.

I have no idea what I'd use one for... but I want one ; )

[borkbork]

Re: Stampede

Says 6MW over here:

http://semiaccurate.com/2012/11/12/peeking-at-the-tacc-stampede-supercomputer/

That's more than the nuclear reactor next door, unless you meant 1.1 gigawatts :)

[markx]

Re: Stampede

No, not right. Stampede is a work in progress. If you go back to the original article you will see it says:

*Once completed* [it] will pull 6MW, and that is for an estimated 9PFlops. But that is probably peak rather than Linpack flops. Anyway, it is certainly not as bad as your numbers suggest...

[Anonymous Coward]

Re: Stampede

Hey, wasn't trying to slag anyone! Being an alumni of the university in question, their efforts are near and dear to me...

I would be happy to re-do that chart if anyone can help me find a better power number to use.

Versus the Titan - the new power efficiency king - their coprocessors (Phi) are less efficient, but their CPUs (Intel) could be more efficient and balance that out. Regardless, for a current gen install I'd expect the power efficiency to at least be in the ballpark of 2 Tflop/s per kW. 9 PFlops at 6MW would still only be 1.5 Tflop/s per kW - which would put them at #6 out of the top 10. It's better than they had but nowhere near the Cray and BlueGenes which are all above the 2 Tflop/s per kW mark.

[tpm]

Re: Re: Stampede

Actually, I put the most recent data into the Top500 story here:

http://www.theregister.co.uk/2012/11/12/top_500_supercomputers_nov_2012/

Dell and TACC said Stampede would be around 9 petaflops, but as you can see in more current documents I cited at https://www.xsede.org/documents/234989/378230/XSEDE12_Stampede_and+MIC_Overview.pdf, the plan is to reach around 10 petaflops peak with a 6 megawatt power draw.

[Anonymous Coward]

Re: Stampede

Cool, so "1,667 megaflops per watt peak"... at least when they get fully built out. Here's the chart redone, ranked by power efficiency:

Rank/Name, TFlop/s, kW, Tflop/s per kW

1. Titan, 17590, 8209, 2.14277013

5. JUQUEEN, 4141, 1970, 2.102030457

9. Fermi, 1725, 822, 2.098540146

2. Sequoia, 16325, 7890, 2.069074778

4. Mira, 8162, 3945, 2.068948035

7. Stampede, 2660, na, 1.667 estimated

6. SuperMUC, 2897, 3423, 0.846333625

3. K computer, 10510, 12660, 0.830173776

8. Tianhe-1A, 2566, 4040, 0.635148515

10. DARPA Trial, 1515, 3576, 0.423657718

[TeeCee]

Old joke time.

" spinning at 1.053GHz"

Presumably that's the answer you get when you ask a 1GHz chip full of pentium FPUs what it's clocking at.

[Ramiro]

Up to 1.7x

For a finite element code, isn't really that much, is it?

Customers are *not* seeing big performance gains from Xeon Phi coprocessors.

Might as well use another cpu, simpler and possibly more power efficient.