Oak Ridge National Laboratory gets Kepler K20 "Big K"

Jaydip · Oct 31, 2012

ShintaiDK said:
Xeon Phi is doing above 1Tflop in DP. And thats x86. Thats the same ballpark as GK110s DP. Xeon Phi is just much easier to code for and dont have performance penalities in alot of cases like AMD/nVIdia cards will.

Just see on AMD vs nVidia when 1 single instruction is missing in bitcoin mining.

And the power envelop for Xeon Phi? Now it is really easy to develop gpu powered apps using cuda.

ShintaiDK · Oct 31, 2012

Jaydip said:
And the power envelop for Xeon Phi? Now it is really easy to develop gpu powered apps using cuda.

Its the same ballpark as Kepler. Plus the Phi got 8GB.

CUDA compared to x86 is horrible.

sontin · Oct 31, 2012

Phi is not the same "x86" like Intel's CPUs. You can't use existing code and let it perform on Phi. There is the same learning curve like it was with CUDA.

boxleitnerb · Oct 31, 2012

Xeon Phi has significantly less compute power and apparently uses more power than K20:
http://news.softpedia.com/news/Intel-s-Knights-Corner-Burns-258W-at-Idle-292566.shtml

And that was not even the full part but only one using 56 cores instead of 64.

ShintaiDK · Oct 31, 2012

sontin said:
Phi is not the same "x86" like Intel's CPUs. You can't use existing code and let it perform on Phi. There is the same learning curve like it was with CUDA.

The basic x86+x87+x64 is there. MMX, XMM and YMM registers are replaced however with the 512bit ZMM.

ShintaiDK · Oct 31, 2012

boxleitnerb said:
Xeon Phi has significantly less compute power and apparently uses more power than K20:
http://news.softpedia.com/news/Intel-s-Knights-Corner-Burns-258W-at-Idle-292566.shtml

And that was not even the full part but only one using 56 cores instead of 64.

Kinda went terrible wrong here didnt it?

In an interesting development, hardware experts at Semiaccurate.com

And just for the laugh:

http://news.softpedia.com/news/Inte...ping-New-Specifications-Revealed-284222.shtml

For more information on Intel’s Larabee architecture and Xeon Phi cards, check out Theo Valich’s detailed report.

boxleitnerb · Oct 31, 2012

So they faked the screenshots?

ShintaiDK · Oct 31, 2012

boxleitnerb said:
So they faked the screenshots?

And the screenshot shows exactly what?

And why would one identical card use so much less than the other? Maybe Charlie didnt tell the entire story...as always. Is the utility even measuring correctly or did the speaker say its not final?

http://goparallel.sourceforge.net/how-xeon-phi-stacks-up-gpus/

“We put Intel’s world-class power management technology into the Xeon Phi,” Chrysos concludes. “When individual cores are idle, or the Xeon Phi is not processing anything, we reduce its power consumption proportionately.”

I think I take Xeon Phi lead architect George Chrysos over Charlie.
Xeon Phi even support C1, C3 and C6 states. So if all the cores are gated, what should use the power?
http://newsroom.intel.com/servlet/J...on_Phi_Hotchips_architecture_presentation.pdf

Jaydip · Oct 31, 2012

ShintaiDK said:
Its the same ballpark as Kepler. Plus the Phi got 8GB.

CUDA compared to x86 is horrible.

Thats impressive.But why do you say CUDA compared to x86 is horrible?the learning curve is mostly same.

boxleitnerb · Oct 31, 2012

ShintaiDK said:
Its the same ballpark as Kepler. Plus the Phi got 8GB.

Btw do you have any sources for that? What is ballpark, 225W, 250W, 300W?

Jaydip · Oct 31, 2012

It is apparently 300W
http://www.anandtech.com/show/6265/intels-xeon-phi-in-10-petaflops-supercomputer

boxleitnerb · Oct 31, 2012

If that were true, a big big ouch. K20 uses 30% less power AND has 30% higher performance (at least theoretical values). Okay, Xeon Phi has up to 33% higher bandwidth and 33% more memory.

ShintaiDK · Oct 31, 2012

Jaydip said:
It is apparently 300W
http://www.anandtech.com/show/6265/intels-xeon-phi-in-10-petaflops-supercomputer

On 1 6 pin and 1 8pin connector?

railven · Oct 31, 2012

ShintaiDK said:
On 1 6 pin and 1 8pin connector?

Using Jaydip's link:

Anyway, the Xeon Phi is definitly not limited to ultra expensive supercomputers. Supermicro showed us the Superserver 2027GR-TRF which contain 4 Xeon Phi cards thanks to two redundant 1800W (Platinum) PSUs. The rest of the server consists of two Xeon E5 and 16 DIMM slots in total, supporting up to 256 GB. So it seems that one Xeon Phi card consume about 300W.

Seems like an estimate, not exact figure.

Jaydip · Oct 31, 2012

They guessed it but x86 is not much power efficient.
1 6 pin and 1 8pin connector can draw up to 300watts(assuming pcie ofc).

ShintaiDK · Oct 31, 2012

K20 got 1 6pin and 1 8pin.

Just like Xeon Phi.

tviceman · Oct 31, 2012

ShintaiDK said:
On 1 6 pin and 1 8pin connector?

There has been multiple links given to you now that verifies what you say isn't very accurate. Sure semiaccurate and brightside aren't the most reliable sources of news, but when 1, 2, 3, 4 different sites - some of which are very reputable, report the SAME THING, then it's more than likely accurate. Yet you continue to argue against what several other people have shown you to be the apparent truth, and at the same time you cannot provide any links or concrete info to back up what you say. Denial?

ShintaiDK · Oct 31, 2012

tviceman said:
There has been multiple links given to you now that verifies what you say isn't very accurate. Sure semiaccurate and brightside aren't the most reliable sources of news, but when 1, 2, 3, 4 different sites - some of which are very reputable, report the SAME THING, then it's more than likely accurate. Yet you continue to argue against what several other people have shown you to be the apparent truth, and at the same time you cannot provide any links or concrete info to back up what you say. Denial?

You mean the sites that uses SA etc as source? Or AT that just makes a guess?

Jaydip · Oct 31, 2012

@Shintai
K20 is not rated for 300W.AT guessed it but they are probably spot on.Power efficiency is not a great trait of x86. Xeon phi may be great but when you use openmp to tap that potential it isn't any easier than cuda.
http://forums.anandtech.com/member.php?u=307779

Sunny129 · Oct 31, 2012

moonbogg said:
I wonder how much RAC that thing could BOINC up...

+1...this is all i care about

sontin · Oct 31, 2012

Jaydip said:
@Shintai
K20 is not rated for 300W.AT guessed it but they are probably spot on.Power efficiency is not a great trait of x86. Xeon phi may be great but when you use openmp to tap that potential it isn't any easier than cuda.

Perf/watt of Phi is not really great. Intel is not even really beating Fermi with Phi:

intel_xeon_phi_hotchips_architecture_presentation_page_05.jpg

http://www.techpowerup.com/171436/I...e-Details-of-Intel-Xeon-Phi-Co-Processor.html

ShintaiDK · Oct 31, 2012

sontin said:
Perf/watt of Phi is not really great. Intel is not even really beating Fermi with Phi:

http://www.techpowerup.com/171436/I...e-Details-of-Intel-Xeon-Phi-Co-Processor.html

Your graph does show its beating Fermi.

boxleitnerb · Oct 31, 2012

Your picture shows exactly the opposite. Do you look at your pictures, before you post them?

Jaydip · Oct 31, 2012

That graph is wrong.Apparently the intel powered system consume 100w not 73w.Check this
http://top500.org/list/2012/06/200

sontin · Oct 31, 2012

ShintaiDK said:
Your graph does show its beating Fermi.

Yes, it's beating a two old architecture on a three old process with a 9% better perf/watt.

K20 will have a 2 times better theoretical perf/watt and 3 times better practical Perf/watt than M2090.

Phi is no competition to Kepler.

Oak Ridge National Laboratory gets Kepler K20 "Big K"

Diamond Member

Lifer

Diamond Member

Platinum Member

Lifer

Lifer

Platinum Member

Lifer

Diamond Member

Platinum Member

Diamond Member

Platinum Member

Lifer

Diamond Member

Diamond Member

Lifer

Diamond Member

Lifer

Diamond Member

Diamond Member

Diamond Member

Lifer

Platinum Member

Diamond Member

Diamond Member