What about Intel's new Phi?

Message boards : Number crunching : What about Intel's new Phi?
Message board moderation

To post messages, you must log in.

AuthorMessage
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 1306484 - Posted: 15 Nov 2012, 18:01:02 UTC

I see Intel is releasing a new coprocessor that offers extremely high performance and what seems to be a very easy port. How will Seti support this in the coming new year?

http://newsroom.intel.com/community/intel_newsroom/blog/2012/11/12/intel-delivers-new-architecture-for-discovery-with-intel-xeon-phi-coprocessors
ID: 1306484 · Report as offensive
Claggy
Volunteer tester

Send message
Joined: 5 Jul 99
Posts: 4654
Credit: 47,537,079
RAC: 4
United Kingdom
Message 1306523 - Posted: 15 Nov 2012, 19:13:26 UTC - in response to Message 1306484.  

I see Intel is releasing a new coprocessor that offers extremely high performance and what seems to be a very easy port. How will Seti support this in the coming new year?

http://newsroom.intel.com/community/intel_newsroom/blog/2012/11/12/intel-delivers-new-architecture-for-discovery-with-intel-xeon-phi-coprocessors

Ask at the Boinc Dev Forum, until Boinc supports it, Seti can't support it,

Claggy
ID: 1306523 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 1306529 - Posted: 15 Nov 2012, 19:19:52 UTC - in response to Message 1306523.  

Would you be so kind as to point me to the Boinc Dev Forum?

ID: 1306529 · Report as offensive
Claggy
Volunteer tester

Send message
Joined: 5 Jul 99
Posts: 4654
Credit: 47,537,079
RAC: 4
United Kingdom
Message 1306539 - Posted: 15 Nov 2012, 19:32:30 UTC - in response to Message 1306529.  

Would you be so kind as to point me to the Boinc Dev Forum?

http://boinc.berkeley.edu/dev/index.php

You'll need to creat a new account to post, Best ask in this thread since it has no anwers: Xeon Phi (MIC): Will it crunch?

For Further Info, there is a thread here too: Xeon Phi (aka Knights Corner, MIC)

Claggy
ID: 1306539 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 1306550 - Posted: 15 Nov 2012, 20:03:55 UTC - in response to Message 1306539.  

Thank you Claggy

This link is from someone who has traditionally been not very favorable to Intel but this time he's about as optimistic as anyone can get:

http://semiaccurate.com/2012/11/13/what-will-intel-xeon-phi-do-to-the-gpgpu-market/


ID: 1306550 · Report as offensive
Keith White
Avatar

Send message
Joined: 29 May 99
Posts: 392
Credit: 13,035,233
RAC: 22
United States
Message 1306551 - Posted: 15 Nov 2012, 20:07:02 UTC
Last modified: 15 Nov 2012, 20:12:52 UTC

From what I read it's a cluster of roughly 60 Pentium (the original) class CPU cores but is devoid of the usual MMX, SSE and the new AVX instructions in favor of a specialized vector unit (512 bit SIMD design), all running at around 1GHz. In theory each core could do up to 4 threads but that's rarely occurs in reality due to other issues.

The card itself is $2700 and peaks at 225 watts. It's only 2-2.5x faster than the top end Xeon CPU in highly parallelized tasks.
"Life is just nature's way of keeping meat fresh." - The Doctor
ID: 1306551 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 1306603 - Posted: 15 Nov 2012, 23:24:34 UTC - in response to Message 1306551.  

The card itself is $2700 and peaks at 225 watts. It's only 2-2.5x faster than the top end Xeon CPU in highly parallelized tasks.


I believe it is over 1TeraFlop double precision. That's more than 2.5X a Xeon.


ID: 1306603 · Report as offensive
Profile ivan
Volunteer tester
Avatar

Send message
Joined: 5 Mar 01
Posts: 783
Credit: 348,560,338
RAC: 223
United Kingdom
Message 1306613 - Posted: 16 Nov 2012, 0:05:51 UTC - in response to Message 1306484.  

I see Intel is releasing a new coprocessor that offers extremely high performance and what seems to be a very easy port. How will Seti support this in the coming new year?

Ask me again when I get my hands on one. :-) Possibly before February if I made a good enough impression on our preferred HPC supplier last week...
ID: 1306613 · Report as offensive
Keith White
Avatar

Send message
Joined: 29 May 99
Posts: 392
Credit: 13,035,233
RAC: 22
United States
Message 1306697 - Posted: 16 Nov 2012, 7:03:35 UTC - in response to Message 1306603.  
Last modified: 16 Nov 2012, 7:10:40 UTC

The card itself is $2700 and peaks at 225 watts. It's only 2-2.5x faster than the top end Xeon CPU in highly parallelized tasks.


I believe it is over 1TeraFlop double precision. That's more than 2.5X a Xeon.


The 2.5x is from Intel's own page on the Phi, comparing two E5-2687 (3.1GHz, 8 cores) to one Phi card (1GHz, 60 cores). (footnote 5).

They did see some Monte Carlo simulations run at 10.75x faster comparing two E5-2670 (2.6GHz, 8 cores) to a Phi SE10P (1.1GHz, 61 cores). (footnote 9)

The 1 TFlop DP is theoretical peak performance.
"Life is just nature's way of keeping meat fresh." - The Doctor
ID: 1306697 · Report as offensive
Profile Raistmer
Volunteer developer
Volunteer tester
Avatar

Send message
Joined: 16 Jun 01
Posts: 6325
Credit: 106,370,077
RAC: 121
Russia
Message 1306722 - Posted: 16 Nov 2012, 10:04:05 UTC

If it will be programmable via OpenCL we can make use of it even w/o BOINC support as it was done already with early CUDA and early ATi Brook+/OpenCL support.
SETI@home (at least in it's optimized part) usually goes ahead of BOINC "official" support.
What we really need is hardware availability for test and free development tools for it.

SETI apps news
We're not gonna fight them. We're gonna transcend them.
ID: 1306722 · Report as offensive
Nick

Send message
Joined: 17 May 99
Posts: 96
Credit: 17,356,094
RAC: 0
United States
Message 1306761 - Posted: 16 Nov 2012, 14:16:09 UTC - in response to Message 1306697.  
Last modified: 16 Nov 2012, 14:18:15 UTC

The 2.5x is from Intel's own page on the Phi, comparing two E5-2687 (3.1GHz, 8 cores) to one Phi card (1GHz, 60 cores). (footnote 5)


Yes, you're right. That comparison is watt for watt. It would be 5X faster than a single 8 core Xeon but the power comparison seems more appropriate.
ID: 1306761 · Report as offensive
Profile HAL9000
Volunteer tester
Avatar

Send message
Joined: 11 Sep 99
Posts: 6534
Credit: 196,805,888
RAC: 57
United States
Message 1306833 - Posted: 16 Nov 2012, 17:49:46 UTC - in response to Message 1306722.  

If it will be programmable via OpenCL we can make use of it even w/o BOINC support as it was done already with early CUDA and early ATi Brook+/OpenCL support.
SETI@home (at least in it's optimized part) usually goes ahead of BOINC "official" support.
What we really need is hardware availability for test and free development tools for it.

I think standard x86 CPU supports OpenCL. So I would guess x86 based coprocessor cards would also.
SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[
ID: 1306833 · Report as offensive
Mark Lybeck

Send message
Joined: 9 Aug 99
Posts: 245
Credit: 216,677,290
RAC: 173
Finland
Message 1306881 - Posted: 16 Nov 2012, 19:44:27 UTC - in response to Message 1306697.  

The card itself is $2700 and peaks at 225 watts. It's only 2-2.5x faster than the top end Xeon CPU in highly parallelized tasks.


I believe it is over 1TeraFlop double precision. That's more than 2.5X a Xeon.


The 2.5x is from Intel's own page on the Phi, comparing two E5-2687 (3.1GHz, 8 cores) to one Phi card (1GHz, 60 cores). (footnote 5).

They did see some Monte Carlo simulations run at 10.75x faster comparing two E5-2670 (2.6GHz, 8 cores) to a Phi SE10P (1.1GHz, 61 cores). (footnote 9)

The 1 TFlop DP is theoretical peak performance.


Hello,

You get almost One teraflop with one GTX 500 series (5xx) card at a fraction of the price. What is the business case here?
ID: 1306881 · Report as offensive
Josef W. Segur
Volunteer developer
Volunteer tester

Send message
Joined: 30 Oct 99
Posts: 4504
Credit: 1,414,761
RAC: 0
United States
Message 1306891 - Posted: 16 Nov 2012, 20:20:16 UTC - in response to Message 1306881.  

...
The 1 TFlop DP is theoretical peak performance.

Hello,

You get almost One teraflop with one GTX 500 series (5xx) card at a fraction of the price. What is the business case here?

Double precision. For full-up supercomputers, that's an important consideration and the PHI 1 TFLOPS is a lot better than the direct competitor NVIDIA Tesla K10 at about 0.19 TFLOPS DP.

For SETI@home which uses mostly single precision, and users who are willing to go with consumer grade GPUs, top GPUs from either AMD or NVIDIA are likely much better than PHI on a cost/performance basis. But it would be good if a system with PHI bought for other purposes could also be used for SETI crunching. As Raistmer said, it depends on Intel providing the OpenCL drivers and low/no cost development software.

Hal is right that the high level OpenCL can be translated to x86, both AMD and Intel have drivers which do that for some range of their CPUs. But doing OpenCL on a Sandy Bridge chip is somewhat different than the same on 64 cores each with 512 bit SIMD but lacking some other refinements.
                                                                  Joe
ID: 1306891 · Report as offensive
Profile HAL9000
Volunteer tester
Avatar

Send message
Joined: 11 Sep 99
Posts: 6534
Credit: 196,805,888
RAC: 57
United States
Message 1306894 - Posted: 16 Nov 2012, 20:27:50 UTC - in response to Message 1306881.  

The card itself is $2700 and peaks at 225 watts. It's only 2-2.5x faster than the top end Xeon CPU in highly parallelized tasks.


I believe it is over 1TeraFlop double precision. That's more than 2.5X a Xeon.


The 2.5x is from Intel's own page on the Phi, comparing two E5-2687 (3.1GHz, 8 cores) to one Phi card (1GHz, 60 cores). (footnote 5).

They did see some Monte Carlo simulations run at 10.75x faster comparing two E5-2670 (2.6GHz, 8 cores) to a Phi SE10P (1.1GHz, 61 cores). (footnote 9)

The 1 TFlop DP is theoretical peak performance.


Hello,

You get almost One teraflop with one GTX 500 series (5xx) card at a fraction of the price. What is the business case here?

The number you are looking at on those cards is probably the SP(single precision) rating. Intel is stating TF for DP(double precision). Which generally is less than 1/4 of the SP rating. For instance a Radeon HD 6990 is rated 5099 GF running SP & 1276.88 GF running DP.

SETI@Home application currently only use SP. So the SP performance is what you would want to compare. Currently that isn't really helpful across manufacturers yet.
SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[
ID: 1306894 · Report as offensive
Keith White
Avatar

Send message
Joined: 29 May 99
Posts: 392
Credit: 13,035,233
RAC: 22
United States
Message 1306967 - Posted: 16 Nov 2012, 23:49:29 UTC - in response to Message 1306891.  

...
The 1 TFlop DP is theoretical peak performance.

Hello,

You get almost One teraflop with one GTX 500 series (5xx) card at a fraction of the price. What is the business case here?

Double precision. For full-up supercomputers, that's an important consideration and the PHI 1 TFLOPS is a lot better than the direct competitor NVIDIA Tesla K10 at about 0.19 TFLOPS DP.

For SETI@home which uses mostly single precision, and users who are willing to go with consumer grade GPUs, top GPUs from either AMD or NVIDIA are likely much better than PHI on a cost/performance basis. But it would be good if a system with PHI bought for other purposes could also be used for SETI crunching. As Raistmer said, it depends on Intel providing the OpenCL drivers and low/no cost development software.

Hal is right that the high level OpenCL can be translated to x86, both AMD and Intel have drivers which do that for some range of their CPUs. But doing OpenCL on a Sandy Bridge chip is somewhat different than the same on 64 cores each with 512 bit SIMD but lacking some other refinements.
                                                                  Joe

The Tesla K20 is 1.17 TFlop peak DP. Also a 225 watt card. The K20X is 1.31 TFlop peak DP at 235 watts. The K20X are used in the new #1 in the Top500 supercomputer list.

The AMD S10000 is 1.48 TFlop peak DP at 375 watts.

Everyone has a dog in the HPC hunt now.
"Life is just nature's way of keeping meat fresh." - The Doctor
ID: 1306967 · Report as offensive
Grant (SSSF)
Volunteer tester

Send message
Joined: 19 Aug 99
Posts: 13715
Credit: 208,696,464
RAC: 304
Australia
Message 1306980 - Posted: 17 Nov 2012, 0:12:58 UTC - in response to Message 1306722.  

If it will be programmable via OpenCL we can make use of it even w/o BOINC support as it was done already with early CUDA and early ATi Brook+/OpenCL support.

Keep in mind these are fully compatable x86 CPU cores.
If the software runs on a Xeon system, it will run on a Phi card without any changes. Of course if you do work the code to take advantage of Phi then you'll get even greater improvements in performance than just by running it as is.
Grant
Darwin NT
ID: 1306980 · Report as offensive

Message boards : Number crunching : What about Intel's new Phi?


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.