mersenneforum.org Intel Xeon PHI?
 Register FAQ Search Today's Posts Mark Forums Read

 2020-07-23, 05:10 #1 Cheetahgod   May 2020 318 Posts Intel Xeon PHI? It would be interesting to see if you could use a intel phi co-processor to do prp tests. You can get them on ebay for around $80 -$100 https://www.ebay.com/itm/Intel-Xeon-...wAAOSwHSxdp2xv Last fiddled with by Uncwilly on 2020-07-23 at 14:02 Reason: added Xeon to title
 2020-07-23, 14:08 #2 Uncwilly 6809 > 6502     """"""""""""""""""" Aug 2003 101×103 Posts 2×3×1,579 Posts It is a Xeon. You might want to look at this thread: https://www.mersenneforum.org/showthread.php?t=16912
2020-07-23, 14:56   #3
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

500910 Posts

Quote:
 Originally Posted by Cheetahgod https://www.ebay.com/itm/Intel-Xeon-...wAAOSwHSxdp2xv
Any model that ends in P is passively cooled, expecting the case it's installed in to blow cooling air through it. Note there's no fan visible in the listing you posted. The models ending in A that have their own cooling fan suitable for consumer chassis use are rare but do exist.
There are multiple generations of Xeon Phi with substantially different performance. https://en.wikipedia.org/wiki/Xeon_Phi

Evaluating the 7120A I bought is on my to-do list. https://www.ebay.com/itm/Intel-Xeon-...r/233581462183

Other listings for the same model with stock on eBay are currently at US$600 and up. https://www.ebay.com/itm/New-Dell-In...D/192724974858 is a less costly possibility, with its own fan. Otherwise figure out how you're going to cool a passively cooled model. Systems designed to support the passively cooled Xeon Phis can be expensive https://www.ebay.com/itm/SuperMicro-...E/254659755925 To compile code for Xeon Phi and their embedded Linux OS requires special development software from Intel which is$3000/license, or a busy 30-day free evaluation period.

I estimate it to be somewhere around 1/4 the speed of a Radeon VII, yet use an equal power budget.

Last fiddled with by kriesel on 2020-07-23 at 14:57

 2020-07-23, 20:58 #4 tServo     "Marv" May 2009 near the Tannhäuser Gate 62210 Posts GIMPS had a Xeon Phi system ( at one time ) A bunch of people chipped in and a system was purchased. As mentioned, the software license has probably expired. The person who hosted the system is inactive in this forum so I'm not sure the system can be accessed. Ernst did a lot of work on the system, as I recall. However, the results from this hardware were unimpressive. See this thread: https://www.mersenneforum.org/showthread.php?t=21544 BTW, I have seen blog posts showing IBM is getting ready to try again with a massive new chip. Perhaps third time's a charm. I wonder who would bother to try it since their reputation is somewhat tarnished.
2020-07-23, 22:04   #5
ewmayer
2ω=0

Sep 2002
República de California

101101011010002 Posts

Quote:
 Originally Posted by tServo A bunch of people chipped in and a system was purchased. As mentioned, the software license has probably expired. The person who hosted the system is inactive in this forum so I'm not sure the system can be accessed. Ernst did a lot of work on the system, as I recall. However, the results from this hardware were unimpressive.
The KNL was actually quite good in terms of performance for the work I was using it for - one of my two side-by-side (one @60M FFT, one @64M) Pépin primality tests of F30, the 64M was running 64-threaded (i.e. 1 thread per hardware core) on the KNL, getting ~68 ms/iter at time the physical host of the system, David Stanfill, went AWOL. Getting good || scaling for this kind of GIMPS-style work on more than 8-16 cores was a problem, but I found that the bigger the FFT memory-footprint, the better the || scaling, so at 64M I was able to get (IIRC) more than 50% the extrapolated "perfect parallel" scaling. I expect the above runtime has since been eclipsed by top-tier GPUs, but not bad for a $5000 system at the time we bought it. And even more valuable was the ability to do AVX-512 code prototyping & testing on actual hardware, rather than a simulator. Re. licenses, I only ever used GCC on the system, what proprietary licenses might be needed in that context? Would a standard Linux install work on this kind of hardware? Last fiddled with by ewmayer on 2020-07-23 at 22:06  2020-07-23, 22:09 #6 VBCurtis "Curtis" Feb 2005 Riverside, CA 2×5×11×43 Posts If I understand Phi correctly, the chip linked in the first post is on a PCI card and needs a special compiler and/or OS, while the system GIMPS purchased was the sole CPU on a motherboard and ran a "regular" linux and GCC. The cheap used parts are the addin cards, and are much tougher to get performance out of; I would be surprised if a Phi outran a same-price CUDA card on any software run for projects in this forum. Cool tech from 5+ years ago may still be cool now, but is almost-never a faster solution than new commodity stuff. 2020-07-24, 02:40 #7 kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 5,009 Posts Quote:  Originally Posted by ewmayer Re. licenses, I only ever used GCC on the system, what proprietary licenses might be needed in that context? Would a standard Linux install work on this kind of hardware? For the addin PCIe Xeon Phi, there's a uOS linux variant embedded on the card. Drivers to talk to it from the host OS are needed Or something like that. Supposedly there are Windows drivers for it. There appears to also be Linux host support. The license is for Intel's Parallel Studio and related tools. Puget Systems has some old articles about it. https://www.pugetsystems.com/all_hpc.php?query=phi Last fiddled with by kriesel on 2020-07-24 at 02:41 2020-07-24, 20:58 #8 ewmayer 2ω=0 Sep 2002 República de California 23·1,453 Posts Quote:  Originally Posted by kriesel For the addin PCIe Xeon Phi, there's a uOS linux variant embedded on the card. Drivers to talk to it from the host OS are needed Or something like that. Supposedly there are Windows drivers for it. There appears to also be Linux host support. The license is for Intel's Parallel Studio and related tools. Puget Systems has some old articles about it. https://www.pugetsystems.com/all_hpc.php?query=phi Well, for under$100, it seems worth a try - I'd snap it up but I'm fresh out of full-width PCI slots in my 2 systems.

I just contacted the seller re. driver and other-software support, reply said they'd been referring people to this Intel page -- Looks like both Linux and Windows support stacks there, The Linux ones involve specific enterprise-server distros, but I see no licensing hurdles.

2020-07-24, 21:05   #9
chalsall
If I May

"Chris Halsall"
Sep 2002

224558 Posts

Quote:
 Originally Posted by ewmayer The Linux ones involve specific enterprise-server distros, but I see no licensing hurdles.
One could reasonably presume that includes RedHat, which is (legally) cloned in CentOS.

You only really need to pay for it when the PHB's insist on it (doesn't hurt that they're now owned by IBM).

 2020-07-25, 12:52 #10 Xyzzy     "Mike" Aug 2002 1F5916 Posts
 2020-07-25, 21:36 #11 ewmayer ∂2ω=0     Sep 2002 República de California 23·1,453 Posts Mike, this sounds like the kind of software-stack-to-run-interesting-hardware project you enjoy, especially with the GPUs you tried out last month idled due to excess heat. If someone manages to get one of these up and running, and the throughput from running mprime or Mlucas on all cores is within 10x of a Radeon VII - which seems entirely plausible, based on my earlier runs on the GIMPS KNL, you've already broken even in terms of what even a used R7 currently costs, assuming the wattage is similar. C'mon, you know you want to...

 Similar Threads Thread Thread Starter Forum Replies Last Post dtripp Software 3 2013-02-19 20:20 nucleon Hardware 2 2012-05-10 23:53 R.D. Silverman Programming 19 2011-09-17 01:43 mack Information & Answers 7 2009-09-13 01:48 penguain NFSNET Discussion 0 2006-06-12 01:31

All times are UTC. The time now is 11:15.

Sat Apr 10 11:15:43 UTC 2021 up 2 days, 5:56, 1 user, load averages: 1.46, 1.29, 1.29