Intel Xeon Phi co-processor (GPGPU)

Discussion in 'Intel x86 CPUs and chipsets' started by stmok, Jun 19, 2012.

  1. stmok

    stmok Member

    Joined:
    Jul 24, 2001
    Messages:
    8,878
    Location:
    Sydney
    => http://www.techpowerup.com/167870/I...option-Announcing-Xeon-Phi-Co-Processors.html

    Remember Intel's Larrabee project? It was to be their first discrete video card in a long time. Unfortunately, it failed as a GPU and was cancelled in 2010. However, the research and resources from that project wasn't thrown away...It was adapted to the high-performance computing role (supercomputing) under a new name: Intel Many Integrated Core Architecture (Intel MIC).

    The first card was codenamed "Knights Ferry". This was a development prototype solution that was never sold commercially. It was given to various commercial and academic developers around the world to tinker with.

    According to Wikipedia: It was a PCIe card with 32 in-order cores at up to 1.2 GHz with 4 threads per core, 2 GB GDDR5 memory, and 8 MB coherent L2 cache (256 kB per core with 32 kB L1 cache), and a power requirement of ~300 W, built at a 45 nm process

    Performance was approx 750 GFLOPS under single precision.


    Today, (at Hamburg International Supercomputing Conference 2012), Intel has announced the commercialised implementation of their MIC architecture. Codenamed "Knights Corner", it will be marketed under the Intel Xeon Phi brand.

    According to Top500's current list, its now ranked 150th. :)
    => http://www.top500.org/list/2012/06/200

    If you're a developer with access to the hardware, the Knights Corner instruction set documentation is found here.
    => http://software.intel.com/en-us/forums/showthread.php?t=105443


    ...For those who must ask: No, you won't be able to play Crysis on it. But you will be able to Fold with it. (Presuming you can afford one when its released!) :lol:
     
  2. Reaper

    Reaper Member

    Joined:
    Jun 27, 2001
    Messages:
    10,769
    Location:
    Brisbane, Qld, Australia
    Been able to teraflop fold for years. Intel is playing catch up. I wouldn't bother, buy more radeons instead. You'd be able to fill your computer with those for less than what one of these devices would cost.
     
  3. bobbth

    bobbth Member

    Joined:
    Nov 1, 2010
    Messages:
    680
    Location:
    Melbourne
    so it's useless then? :lol:



    It could be competitive depending on the performance/watt and its bandwidth requirement when compared to the traditional GPGPU competitors (i.e. quattro/firepro).
    One of the big selling points of GPGPU on a large scale basis is that the core density is much, much greater that that of the traditional cpu based systems, if this can execute X86/X64 instructions then this could be a real competitor in the HPC sector.

    Edit:


    http://theburnerishot.com/photo/Knights-Ferry-8-Way.jpg
     
    Last edited: Jun 19, 2012
  4. proffesso

    proffesso Member

    Joined:
    Jan 5, 2002
    Messages:
    8,978
    Location:
    Watsonia, Melbourne
    if it can execute x86 instructions, then it might be good for renderes that dont support gpgpu...but i highly doubt it would...I dream of cpu add-in cards :/
     
  5. Phido

    Phido Member

    Joined:
    Jun 20, 2003
    Messages:
    4,394
    Location:
    Dark City
    A long time ago in a galaxy far far away....

    Intel sold the i960 (no not the 960 your thinking of) as a coprocessor on a card.
    http://en.wikipedia.org/wiki/Intel_i960
    Back when people were sporting 286's and 386's having a 32bit risc processor on a ISA slot was woohoo. Last time they were used were in things like printers.

    Intel's key strength is that they write compilers that people use and they can get people coding for things.

    I think Intel has missed the boat. GPU's are now the default daughtercard processors, and the design of GPU's have been evolved to make them more useful at non rendering data processor. People will spend $500 or $5000 on a GPU that also kicks butt in data and graphics rather than $5000 for something thats only going to be used in some software that doesn't exist.
     
  6. proffesso

    proffesso Member

    Joined:
    Jan 5, 2002
    Messages:
    8,978
    Location:
    Watsonia, Melbourne
    a long-ish time ago there was this as well

    http://www.protograph.co.uk/artvps.html

    basically a bunch of fpga's on a board to accelerate raytracing.

    these guys were ahead of the gpgpu curve though, and charged way too much for it. which is a pity, if it caught on then we might have better options today
     
  7. Phido

    Phido Member

    Joined:
    Jun 20, 2003
    Messages:
    4,394
    Location:
    Dark City
    I hadn't heard much about those guys. Still kicking I see. Seemed like they had some interesting features going there.

    However these days most people seem better off with a few more GPU's or a renderfarm or what not. The thing about having more x86 power is that you can run whatever you dam want on it and not be tied to a path that goes nowhere.

    You could build a 8-12 node AMD (6 or 8 core) or intel (4 core) farm. Which could be used for rendering, post processing, encoding, etc. Able to split between workloads etc as per required.

    8120FX cpus are like $160 now. 1045t can be had ~$110. $200 a node?.
     
  8. Luke212

    Luke212 Member

    Joined:
    Feb 26, 2003
    Messages:
    9,136
    Location:
    Sydney
    why do people think it will be expensive? i was hoping they will be cost competitive with to gpu solutions
     
  9. OP
    OP
    stmok

    stmok Member

    Joined:
    Jul 24, 2001
    Messages:
    8,878
    Location:
    Sydney
    Because its targeted for the "high performance computing" market.

    Other cards in the same market are Nvidia's Tesla series and AMD's FireStream series...Just look up the prices for those lines, and you'll understand. ;)

    ie:
    NVIDIA Tesla C870 => ~AUD$500 to AUD$700
    NVIDIA Tesla C1060 => ~AUD$1500 to AUD$1600
    NVIDIA Tesla C2075 => ~AUD$2500 to AUD$2600
     
  10. akashra

    akashra Member

    Joined:
    Apr 25, 2003
    Messages:
    3,644
    Location:
    Melbourne, AU
    Pft, we just installed a system that now sits in position 29 almost right under my feet, two floors down :D
    65 TB of RAM, 65,536 cores. Wonder if my account allows me to ssh in to avoca yet ;D
    Doubt it :(
     
  11. AfterBurner1

    AfterBurner1 Member

    Joined:
    Oct 29, 2007
    Messages:
    1,002
    Location:
    Adelaide
    Either way, I'd think that these systems would really get some folding done when set up right and singing.
     
  12. Myne_h

    Myne_h Member

    Joined:
    Feb 27, 2002
    Messages:
    7,372
    Seems like a damn good idea for pure grunt work.
    The Pentium core and instruction set are well understood. They were a comparatively simple processor and arguably the first really modern CPU. The 486 you might think had the FPU, but it spoke directly to its ram and cache and I think the Vesa Local Bus without a chipset per se, so it cannot claim this title.

    Obviously x86 knowledge is everywhere, so optimisation is easy.
    Investment into this platform will last. Nvidia could arbitrarily decide to stop support for CUDA at their next architectural refresh and no one would be able to do jack. Whatever you write for this will run on any cpu for the forseeable future.

    Presumably multiple units can be installed in a machine. It's hard to top the flexibility of 200 CPU's over a graphics card.
    Can a graphics card run a bunch of Virtual Machines? I doubt it, but this puppy should be able to.

    Graphics cards, even with their abilities are still specialist devices requiring specialist coders and there's always the risk that Nvidia or AMD could simply go under.


    That's why it will sell. Because in the high density space, it is pure generic grunt. It's a tractor, not a dune buggy.

    Nice concept. Nice to know someone still has respect for Keep It Simple Stupid.
     
  13. OP
    OP
    stmok

    stmok Member

    Joined:
    Jul 24, 2001
    Messages:
    8,878
    Location:
    Sydney
    Intel Xeon Phi steals top HPC tenders, more SKUs to come
    => http://vr-zone.com/articles/intel-xeon-phi-steals-top-hpc-tenders-more-skus-to-come/16410.html

    According to the above link, the Xeon Phi offers advantages over current GPU based solutions.

    (1) Easier to port code to...Because it uses x86 + SIMD instead of OpenCL or CUDA.

    (2) Likely to match or exceed rival GPGPU solutions. (Based on VR-Zone's sources.)


    ...As a result, organisations (supercomputing customers) are considering Intel's Xeon Phi as part of their future upgrade to pursue more computational performance. (Over Nvidia's Tesla...Which is currently being used by the top 10 supercomputers in the world.)



    I actually would love to see Intel produce a cost-friendly, consumer-oriented version. (For enthusiasts, DIY'ers, students, etc...As this group can come up with some creative uses for the Xeon Phi.)
     
  14. OP
    OP
    stmok

    stmok Member

    Joined:
    Jul 24, 2001
    Messages:
    8,878
    Location:
    Sydney
    Intel Xeon Phi (B0 Stepping): The Knight in Shining Armor?
    => http://vr-zone.com/articles/intel-xeon-phi-b0-stepping--the-knight-in-shining-armor-/16871.html

    The gist...
    (Too much chit-chat in the article.)

    A0 silicon
    * rough 22nm design.
    * 48, 52, or 60 cores. (At 1Ghz.)
    * 1.5 to 1.9MB of L1 cache.
    * 24 to 30MB of L2 cache.
    * 4 or 8GB GDDR5 memory. (600Mhz to 1.125GHz in QDR mode.)
    * Not competitive against flagship model of Nvidia's Tesla GPGPU line.

    B0 silicon
    * 57, 60, and 61 cores.
    * 1.8 to 1.9MB of L1 cache.
    * 28 to 30.5MB of L2 cache.
    * 3GB, 6GB and 8GB of GDDR5 memory. (1.25 to 1.37GHz QDR mode.)
    * Has Turbo mode. (Speeds undisclosed.)
    * There are 5 models:
    (1) 57 cores with 3GB GDDR5
    => 600MHz to 1.1GHz
    => 245W TDP
    (2) 57 cores with 6GB GDDR5
    => 600MHz to 1.1GHz
    => 300W TDP
    (3) 60 cores with 6GB GDDR5
    => 630MHz+ to 1.05 or 1.09GHz
    => 245W TDP
    (4) 61 cores with 8GB GDDR5
    => 630MHz+ to 1.05 or 1.09GHz
    => 300W TDP
    => Two models: One comes with a passive heatsink, the other does not.
    * Available with or without passive heatsink.
    => Ones without heatsink are destined for systems using third-party custom cooling implementations or liquid cooling.
    * Active cooling (fan) heatsink only available with the 300W, 57 core + 6GB GDDR5 version.

    Intel's goal is to push for 1 TFLOPS double precision and 2 TFLOPS single precision.



    This is related in terms of real world application of the Xeon Phi...

    Cray Signs Contract to Install Cascade Supercomputer and Sonexion Storage System at the Pawsey Centre
    => http://investors.cray.com/phoenix.zhtml?c=98390&p=irol-newsArticle&ID=1717600

     
  15. bobbth

    bobbth Member

    Joined:
    Nov 1, 2010
    Messages:
    680
    Location:
    Melbourne
    300w sounds like a lot to me, especially for 22nm and there is supposed to be a passive cooled option?

    How do you passively cool ~300w? you would need a heatsink the size of a dog
     
  16. NSanity

    NSanity Member

    Joined:
    Mar 11, 2002
    Messages:
    16,500
    Location:
    Canberra
    Its not really the size of the heatsink (beyond a certain point), but the CFM of the fans.

    Remember anything passively cooled will be a rack mounted chassis option (not a desktop) and will be in an actively cooled environment.
     
  17. bobbth

    bobbth Member

    Joined:
    Nov 1, 2010
    Messages:
    680
    Location:
    Melbourne
    I'll be honest I pictured this in a reasonably mid-to-low airflow environment (i.e. the average desktop), I don't know much about commercial hardware, is this standard practice?
    to me it sounds extreme to cool the load of a gtx480 passively.
     
  18. NSanity

    NSanity Member

    Joined:
    Mar 11, 2002
    Messages:
    16,500
    Location:
    Canberra
    The heatsinks will be large and unweildey - but its all ducted and the like. No doubt it will use heatpipes etc.

    But passively cooling multiple 150w TDP cpu's is fairly standard, even in 1U form factors.
     
  19. bobbth

    bobbth Member

    Joined:
    Nov 1, 2010
    Messages:
    680
    Location:
    Melbourne
    That sounds intense, those machines must be loud! (not that it matters I suppose)
     
  20. Aetherone

    Aetherone Member

    Joined:
    Jan 15, 2002
    Messages:
    8,459
    Location:
    Adelaide, SA
    Ear bleeding noise? Sure is :) :wired: :shock:
     

Share This Page