Xeon Phi: Intel's Larrabee-Derived Card In TACC's Supercomputer

Intel Xeon Phi Architecture

Intel has a vast portfolio of technology its engineers have developed, and Xeon Phi unquestionably taps some of that. However, the Many Integrated Core architecture is notably more than a bunch of modified Pentium processors manufactured at 22 nm. Some of its notable attributes include:

  • An in-order, dual-issue x86 design with 64-bit support
  • Four threads per core, and up to 61 cores per coprocessor
  • 512-bit SIMD for wider vectors
  • 512 KB of L2 cache per core (up to 30.5 MB per Xeon Phi)
  • 22 nm tri-gate transistors
  • Red Hat Enterprise Linux 6.x or SuSE Linux 12+ support
  • 6 or 8 GB of GDDR5 per card


You'll notice that even the highest-end Xeon Phi wields far fewer cores than a typical GPU. However, you cannot compare an MIC core to a CUDA core, for example, on a 1:1 basis. Just one Phi core is quad-threaded with a 512-bit SIMD unit. A fair comparison requires getting past marketing's definition of a "core."

It's also interesting that the card runs Linux. This probably isn't a solution you'd want to run a LAMP package on, but I would also guess that someone will try to do it anyway. You can SSH into the Xeon Phi card, though, to find out more about the hardware. We were advised that the following screenshot came from a pre-production board.

In the following diagram of a MIC architecture core, Intel claims that less than two percent of the core and cache die area is x86-specific logic. Although the Xeon E5-2680 CPUs also found in the Stampede supercomputer are made up of 2.27 billion transistors each, the lineage of x86 comes from the 20 000- to 30 000-transistor 8086 processor.

Of course, even today's desktop CPUs are incredibly complex, emphasizing the importance of getting data to and from where it needs to go as expediently as possible. Like the Sandy and Ivy Bridge-based CPUs, the prototype product code-named Knights Corner employs a ring bus interconnect to most effectively maximize throughput and available die area. By also giving each core lots of cache, the processor is able to avoid the performance hit it'd take if each core instead needed to be fed constantly from the GDDR memory controller stops.

Create a new thread in the US Reviews comments forum about this subject
This thread is closed for comments
70 comments
    Your comment
    Top Comments
  • mocchan
    Articles like these is what makes me more and more interested in servers and super computers...Time to read up and learn more!
    20
  • wannabepro
    Highly interesting.
    Great article.

    I do hope they get these into the hands of students like myself though.
    17
  • Other Comments
  • esrever
    meh
    -25
  • tacoslave
    i wonder if they can mod this to run games...
    -22
  • mocchan
    Articles like these is what makes me more and more interested in servers and super computers...Time to read up and learn more!
    20
  • wannabepro
    Highly interesting.
    Great article.

    I do hope they get these into the hands of students like myself though.
    17
  • ddpruitt
    Intriguing idea....

    These X86 cores have the uumph to run something a little more complex than what a GPGPU can. But is it worth it and what kind of effort does it require. I'd have to disagree with Intel's assertion that you can get used to it by programming for an "i3". Anyone with a relatively modern graphics card can learn to program OpenCL or CUDA on there own system. But learning how to program 60 cores efficiently (or more) from an 8 core (optimistically) doesn't seem reasonable. And how much is one of these cards going to run? You might get more by stringing a few GPUs together for the same cost.

    I'm wonder if this is going to turn into the same time of niche product that Intel's old math-coprocessors did.
    9
  • CaedenV
    man, I love these articles! Just the sheer amounts of stuffs that go into them... measuring ram in hundreds of TBs... HDD space in PBs... it is hard to wrap one's brain around!

    I wonder what AMD is going to do... on the CPU side they have the cheaper (much cheaper) compute power for servers, but it is not slowing Intel sales down any. Then on the compute side Intel is making a big name for themselves with their new (but pricy) cards, and nVidia already has a handle on the 'budget' compute cards, while AMD does not have a product out yet to compete with PHI or Tesla.
    On the processor side AMD really needs to look out for nVidia and their ARM chip prowess, which if focused on could very well eat into AMD's server chip market for the 'affordable' end of this professional market... It just seems like all the cards are stacked against AMD... rough times.

    And then there is IBM. The company that has so much data center IP that they could stay comfortably afloat without having to make a single product. But the fact is that they have their own compelling products for this market, and when they get a client that needs intel or nvidia parts, they do not hesitate to build it for them. In some ways it amazes me that they are still around because you never hear about them... but they really are still the 'big boy' of the server world.
    8
  • A Bad Day
    esrevermeh


    *Looks at the current selection of desktops, laptops and tablets, including custom built PCs*

    *Looks at the major data server or massively complex physics tasks that need to be accomplished*

    *Runs such tasks on baby computers, including ones with an i7 clocked to 6 GHz and quad SLI/CF, then watches them crash or lock up*


    ENTIRE SELECTION IS BABIES!

    tacoslavei wonder if they can mod this to run games...


    A four-core game that mainly relies on one or two cores, running on a thousand-core server. What are you thinking?
    -3
  • ThatsMyNameDude
    Holy shit. Someone tell me if this will work. Maybe, if we pair this thing up with enough xeons and enough quadros and teslas, we can connect it with a gaming system and we could use the xeons to render low load games like cod mw3 and tf2 and feed it to the gaming system.
    -1
  • mayankleoboy1
    Main advantage of LRB over Tesla and AMD firepro S10000 :

    A simple recompile is all thats needed to use PHI. Tesla/AMD needs a complete code re write. Which is very very expensive .
    I see LRB being highly successful.
    1
  • PudgyChicken
    It'd be pretty neat to use a supercomputer like this to play a game like Metro 2033 at 4K, fully ray-traced.

    I'm having nerdgasms just thinking about it.
    5
  • palladin9479
    Each of these cards is their own computer and runs their own OS. The host system will need to manage the environment of the card and give it some form of permanent storage, most likely through abstraction.

    I've worked with specialized cards like these, through they were for running non-native x86 code on a SPARC platform. They act like a totally separate system with their own IP / virtual frame buffer / IO space, the whole works. You treat them just like you would a dedicated server. Which makes their choice of linux pretty clear, you can easily cluster multiple non-uniform linux servers for distributed processing.

    These are very interesting, won't be quite as powerful as a dedicated vector processor but handles general computing tasks well enough and can run code natively without needing to rewrite the program.
    5
  • PreferLinux
    CaedenVman, I love these articles! Just the sheer amounts of stuffs that go into them... measuring ram in hundreds of TBs... HDD space in PBs... it is hard to wrap one's brain around!I wonder what AMD is going to do... on the CPU side they have the cheaper (much cheaper) compute power for servers, but it is not slowing Intel sales down any. Then on the compute side Intel is making a big name for themselves with their new (but pricy) cards, and nVidia already has a handle on the 'budget' compute cards, while AMD does not have a product out yet to compete with PHI or Tesla.On the processor side AMD really needs to look out for nVidia and their ARM chip prowess, which if focused on could very well eat into AMD's server chip market for the 'affordable' end of this professional market... It just seems like all the cards are stacked against AMD... rough times.And then there is IBM. The company that has so much data center IP that they could stay comfortably afloat without having to make a single product. But the fact is that they have their own compelling products for this market, and when they get a client that needs intel or nvidia parts, they do not hesitate to build it for them. In some ways it amazes me that they are still around because you never hear about them... but they really are still the 'big boy' of the server world.

    1. Not really, their price/performance ratio is fairly similar. They're a lot cheaper, but also perform a lot worse. Add in the higher power consumption (and it surely matters in anything at this level), and if anything Intel is better.

    2. AMD released the FirePro S10000 less than 24 hours ago. It is competing directly against this and Tesla.

    -----------------------------------

    On a completely separate note, I'm wondering what the price of these is. nVidia's latest K20 costs about $3200 and is rated at 1.17 TFLOPS peak and 225 W. This is rated at 1.01 TFLOPS peak, with the same power rating. It wouldn't be hard for Intel to beat nVidia on price...
    2
  • Alphi
    aye. they can accurately model the weather... but during downtime they cannot play 3D games at ridiculously high frame rates. what geek would recommend buying such a "work tool"?
    0
  • blazorthon
    PreferLinux1. Not really, their price/performance ratio is fairly similar. They're a lot cheaper, but also perform a lot worse. Add in the higher power consumption (and it surely matters in anything at this level), and if anything Intel is better.2. AMD released the FirePro S10000 less than 24 hours ago. It is competing directly against this and Tesla.-----------------------------------On a completely separate note, I'm wondering what the price of these is. nVidia's latest K20 costs about $3200 and is rated at 1.17 TFLOPS peak and 225 W. This is rated at 1.01 TFLOPS peak, with the same power rating. It wouldn't be hard for Intel to beat nVidia on price...


    Some of AMD's Opterons have far greater performance for the money than any Xeons do, so I'd have to disagree (at least partially) with you. The same is probably not true for all of AMD's Opterons, but it is true for at least some of them.

    For example, here:
    http://www.newegg.com/Product/Product.aspx?Item=N82E16819113038
    $600 for a CPU that in fully threaded work can compete with Intel's six-core and eight-core Xeon models quite well, if not beat them, in work that scales across as many cores as you throw at it and it does so while using not too much power at all.

    Other good examples:
    http://www.newegg.com/Product/Product.aspx?Item=19-113-036

    http://www.newegg.com/Product/Product.aspx?Item=N82E16819113030

    Pretty much all of these models:
    http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100008494+600061216&QksAutoSuggestion=&ShowDeactivatedMark=False&Configurator=&IsNodeId=1&Subcategory=727&description=&hisInDesc=&Ntk=&CFG=&SpeTabStoreType=&AdvancedSearch=1&srchInDesc=

    They all have better price/performance than the multi-socket Xeons. Lower power efficiency, definitely so in most cases, but not in all.
    4
  • IndignantSkeptic
    Please forgive my ignorance but why do we need x86 processors still when a compiler now is supposed to automatically generate all the necessary machine codes from the same high level source code?

    Also will we ever be able to upgrade supercomputers simply by replacing small parts of it at a time with newer better components and basically having different parts of the supercomputer running at different speeds?

    Also why the hell is CUDA and OpenCL advocated repeatedly in this article when OpenMP was supposed to replace them, as far as I understand, and is actually mentioned in one of the pictures?!
    0
  • memadmax
    Ok, the hardware looks decent.
    But, there's one problem: Intel is relying on programmers being able to optimize to a high degree for this setup...

    And that is the problem, programmers these days are lazy...
    It used to be you would spend a whole day optimizing loops using counters and timers to squeeze every last drop out of that loop...

    Now, programmers are like kids with legos, they just merge crap together that was written by someone else and as long as it runs decent, they call it good...
    -2
  • PreferLinux
    blazorthonSome of AMD's Opterons have far greater performance for the money than any Xeons do, so I'd have to disagree (at least partially) with you. The same is probably not true for all of AMD's Opterons, but it is true for at least some of them.For example, here:http://www.newegg.com/Product/Prod [...] 6819113038$600 for a CPU that in fully threaded work can compete with Intel's six-core and eight-core Xeon models quite well, if not beat them, in work that scales across as many cores as you throw at it and it does so while using not too much power at all.Other good examples:http://www.newegg.com/Product/Prod [...] 19-113-036http://www.newegg.com/Product/Prod [...] 6819113030Pretty much all of these models:http://www.newegg.com/Product/Prod [...] rchInDesc=They all have better price/performance than the multi-socket Xeons. Lower power efficiency, definitely so in most cases, but not in all.

    Some... Good point.

    IndignantSkepticPlease forgive my ignorance but why do we need x86 processors still when a compiler now is supposed to automatically generate all the necessary machine codes from the same high level source code?Also will we ever be able to upgrade supercomputers simply by replacing small parts of it at a time with newer better components and basically having different parts of the supercomputer running at different speeds?Also why the hell is CUDA and OpenCL advocated repeatedly in this article when OpenMP was supposed to replace them, as far as I understand, and is actually mentioned in one of the pictures?!

    Because if you've got x86 code and you want to use OpenCL or CUDA you have to completely re-write it. Compilers can't just translate between them. Besides, a lot of HPC stuff would be highly-optimised, and that often means low-level stuff (Assembly, even).

    http://en.wikipedia.org/wiki/OpenMP OpenMP is for multi-threading, generally using the CPU. It also works with clusters over a network, hence it is great for this (Xeon Phi). But it doesn't work with GPUs.

    memadmaxOk, the hardware looks decent.But, there's one problem: Intel is relying on programmers being able to optimize to a high degree for this setup...And that is the problem, programmers these days are lazy...It used to be you would spend a whole day optimizing loops using counters and timers to squeeze every last drop out of that loop...Now, programmers are like kids with legos, they just merge crap together that was written by someone else and as long as it runs decent, they call it good...

    Um, I get the distinct impression it needs very little optimisation. I think programmers will find it far easier to optimise for this than for GPGPU which requires a complete re-write.

    PreferLinux-On a completely separate note, I'm wondering what the price of these is. nVidia's latest K20 costs about $3200 and is rated at 1.17 TFLOPS peak and 225 W. This is rated at 1.01 TFLOPS peak, with the same power rating. It wouldn't be hard for Intel to beat nVidia on price...

    As a follow-up to this, (one of) the SemiAccurate articles gave the price, and it was mid-$2000s – much better than nVidia's K20, for a similar performance, and most likely a much better architecture to program for (or make use of).
    2
  • cats_Paw
    Is there any way to convert that into gaming power? I mean, 60 cpus for gaming is definitly overkill, but would it be posibl to use that and chanel all that raw power into really amazing games?

    I ask cus, in reality, gaming industry does seem limited to me ...
    0
  • IndignantSkeptic
    @PreferLinux, I'm shocked programmers still program in low level. Also how can OpenMP not work on GPUs when a section of OpenMP is OpenACC?!!
    0
  • blazorthon
    Cats_PawIs there any way to convert that into gaming power? I mean, 60 cpus for gaming is definitly overkill, but would it be posibl to use that and chanel all that raw power into really amazing games?I ask cus, in reality, gaming industry does seem limited to me ...


    I doubt that you could do a whole lot with that much performance even if we made games that could somehow utilize that many cores. You might be able to do something like one hell of a physics processing game with very high FPS that way, but otherwise, IDK how much you could actually do with it.
    0