Intel quietly kills off next-gen Knights Hill Xeon Phi chips

The SC '17 conference is taking place this week, and while many companies are taking the opportunity to launch new products and services at the show, Intel is quietly taking the opportunity to announce the removal of a chip from its roadmap. As part of a discussion of its plans for exascale computing, or systems capable of processing one thousand petaFLOPS, Intel noted that it will abandon its next-generation Xeon Phi accelerator, code-named Knights Hill, in favor of "a new microarchitecture and a new platform specifically designed for exascale."

Although we didn't examine news of Knights Hill in depth when it first emerged, the few details that ever did come out regarding the chip suggested it would have been a follow-on to the Knights Landing Xeon Phi family fabricated on Intel's 10-nm process technology that would incorporate Intel's second-generation Omni-Path interconnect for supercomputing systems. Although it's treacherous to guess at the reasons behind any move like this one, the company's struggles with bringing its 10-nm products to market seem the most likely cause.

The first consumer CPUs fabricated on a 10-nm process, code-named Cannon Lake, are set to begin shipping "near the end of [this] year," while volume shipments are officially slated for the first half of next year. The only Cannon Lake CPU that Intel has officially shown operating in any form so far has been one for ultrabook form factors, suggesting small dies will lead the way on the new process. Given the monster die size of past Xeon Phis, yield or performance challenges for Knights Hill on the new process could have interfered with a timely introduction of the product.

It's unclear what this roadmap change means for the Knights Mill Xeon Phi, an accelerator that Intel added to its roadmap last August. Knights Mill was touted as an ideal platform for deep-learning tasks with its improved variable-precision support, among other features. Knights Mill is slated for shipment this quarter, so we may hear more about it this week. Cheers to @witeken on Twitter for the tip.

Comments closed
    • psuedonymous
    • 2 years ago

    Probably a good sign that whatever GPU architecture Intel are going for, it’s not going to be a Ghost of Larrabee design based on clustered tiny x86 cores.

      • tipoo
      • 2 years ago

      Yeah. Sounds like the scaled up GPU is set to replace it.

      Xeon Phi was very flexible but not that competitive in pure performance per dollar with workstation grade graphics.

    • snowMAN
    • 2 years ago

    Is it possible that this is connected to the high-end GPUs they want to develop? Perhaps they are planning on using those for big number-crunching applications.

    • SkyWarrior
    • 2 years ago

    [s<]INTELAMD[/s<] [s<]INTELEON GPU[/s<] HALFLIFE3 CONFIRMED Yeah whatever...

    • NoOne ButMe
    • 2 years ago

    what a shocker.

    they killed a slow, large and expensive product.

    slow as in slower than Nvidia (and probably AMD) despite being a larger chip that probably had more transistors.

    And expensive as it had to consumer offroot.

      • Klimax
      • 2 years ago

      Evidence for slow?

        • NoOne ButMe
        • 2 years ago

        Raw numbers are bad for it, both the peak compute numbers, and all the numbers I can find online which also benchmark NVidia GPUs next to them.

        I’m struggling to find any site that has access to both.

        I also generally agree with the sentiment that if a company doesn’t want their product compared to competition, it probably is because it looks bad. For professional parts, you don’t see many, in part because they are hard to benchmark well. But you still see Nvidia, and even sometimes AMD pro cards benchmarked against the competition.

        Even on youtube results are nonexistent. 😐

        I am of course not counting any results Nvidia or Intel are putting out themselves.

          • chuckula
          • 2 years ago

          I don’t know where you are getting this idea that Intel doesn’t want people using the Xeon Phi.

          The latest Top 500 list has 4 of the top 10 systems in the world using Xeon Phi.

          [url<]https://www.top500.org/lists/2017/11/[/url<]

            • the
            • 2 years ago

            I do believe that Intel doesn’t want the Xeon Phi to be used [i<]outside of HPC.[/i<] The last part is critical as there are a handful of applications where despite its low single threaded performance, it is 'good enough' and makes up ground by throwing an enormous amount of threads at a problem. Ditto for some select memory bandwidth bound applications (hello HCM). Intel would prefer that you purchase a dual Xeon SP box that not only costs more in hardware but generally more in software licensing too.

    • the
    • 2 years ago

    Intel’s 10 nm process has claimed another victim. For those keeping track, it appears that desktop Cannon Lake was the first (Ice Lake looks to be the first 10 nm desktop part due next year).

    I don’t think that the Knight’s family itself is dead. Knights Mill is still on track for example but is a 14 nm part. So what is next for general HPC? Probably a series of parts comprised of several tiles joined at the hip by EMIB and Intel flat out not caring about aggregate die size to win back market share. Using the current dual core per tile topology, I would predict a 4 x 4 tile arrangement per die which equates to 32 cores per die. From there, it is just IO around the perimeter and several stacks of memory to complete things. Not cheap to produce but these are the ultra high margin HPC parts.

      • NoOne ButMe
      • 2 years ago

      Knights Mill is probably already made.

      This chip might never have gotten masks even.

      The Knight’s family is effectively dead, I suspect. Just a matter of what parts are far enough along that that can eek some profit out.

      Finally after 10-15 years Intel looks to seriously be making a non-x86 product inhouse.

        • the
        • 2 years ago

        Knight’s Hill likely had some masks made as Intel has been prototyping their 10 nm process for some time and this was to be their first big die on it. However, that process is roughly ~18 months behind schedule which is a good part of a product generation. By the time Intel’s 10 nm process was ready for a big die at acceptable yields, Knight’s Hill as a product would be behind AMD and nVidia in the HPC space. Intel has to realign their HPC roadmap and Knight’s Milll appears to be a good stop gap as it appears to be basically Knight’s Landing with some additional ISA extensions for variable precision to please the deep learning crowd. A few small surprises might make their way into Knight’s Mill as well (NVDIMM support, TSX maybe?) that would help mature the product.

        As for generations after Knight’s Hill, it is curious that Intel is still investing into the Atom lineage. If the Knight’s family is genuinely dead, it strikes me odd that resources are still allocated here as a generation has been skipped in both mobile and HPC. With Intel widening the front end to 4-way, this should help improve HPC workloads, especially considering the 4-way SMT will likely remain a feature.

        As for Intel adopting non-x86, yes but not what you think. Their FPGA business and Nervana tech are pursuing emerging markets that can easily leverage Intel’s foundries. This were acquisitions as Intel saw an opportunity to profit by removing the middle men in bringing future generations of this technology to market themselves. Intel management today doesn’t seem to be afraid of ditching their investments either if they don’t bare fruit.

          • DavidC1
          • 2 years ago

          Yes.

          It’s not necessarily Intel adopting non-x86 either. Those products are still just additions to Intel’s primary lines.

          The Altera FPGA business will also come in a on package format later this year.

          After Lake Crest, they’ll come with Knights Crest. They call it a “Bootable Intel Xeon Processor with Integrated Acceleration”. It sounds like it could be a Xeon Phi version with Nervana IP instead of AVX-512 accelerators.

          Few years ago the cable modem division moved to Atom-based cores, which are x86. Xeon Scalable’s Lewisburg chipset uses a Quark x86 chip for ME/IE features.

          In the future, I can see the Altera FPGA’s ARM cores being replaced with Quark, or Atom cores.

          • NoOne ButMe
          • 2 years ago

          with regards to Intel and non-x86. I mean from an inhouse perspective.

          Intel has done non-x86 stuff inhouse, but generally it got killed (Larrabee, revived as x86) or has bought companies that don’t do x86, and let them ride.

            • the
            • 2 years ago

            The shader cores of Larrabee were always x86. The catch there is that Intel eventually added some fixed function TMUs and ROPs to quickly handle those operations at lower power consumption. The software only solution Intel was proposing didn’t have enough processing power to do both shader and ROP tasks in a timely manner.

            • NoOne ButMe
            • 2 years ago

            that’s interesting, I didn’t realize that Larabee had x86. Thanks for the neat bit of info 🙂

          • Ninjitsu
          • 2 years ago

          I’m not saying GPGPU, but GPGPU 😛

            • chuckula
            • 2 years ago

            The only way you are going to convince me that aliens didn’t design those Intel chips is when the aliens land here and tell us they didn’t do it!

    • Neutronbeam
    • 2 years ago

    Tis a sad farewell to the knights who say xe.

      • chuckula
      • 2 years ago

      Defeat at the castle seems to have utterly disheartened Intel. The ferocity of the French taunting took Intel completely by suprise, and Intel became convinced that a new HPC microarchitecture was required if the Quest for the Exascale Grail were to be brought to a successful conclusion. Intel, having consulted its closest poached engineering hires, decided that they should separate, and search for the Exascale Grail individually. Now this is what they did….

      • K-L-Waster
      • 2 years ago

      No one brought them a shrubbery it would seem.

        • GrimDanfango
        • 2 years ago

        Ah Monty Python… the “{reference to memes!!!}” from before the internet existed 😛

        …Still a better love story than Twilight.

          • rrrhal
          • 2 years ago

          Brave Sir Xeon ran away
          Bravely ran away away
          When danger reared its ugly head
          He bravely turned his tail and fled
          Yes, brave Sir Xeon turned about)
          And gallantly he chickened out

          Bravely taking to his feet
          He beat a very brave retreat
          Bravest of the brave, Sir Xeon!

        • Chrispy_
        • 2 years ago

        The problem was that people brought them a shrubbery as asked, but then they just wanted another one, beside the first one, only slightly higher so you get a two-level effect with a path through the middle.

        Intel should have been clearer about the system requirements up front, rather than playing this switcharoo with the number of shrubberies.

        Don’t even get me [i<]started[/i<] on the herring....

          • chuckula
          • 2 years ago

          I like the laurels particularly.

            • Liron
            • 2 years ago

            Intel likes to rest on them too.

          • Growler
          • 2 years ago

          Oh, what sad times are these when passing ruffians can say Xe at will to old system builders. There is a pestilence upon this land, nothing is sacred. Even those who arrange and design shrubberies are under considerable economic stress in this period in history.

          • just brew it!
          • 2 years ago

          What about the Killer Ryzen of Caerbannog?

            • Mr Bill
            • 2 years ago

            [url=https://en.wikipedia.org/wiki/Rabbit_of_Caerbannog#/media/File:Paris_-_Cath%C3%A9drale_Notre-Dame_-_Portail_du_Jugement_Dernier_-_PA00086250_-_154.jpg<]This one?[/url<] Send for The Holy Hand Grenade from Icy Lake Antioch!

      • tipoo
      • 2 years ago

      Have my upvote and then see yourself out

    • chuckula
    • 2 years ago

    [quote<]It's unclear what this roadmap change means for the Knights Mill Xeon Phi, an accelerator that Intel added to its roadmap last August. Knights Mill was touted as an ideal platform for deep-learning tasks with its improved variable-precision support, among other features. [/quote<] According to Serve the Home Knights Mill is being shown extensively around SC 2017 so it's probably not canceled like Knights Hill. [url<]https://www.servethehome.com/intel-at-sc17-no-official-knights-mill-launch-and-a-path-to-exascale/[/url<]

    • DancinJack
    • 2 years ago

    Nervana!

    • chuckula
    • 2 years ago

    And you thought Intel’s hire of Raja was just about finally getting Crysis running on Intel graphics.

      • tipoo
      • 2 years ago

      I’m still bitter they bought out Project Offset to show off Larrabee and then killed that with it. The tech demos looked impressive for *years* after its debut.

      [url<]https://www.youtube.com/watch?v=rkoi6WCCJHY[/url<] [url<]https://www.youtube.com/watch?v=HX0jO3y02iI[/url<]

Pin It on Pinterest

Share This