In 2006, Intel determined to construct a brand new GPU structure, codenamed Larrabee. It didn’t work, so Intel determined to re-spin it as a manycore structure dubbed Knights Ferry. Whereas the primary KNF processors have been prototypes with literal GPU silicon nonetheless bolted on-die, follow-up chips like Knights Nook switched to a modified Pentium (P54C) structure with x86 ISA help, Four-way SMT, and as much as 1.2TFLOPS of double-precision efficiency. The Knights Touchdown structure up to date the CPU core to a closely modified model of Intel’s Silvermont uarch, elevated the core depend as much as 72 cores, added MCDRAM (aka Hybrid Reminiscence Dice), AVX-512 help, and was presupposed to set the stage for a 10nm refresh with 2nd-generation Omnipath — besides Intel simply quietly killed it.
You possibly can be forgiven for not noticing, even for those who learn Intel’s weblog publish on SC17, the annual supercomputing conference. Out of the corporate’s 828-word weblog publish, simply 30 of them cope with the cancellation of Knights Hill, and it’s buried in a bit labeled “Enabling Exascale.” (PR professional tip: If you wish to cover the adverse information, put it underneath a optimistic header). It reads: “One step we’re taking is to interchange one of many future Intel Xeon Phi processors (code title Knights Hill) with a brand new platform and new microarchitecture particularly designed for exascale.”
Intel additionally had an element named Knights Mill in improvement, however we don’t know a lot about that chip but, past that it helps half-precision work like GPUs from AMD and Nvidia and is meant for AI and deep studying workloads. Whereas we don’t know a lot in regards to the low-level structure of Knights Hill, we do know one thing about what sort of deployment Intel anticipated the chip to see, through the Aurora supercomputer undertaking.
There are a number of alternative ways to learn this announcement, and we don’t know which is correct but.
Essentially the most direct is that Intel is canceling Knights Hill as a separate product as a result of it believes it could goal each areas (HPC, AI/DL) with the identical half. Years in the past, we talked about experimental analysis Intel had completed into variable floating level computing, which saved on energy by giving programmers the choice to decide on how a lot precision to retain on a case-by-case foundation.
Second, it’s attainable that one thing has gone badly fallacious with some facet of the chip’s design, to such an unlimited diploma that cancelling the undertaking and beginning over is the one strategy to repair it. This wouldn’t routinely level to one thing fallacious with the CPU core, which seemed like an inexpensive die shrink with some new options — it could possibly be a scaling or energy consumption problem with another facet of the design. There’s no proof to recommend that is true, it’s only a technical chance.
Third, there’s the possibility that Intel has determined Knights Hill isn’t going to hit the aggressive efficiency or functionality targets it desires and is re-spinning the design to take action. Xeon Phi merchandise have at all times had a for much longer cadence than different elements within the client and even enterprise — Intel has refreshed its Xeon processors far more shortly, for instance — and so, it could be pondering long-term about the way forward for exascale and selecting to take successful now slightly than danger being dethroned afterward. Both method, don’t be stunned if Nvidia and even AMD play up this strategic hole with new launches of their very own. Hat tip to Tech Report for sussing out Intel’s blink-and-you’ll-miss-it announcement.