This is subtle and breathtakingly obvious (in retrospect, now a band of geniuses have thought of it): reversing the tables on what is the “main processor” and what is the “accelerator”, and making the customised computation silicon the baseline resource that utilises the CPUs if the need to do so arises rather than the other way around.
I am very impressed. And I am also impressed that their next machine will have ARM processors serving as the CPUs.
Well sure, but keep in mind that this is for supercomputing workloads which run extremely parallel workloads for the most part and only really do serial, CPU-ish workloads rarely.
Slight point of nuance: The Xeon Phi model mentioned in the article is not a coprocessor. The original Xeon Phi MIC coprocessors, which have been EoL since 2013 or 2014 [0], were known as Knight's Corner or "KNC" for short.
The Xeon Phi which was code named Knight's Landing ("KNL") did ship some coprocessor models, but these were not adopted widely compared to the models which function as primary processor for the compute node (in a LGA3647 socket). The KNL coprocessors have also been EoL'd, while the LGA3647 models have not [1].
So Intel started with this strategy, and then backed away from it after one-and-a-half generations. Now they are (possibly?) canceling the successor to even the socketed version [2] and spinning up a compute GPU division. Quite a winding road in such a short span of time.
> So Intel started with this strategy, and then backed away from it after one-and-a-half generations. Now yet they are (possibly?) canceling the successor to even the socketed version [2] and spinning up a compute GPU division. Quite a winding road in such a short span of time.
It looks like they search for a product/market fit, kind of like a startup which starts with one thing and then finds out that something else is far better and does that instead. Which is usually encouraged here. So, good for them? They experiment. Maybe they find something.
Oh, definitely. My post was not intended to be criticism of Intel. I sure as hell couldn't offer a strategy to a multinational semiconductor giant and hope to do any better. Mostly I am amused at how things come full circle: the Knight's Corner project came out of a canceled GPU project, and now they're back to GPUs.
Wonderful, it sounds like a lot of floating-point horsepower.
I can get a cheap Nvidia card with CUDA support and start playing with it under the day, but how does an individual developer, or even a small company, get hold of one of these?
This is the followon to SX-10 core. I really wish NEC would make this chip more available...the SX is a well understood, well supported, very capable architecture. They had the SX-6i and SX-8i "deskside supercomputer" versions in the past, so if they've reduced SX-10 to a PCI-e card, please package it with a compiler and let us have it.
I am very impressed. And I am also impressed that their next machine will have ARM processors serving as the CPUs.