Intel Quietly Drops Xeon Phi 7200 Coprocessors

[AndyChow]

called it 20x over for the last 8 months - funny how so many people in meatspace didn't believe me. Nobody was buying them, nobody was developing against them except a few national labs for what I guess is computational physics. The things sucked at almost everything, required you to use openacc or something for "real" performance and really only had an advantage where things coelesced more poorly. They lost in perf/$ and raw performance in any other circumstance. Canceling IDF was just a tip of the hand, but I didn't know anyone using them anywhere and barely got access to them.

Yes, it was obvious when Intel started selling the first one for $100-200 a pop (with an *). Basically, saying "find a use for them, please".

It was a potential powerhouse for something, but that something was never figured out. A solution looking for a problem.

BTW Michael, you got a shout out for this article on Tom's.

[Zucca]

I could still buy a Xeon Phi card (with active cooler preferably) later when (?) the prices drop. Just because then I could have a computer inside my computer and call it "Yo dawg!".
But seriously. I would like to run some long running processes in the OS the card has. The problem is that I'd need to have a motherboard for the task. My server has some AM3+ MB atm.

[nevion]

Interesting but not a surprise. This particular version just didnt have market acceptance. Phi wasnt a bad concept but not really tackeled in a rational manner. Intel should have focused on a bootable processor that leveraged a standard motherboard and socket in yhe first place.

From this article it appears that newer socketed Phis are still in production so we will see how long this is infact a viable product at Intel. I just see the specialized co processor as a dead end. At least when one uses a GPU as a compute source you have a range of commodity hardawre for the simpler acceleration needs and compatible solutions for more advanced solutions.

not a bad concept, just awful execution. Astoundingly bad. And people say the government does a bad job.. All they needed was better performance and software for making use of the compute capability. They couldn't win enough customers off GPUs.

Anyway intel will be making a comeback in the enterprise sphere to a limited extent via FPGAs. When/if they crack the software on that which they're making progress on, they'll have a unique product with a permanent market used to those hardware costs and have some perf/watt to win back from GPUs.

[nevion]

Yeah. AFAIK the PCIe Xeon Phi is basically its own computer running it's own (Linux) OS. I don't remember how much ram those had, but propably not much. Although NFS has swap support (and RDMA), so lending some memory from the "host" OS is possible via the PCIe x16 lane.
It's sad that those Xeon Phi cards don't just work on any regular consumer motherboard. I remember Xeon Phi needing something like larger adressable memory space etc...

generally people didn't use the os hosting capability, it was more of a compute card

[nevion]

Considering there are plenty of supercomputers out there already running the socketed Xeon Phi, it was probably just an issue of very little demand for a PCIe card that lacks some of the more worthwhile features of the socketed product.

were, and while they don't replace fast, it's already clear their time is over looking at newer more high perf builds.

[nevion]

called it 20x over for the last 8 months - funny how so many people in meatspace didn't believe me. Nobody was buying them, nobody was developing against them except a few national labs for what I guess is computational physics. The things sucked at almost everything, required you to use openacc or something for "real" performance and really only had an advantage where things coelesced more poorly. They lost in perf/$ and raw performance in any other circumstance. Canceling IDF was just a tip of the hand, but I didn't know anyone using them anywhere and barely got access to them.

[starshipeleven]

Wonderful. You are living in your mom's basement and don't have a use for an HPC part. Guess what: Intel is running more Xeon Phis in supercomputers *right now* than AMD has shipped Epyc processors period. So by your lack of logic, Epyc has been canceled.

I'm going out and take a wild stab in the dark with the following daring statement:

You didn't understand what his point was about.

[chuckula]

I think your argument is weak.

I would no more use a "Knight's Landing" board as a desktop CPU than I would use a desktop CPU to mine cryptocurrencies. .

Wonderful. You are living in your mom's basement and don't have a use for an HPC part. Guess what: Intel is running more Xeon Phis in supercomputers *right now* than AMD has shipped Epyc processors period. So by your lack of logic, Epyc has been canceled.

[NotMine999]

Considering there are plenty of supercomputers out there already running the socketed Xeon Phi, it was probably just an issue of very little demand for a PCIe card that lacks some of the more worthwhile features of the socketed product.

I think your argument is weak.

I would no more use a "Knight's Landing" board as a desktop CPU than I would use a desktop CPU to mine crypto currencies. Granted, either product can be programmed to do "less than optimal" tasks, but the real point ought to be "use the best tools available for the job". The Intel add-in boards like "Knight's Landing" have always impressed me as "a scaled down compute cluster on a card" while most desktop CPUs are compromises between compute (what CPUs originally did decades ago), I/O (even more prevalent in SoC designs and "merged Northbridge" designs), and graphics (assuming an integrated CPU-GPU design) tasks.

The closest comparable socketed product to this discontinued product would be the Intel Xeon line of processors that are widely found in servers. The latest announcements from Intel regarding their "gold / silver / bronze" Xeon processors and Intel "Denverton" SoCs with higher core counts, newer processor features/flags, and broader instruction sets could be challenging the position of older products like the "Knight's Landing" card. Prospective buyers could be asking themselves this "value proposition" question: "Why should we buy this older technology when this newer tech shows more potential?"

Here's an interesting look at the details of the Intel "Knight's Landing" architecture that is not specific to any given model of "Knight's Landing" Xeon Phi:

https://www.servethehome.com/intel-k...etails-emerge/

Seeing how the article dates back to 2015, it is quite possible that Intel has discontinued the product due to "lifecycle planning". Said another way, I doubt Intel sees any value in continuing to "mass market" old processor designs like 8088, 80286, 80386, and so on since more powerful and efficient processor designs exist today; newer/better processor designs with more & more features seem to evolve very fast.

[sl1pkn07]

and the 5110P?