I’ve searched both LevelOneTechs and various threads I could find via Google on this topic, but I’m wondering if it’d be possible to expand the compute capability (namely, core count) of one of my hosts that has multiple VMs. I use qemu / kvm.
I don’t wish to pass this card through to a VM, I’m basically wondering it’d be possible to assign cores on this card over the slot to a VM. Perhaps wouldn’t be very performant, but it’s a question that came to mind.
===
Edit 1: Here’s another link that’s intrigued me about a particular card I see on eBay right now for like, $80 -
Languages, tools, and applications run smoothly across the full spectrum of Intel Xeon family-based platforms. Plus, discover the flexibility of a coprocessor that can also host an OS.
What do they mean by, it’s possible to host an OS? How about a VM? (Perhaps I should review the supported instruction sets?)\
The coprocessor OS shipped as part of the MPSS is based on the standard Linux kernel source from kernel.org with minimal changes.
Also
One can also add loadable kernel modules to increase the functionality of the OS through kernel drivers. Intel tools such as profiling tools and debugging tools use the loadable kernel modules to interact with the hardware.
So…shouldn’t it be possible to run KVM on this card?
===
Edit 3: It looks like there are cards from '17 that actually have the VT-* instruction set, though earlier revisions did not support this -
The Wikipedia page describes the thing like a math coprocessor on steroids, designed to handle very complex floating point tasks. All the older (cheap) cards are rated over 1000 GFLOP, the newer ones breaking past 3000 GFLOP.
Are your guest VMs doing mathematically complicated tasks like running some nuclear stockpile stewardship applications? (simulating fusion bombs)
They all seem to be designed to go into a server chassis with ducted forced air cooling, or more exotic cooling methods.
I never used the cards for KVM, but did use them for offloading computation (and as a dedicated machine when Knights Landing came about). I’m not sure how much use the Knights Mill cards got, the Xeon Phi program was all but dead by then. I would be very surprised if the Knights Mill processors are add-in PCIe cards, I think they were designed as a last ditch upgrade for socketed Knights Landing cards. (But I may be wrong)
You probably don’t want a Xeon phi, though. The Knights Corner card you posted is not particularly powerful. A lot of the FLOPS come from the 512 bit vector unit which is not AVX, it has Intel’s own vector extension that only the KNC cards used. The cores don’t support out of order computation and are basically old pentium cores from the 90s (P5 or P54C, I can’t remember off the top of my head). What I remember doing was using some of the offload extensions to send kernels to the card and also SSHing to the card and using it as a separate computer.
The Knights Landing cores (next generation after knights corner) are silvermont (Atom) based with extras including AVX-512 but not the same AVX-512 as Xeon, there’s a hideous Venn diagram about the AVX-512 support in different Intel products. Here’s a tweet with it. I used the standalone (i.e. like a normal CPU) version of knights landing as it was “usable”. Very bad single threaded performance, but if you can use many cores and are memory bandwidth bound, the MCDRAM (a type high bandwidth memory), you might be able to take advantage of one. I would not recommend buying one for a home lab.