The Japanese PEZY machine looks incredibly interresting. Too bad it is only used in that one super-computer.

Its overall design sits somewhere between a GPU and CPU. Like a GPU, it supports 8-way hyperthreading (GPUs typically support 10. CPUs usually only support 2... maybe 4 for Larabee). Like a GPU, its programmed through an OpenCL-like language (I know you don't like that, but its how it works).

However, the overall layout is very CPU-like as well: a "City" is composed of "Villages", which is composed of multiple "Processing Elements". Each PE is scalar like a normal CPU, although 8-way SMT. So its not like a GPU's SIMD or SIMT architecture in this regards.

--------

See information here:

https://en.wikichip.org/wiki/pezy/pezy-scx/pezy-sc

It really is an interesting design that toes the line between CPU and GPU.

Is PEZY still a going concern after the arrests?

So that's why Xeon Phi has been so successful!

/s

Real talk, though. Why do you think larabee is the "right" hardware?

I'd assume because of the extra serial processing power of a CPU.

x86 isn't the right choice though. As the chips get smaller, all the x86 cruft starts to become a larger portion of each core.

RISCV with 40-ish instructions plus a SIMD extension would probably be ideal. They've already proven that those designs use less space per core than ARM.

The truth is that for many types/sizes of models AVS2 or AVS512 in Xeons is as fast and GPUs.

I am interested in seeing benchmarks regarding this. I believe this is true for models with truly huge embeddings, but otherwise if your models are even a little compute dense then GPUs are faster.

My experience is that you need the additional caveat of a streaming fairly homogeneous and highly parallelizable approach or the memory transfer, branching and communication overheads will eat away nearly all the GPU gains.

I've also noticed that a lot of the time, people are comparing GPU to naive implementations (and sometimes, implementations written in dynamic languages) instead of to highly tuned BLAS or MKL implementations. For a large swathe of problem types, using the fastest math library will reduce the CPU/GPU gap to less than an order of magnitude or less.

No, they're not. Straight front the horse's mouth:

https://software.intel.com/en-us/mkl/features/benchmarks

In no case will you see it get close to 10TFLOPS. GPUs easily do this, and can approach 100TFLOPS with tensor cores.

I don't dispute this. If you can remove memory transfer overhead and meet the requirements I mentioned, then GPUs will be much better. But there are many problems that do not fit that (SIMD per warp) regime. In short, GPUs are no panacea and have their own trade-offs and bottleneck sensitivities, just like everything else.

I agree that they are no panacea, but we also have an example of the Intel Xeon phi, which did have high speed hbm memory. That too did not compete with GPUs. I think the majority of it has to do with the GPU chip itself being extremely simple, with no Branch prediction, no branch difference, very simple caching behavior, etc. They are able to allocate a much larger portion of the die to just throw fma units at it.

> comparing GPU to naive implementations (and sometimes, implementations written in dynamic languages) instead of to highly tuned BLAS or MKL implementations

In your experience, are these comparisons using naive GPU implementations as well? If so, that still seems like a valuable comparison.

Well, I'm speaking from personal experience, for many (but not all) text-classification IO-intensive models with large embedding tables Xeon on TF compiled with AVX2 or AVX514 and FMA is as fast as GPU. Obviously these models need to be on the lower spectrum of computational complexity.

This is not true. I've benchmarked it, and you can see yourself by looking at the MKL benchmarks versus cuBLAS. cuBLAS is always significantly faster.