ZLUDA v2 Released For Drop-In CUDA On Intel Graphics

[Paradigm Shifter]

AVX-512 was no match for GPUs before they started adding things like tensor cores.

And AMD's problem (in AI) was always competing against the previous Nvidia GPU. But it looks like their Matrix cores might've finally managed to leap-frog Nvidia, at least for some use cases. Now, if they can just get out of their own way and work on the software situation. However, to be really successful, they're going to have to find a way to build more enthusiasm for their GPU and compute products.

AVX-512 gets "close" for our tasks... provided you throw enough cores and RAM at the problem (min. 48 cores, 96 threads). The main problem I have is that the minimum price of entry for AVX-512 capable system that is actually worth using is ~3,000,000JPY. And that gets "close" to a CUDA-based box at ~40% the price (beats in some things, loses in others).

I agree with your comment about AMD compute. The AMD hardware is nice, but setup and software is more troublesome.

[qarium]

Yes, agreed. It all depends on how it's implemented. We don't use Tensor cores (at least currently) and 99.9% of the time, 11GB/12GB cards are more than enough for what we do... it's just those few occasions where we do need more, we need a lot more.

i think AMD works on such cards... with a lot more "ram" it will be some DDR5 (not GDDR5) cards with "infinity cache"
with a card like this you can have 256GB VRAM...

yes it will not be as fast as GDDR6x or HBM3 but in the case you really need more VRAM it will help.

[coder]

AVX-512 gets "close" for our tasks... provided you throw enough cores and RAM at the problem (min. 48 cores, 96 threads). The main problem I have is that the minimum price of entry for AVX-512 capable system that is actually worth using is ~3,000,000JPY. And that gets "close" to a CUDA-based box at ~40% the price (beats in some things, loses in others).

Wow, I wish I had those kind of hardware budgets, at my job. We're too cheap to buy Intel server CPUs with even dual AVX-512 units per core! At the low end, the server Xeon processors have only one enabled, in each core. Intel also constrains the lower-end HEDT CPUs in the same way, but I think all Xeon W processors have dual.

As for whether AVX-512 can beat GPUs, I guess it has a lot to do with how well your workload vectorizes. GPUs certainly have better raw compute specs.