Intel Shifts Its HPC Max Series Focus To Falcon Shores In 2025

[sabian2008]

Has anyone ever heard of companies actually using Ponte Vecchio?

Benchmarks are excellent... but every git project I have read is trained on a Nvidia x100 cluster. And there were maybe 2 repositories than mentioned ROCm/CDNA on a single line in passing.

You are being too swift on the timescales. Ponte Vecchio just became publicly available in January (although select customers, for instance Argonne, got access to the hardware in the second half of 2022). I'm in the team building up (from the client side) a presumably Top100 system using Sapphire Rapids and Ponte Vecchio, so I hope I can get some numbers of my own by the end of Q2. The GPUs, as far as private benchmarks I was able to peek (both from Intel and selected customers) seem great, specially in flops/$, which is the relevant metric when building a general purpose HPC system (real flops for typical workloads, not theoretical).

[brucethemoose]

You are being too swift on the timescales. Ponte Vecchio just became publicly available in January (although select customers, for instance Argonne, got access to the hardware in the second half of 2022). I'm in the team building up (from the client side) a presumably Top100 system using Sapphire Rapids and Ponte Vecchio, so I hope I can get some numbers of my own by the end of Q2. The GPUs, as far as private benchmarks I was able to peek (both from Intel and selected customers) seem great, specially in flops/$, which is the relevant metric when building a general purpose HPC system (real flops for typical workloads, not theoretical).

Interesting 👍

I have also seen support for Habana from huggingface and maybe somewhere else, now that I think about, but I'm not sure how (at $13.2/hr) it compares to the A100 instances.

[pegasus]

I'm in the team building up (from the client side) a presumably Top100 system using Sapphire Rapids and Ponte Vecchio,

Only question that matters to me at the moment - does it run AlphaFold?