At #ISC23, @intel's Jeff McVeigh going through AI-accelerated #hpc. Either AI helping reduce large problems, or AI hardware being used for reduced precision in HPC 
If you hadn't seen it, Intel's AI roadmap. Falcon Shores is the output of GPU+AI.
GPU Max.
128 Xe Cores
128 GB HBM2e
52 TF FP64
839 TOP BF16
128 Xe Cores
128 GB HBM2e
52 TF FP64
839 TOP BF16
And just to clarify, Falcon shores first gen is a converged Xe + AI (habana) IP design. First point here. Even though its called a GPU overall.
The status on #aurora: 10000 blades delivered, each with 2 CPU + 6 GPU. Blades initially shipped with regular @intel CPU, now 75% replaced with Max CPU (hence why no #Top500 sub).
But 63744 GPUs at 52 TF/GPU means an Rpeak of around 3.3 ExaFLOP FP64!! That's not counting CPUs
But 63744 GPUs at 52 TF/GPU means an Rpeak of around 3.3 ExaFLOP FP64!! That's not counting CPUs
(That's assuming each GPU is the full fat 600W node - I need to go back to my original estimates I did on AnandTech where I calculated 560W ea)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
