Intel Max GPU series for HPC, AI and Data Centers

[DavidGraham]

So, Intel released their AI/Data Center GPU finally, a 128 XE tiled cores GPU with a 600W TDP, 128 GB of HBM memory, 400MB of L2 cache and 64MB of L1 cache. Intel called it the Max 1550 GPU.

The GPU appears to compete with A100, Intel is stating 1.5x performance advantage in physics simulation, and 2.5x in Finance calculations. NVIDIA's H100 is not mentioned though, so it appears the Max 1550 doesn't compete with it.

Intel introduces MAX Series CPUs and GPU featuring Sapphire Rapids-HBM and Ponte Vecchio - VideoCardz.com

Intel Max Series Brings Breakthrough Memory Bandwidth and Performance to HPC and AI New family of products delivers CPU with high bandwidth memory, 4.8x faster than the competition, and Intel’s highest density GPU, to solve the world’s biggest challenges. What’s New: In advance of Supercomputing...

videocardz.com

 

[Frenetic Pony]

It's hilariously late, and not competitive with Nvidia or AMD in their respective specialties of AI/FP64, but it's good to see an Intel Chiplet ultra wide simd arch that's going to be shipping. Hope for the future!

 

[DavidGraham]

The MI250X (2 GPUs) vs Max 1550 vs H100 SMX5 in theoretical metrics:

FP32: 48 vs 52 vs 67
FP64: 48 vs 52 vs 33
FP64 Matrix: 97 vs 0 vs 67
TF32 Matrix: 97 vs 420 vs 495 (419 vs 990 with Sparsity)
FP16 Matrix: 383 vs 840 vs 990 (383 vs 840 vs 1990 with Sparsity)
BF16 Matrix: 383 vs 840 vs 990 (383 vs 840 vs 1990 with Sparsity)
INT8 Matrix: 383 vs 1680 vs 1980 (383 vs 1680 vs 3860 with Sparsity)
INT4 Matrix: 383 vs 3350 vs 3960 (383 vs 3350 vs 7920 with Sparsity)

The MI250X excels in FP64 Matrix throughput, while the Max 1550 excels in the FP64 throughput, the H100 excels in the rest. H100 also enjoys a large lead as it can use Sparcity to boost performance, while the others can't.

 

[trinibwoy]

Max 1100 under the microscope. Truly an engineering marvel. Performance not so much.

https://chipsandcheese.com/2023/09/23/intels-ponte-vecchio-chiplets-gone-crazy/

Intel’s measured FP64 throughput is nowhere near the marketing numbers.