the hopper gh100 is expected to be the largest single-gpu design in nvidia’s history and will support hbm2e ” high-bandwidth cache and other advanced connectivity features
However, this is not the end, considering that Hopper will introduce GPUs for multi-chip modular (MCM) solutions, we can expect two GH100 GPUs fused on a single substrate – the die area alone exceeds 2000 square meters.
Trademark Application Disputes
It is reported that the Hopper GPU will be manufactured using TSMC’s 5nm process, and the combination of two is a total of 288 SM units.the number of cores in each group of sm is unknown
At 64 CUDA units per set, we’d see 18,432 cores — a ferocious spec that’s 2.25 times that of a “full-blooded” Ampere GA 100 GPU.
In addition, NVIDIA can make Hopper GPU use more FP64, FP16 and Tensor cores, which is also related toIntelAn important prerequisite for Ponte Vecchio (expected to be 1:1 FP64) bayonet.
134 of the 144 sm units should be allowed on each gpu module if the possibility of a single core remains
But without GPU Sparsity, this feature may be difficult to achieve withAMDSame FP32/FP64 performance level as the Instinct MI200.
discussion about the single core gh100 continues
nvidia may still be hiding a secret weapon – the hopper gpu based on the copa scheme
earlier this year amd announced two domain-specific copa-gpus that are based on next-generation architectures targeting the high-performance computing hpc and deep learning
hpc derivatives typically opt for a standard solution consisting of mcm gpu designs with hbm mc hbm io chips
In contrast, DL derivatives are much more interesting, like having huge caches on completely separate dies where GPU modules are interconnected.