Recently it was announced that NVIDIA is working on a Hopper H100 PCIe accelerator card featuring up to 120GB of HBM2e memory. To date, the company has released two versions of the Hopper H100 GPU, the SXM5 and PCIe board models. While both SKUs come with 80GB of VRAM, the former uses the new HBM3 standard, while the latter remains HBM2e.
The latest news is that according to s-ss.cc, NVIDIA may be working on a new PCIe expansion card version of the Hopper H100 GPU — but instead of 80GB of HBM2e VRAM, it has been increased to 120GB.
Sources say the new card comes with six HBM2e stacks, 6144-bit bus bit width @ 120GB VRAM, and is complemented by the same GH100 GPU as SXM5. The total of 16,896 CUDA cores, over 3 TB/s of bandwidth, and 30 TFLOPS of performance per precision -- comparable to the SXM5 version.
Based on these specifications, the NVIDIA Hopper GH100 GPU has 144 SM stream processors / 8 GPCs – each GPC consists of 9 TPCs, with each TPC consisting of 2 SM units.
Each group of SM units consists of up to 128 FP32 units, so the full-blooded version would be 18432 CUDA cores.
GH100 GPU full specs reference.
● 8 GPCs, 72 TPCs (9 TPCs/GPCs), 2 SM/TPCs, 144 SM units per GPU ● 128 FP32 CUDA cores per SM unit and 18432 FP32 CUDA cores per GPU. ● 4 fourth-generation tensor cores per SM unit, with a full 576 Tensor Cores per GPU. ● 6 HBM3 or HBM2e memory stacks, complemented by 12 @ 512-bit memory controllers. ● 60 MB of L2 cache
Here are the NVIDIA H100 GPU specifications for the SXM5 form factor.
● 8 GPC / 66 TPC, 2 SM / TPC, 132 SM units per GPU ● 128 FP32 CUDA cores per SM unit and 16896 FP32 CUDA cores per GPU. ● 4 4th generation tensor cores per SM unit with 528 Tensor Cores per GPU. ● 5 @ 80GB HBM3 memory stacks, complemented by 10 @ 512-bit memory controllers. ● Equipped with 50MB of L2 cache ● Supports 4th generation NVLink and PCIe 5.0
It’s unclear if NVIDIA is working on a test prototype or if a future iteration of the Hopper H100 GPU is in the works.
But the company recently said at GTC 2022 that the Hopper GPU is now in full production, with the first products expected to arrive next month.
