Joule 3.0 gpu partition nodes feature two AMD 9534 EPYC processors, 1.5 TB of DDR5 RAM, and eight NVIDIA H100 GPU cards.
Each NVIDIA H100 SXM GPU is composed of multiple GPU Processing Clusters (GPCs), Texture Processing Clusters (TPCs), Streaming Multiprocessors (SMs), L2 cache, and HBM3 memory controllers including:
- 8 GPCs, 66 TPCs, 2 SMs/TPC, and 132 SMs per GPU
- 128 FP32 CUDA Cores per Streaming Processor for a total of 16,896 FP32 CUDA Cores per GPU
- 4 Fourth-generation Tensor Cores per SM for a total of 528 per GPU
- 80 GB of HBM3 Memory
The GPUs are installed on a SXM5 board that provides a high-speed NVLink 4 interconnect. This allows for rapid data transfer between GPUs.
