GannettPeak Hardware Details
The GannettPeak Compute Environment (Advanced Infrastructure to Accelerate Impact of AI through Applications and Innovation for Wyoming) is a high-performance computing (HPC) cluster consisting of NVIDIA’s Hopper GPUs with ARM CPUs onto a single chip with fast memory interconnect. This unique architecture addresses critical computing challenges related to data movement and energy consumption through the Grace Hopper architecture, which integrates high-bandwidth coherent data transfers between CPU and GPU. The cluster was funded through an award from the NSF. This cluster is made of 2 login nodes, 12 single GPU Gracehopper Nodes, and 24 dual GPU Gracehopper Nodes. Nodes are named based on usage and compute components.
Hardware Table
| Hostnames | Usage | Chip GPU | RAM |
|---|---|---|---|
| ai4wy-log1, ai4wy-log2 | Login Nodes | N/A | N/A |
| ai4wy-101 through ai4wy-112 | Compute Nodes | Single GH200 Superchip including 1 NVIDIA H100 GPU & 72 ARM Neoverse V2 Cores | 480 GB ECC On-Board LPDDR5X Memory and 144GB ECC HBM3 |
| ai4wy-201 through ai4wy-224 | Compute Nodes | Dual GH200 Superchip including 2 NVIDIA H100 GPU & 144 ARM Neoverse V2 Cores | 480 GB ECC On-Board LPDDR5X Memory and 144GB ECC HBM3 (144GB per GPU) |
Node Architecture
Single GH200 Node Architecture
The following diagram from NIVIDA details the design of the Gracehopper chip. Our single GH200 Nodes contain 1 Gracehopper chip per node.
Dual GH200 Node Architecture
Dual GH200 Node architecture is similar, but contain two chips.
References
Manufacturer documentation
1. NVIDIA's Gracehopper Superchip Data Sheet
2. NVIDIA's Gracehopper Tuning Guide