| Facility | Comet |
|---|---|
| Facility Status | Planning |
| Node Type | GPU-L |
| Node Quantity | 1 |
| RAM | 768 GBytes |
| CPU | 2x CPU : AMD EPYC 9745 (x86_64) |
| GPU | 4x GPU Type : Nvidia H100 96GB |
| Reserved | 0% for unfunded use |
| Reserved | 100% for funded use |
| CPU Vendor | AMD |
|---|---|
| CPU Model | EPYC 9745 |
| CPU Architecture | x86_64 |
| CPU Cores | 128 |
| CPU Threads | 256 |
| Base Frequency | 2400MHz |
| Turbo Frequency | 3700MHz |
| Cache (L1) | 6144 KBytes |
| Cache (L2) | 131072 KBytes |
| Cache (L3) | 256 MBytes |
| GPU Vendor | Nvidia |
|---|---|
| GPU Model | H100 |
| GPU RAM | 96 GBytes |
| GPU Speed | 3360 GBytes/Sec |
| CUDA Support | 9.0 |
| CUDA Cores | 16896 |
| Speed (FP64) | 31040.0 GFLOPS |
| Speed (FP32) | 62080.0 GFLOPS |
| Speed (FP16) | 248300.0 GFLOPS |
This is the larger type of GPU node as found in the Comet HPC facility. These have the same general CPU performance as all other compute nodes, but are also fitted with Nvidia H100 (96GB) PCIe cards for accelerated compute.
These nodes are also connected by low-latency, high-speed 200gbit/sec Infiniband networking.
We have only a limited number of these cards (four per node), if your workload can run in lower precision FP32 or FP16 mode, then you should consider designing your workflow to use the smaller GPU-S nodes with the L40S cards which are both more numerous and have a lower resource charge cost.
Due to the limited number of H100 cards, these nodes are not currently available for free (unpaid) use.
All registered users of the Comet facility may access any of the hours of compute available in the unfunded queues, as listed below. However, your jobs may need to wait longer to run due to the number of resources available. Consider building in costing for your research activity or project for dedicated HPC resources to enable access to the larger pool of resources dedicated for funded queues.
The maximum hours of GPU compute available across all users of unfunded queues is:
(1 Nodes) x (4 Cards) x (730 Hours/Month) x (0%) = 0.0 GPU Compute Hours/Month
The maximum hours of GPU compute available across all users of funded queues is:
(1 Nodes) x (4 Cards) x (730 Hours/Month) x (100%) = 2920.0 GPU Compute Hours/Month
You can use the for the Comet facility to model the potential contributions towards accessing these resources.
Note: You must be logged in to access the costing tool.