Over the past several years FASRC has seen remarkable growth in GPU demand. The gpu partition has been regularly oversubscribed with near constant 100% utilization. Looking at future trends, GPU computation will only become more popular.

With this in mind, FASRC is pleased to announce the new gpu_h200 partition on Cannon. This partition is made up for following:

• 24 Lenovo SD650-N v3 servers with each with 4 Nvidia H200 GPUs - 96 H200 GPUs total
• 112 Intel Sapphire Rapids cores with 1TB of memory 
• Each H200 has 141 GB of onboard memory 
• A theoretical peak of 51 TFlops at double precision with a total of 4.8  PFLOPS of compute capacity. 

The nodes are connected via NDR Infiniband for low latency interconnectivity. This new partition has a time limit of 3 days.

The new partition permits FASRC to update the gratis fairshare for groups on the cluster. Per our policy we split this additional fairshare across all 500 groups that use Cannon. To that end the new gratis share for Cannon is 250, an increase of 50 shares over the previous 200.

For those looking for additional GPU capacity beyond what is provided by gpu and gpu_h200, we highly recommend looking at gpu_requeue. New GPU’s are being added to that partition constantly as various groups that use the cluster purchase additional hardware. It is a great location to farm cycles for larger projects, just be advised that your job will be preempted by higher priority work.

For a complete list of the GPU’s available on Cannon see the Running Jobs page.