The ITS Research Computing team is excited to announce that NVIDIA H100 GPUs are now available on the OrangeGrid cluster. These nodes are part of the general OrangeGrid pool and are requested the same way as any other GPU resource.
What Is Available
• 2 nodes with 8 NVIDIA H100 SXM GPUs each, for a total of 16 H100s
• 80 GB of HBM3 per GPU, with NVLink between the GPUs inside a node
Three-Day Runtime Limit
Jobs on the H100 nodes are capped at three days of runtime. We have 16 of these cards and demand for them is high. Without a cap, a small number of long-running jobs would hold the entire resource for weeks and prevent others from getting time on it.
Please request these GPUs only when you are ready to use them. Claiming a card ahead of when your work actually needs it disrupts our scarcest resource.
Checkpointing
If your work runs longer than three days, please checkpoint it. Beyond working within the runtime cap, checkpointing is the cheapest protection available against losing a week of compute to a hardware failure, a power event or an unplanned node outage. A checkpointed job can requeue and pick up where it left off rather than starting over. Examples and guidance are available at the following link:
Running Newer CUDA Versions
You are not limited to the CUDA version installed on the node. The CUDA compatibility library lets you install the CUDA version your code needs into a Conda environment and run against it, even when the node’s drivers are older. This matters because packages are increasingly built against CUDA 13. Instructions are available here:
JAX
JAX is a Python library that provides just in time compilation of mathematical operations and can distribute data structures across multiple GPUs. It works as a near drop-in replacement for numpy, so it is straightforward to try against existing code. If you would like a starting point, documentation is available here:
Good Neighbor Policy
We have published a short Research Computing “Good Neighbor” policy covering the basic expectations for shared resources. It addresses keeping login node activity light, not sharing credentials and not claiming scheduled resources before you need them. It also covers tools acting on your behalf, since an agent running under your account is still your responsibility. Please take a few minutes to review it:
Departmental Introduction Sessions
We are scheduling introductory sessions with departments this fall to walk through what Research Computing offers. If your group would like one, please contact researchcomputing@syr.edu.