NVIDIA¶
Work in Progress
This page is not yet complete. Content is being added gradually. Functionality may be incomplete or subject to change. Thank you for your patience!
The FTP provides access to NVIDIA GPU and CPU systems covering both established GPU-accelerated HPC nodes and newer ARM-based platforms. This allows users to evaluate NVIDIA hardware across different generations and integration levels — from discrete GPU acceleration to tightly coupled CPU–GPU designs.
All NVIDIA nodes are part of the ARM cluster and are accessible via the ARM login node.
Hardware Overview¶
Two Grace-based system configurations are available:
| Node type | Description | Queue |
|---|---|---|
| Grace-Grace | Two Grace CPU modules connected via NVLink-C2C, no discrete GPU | grace_grace |
| Grace-Hopper | One Grace CPU module + one Hopper GPU (H100), connected via NVLink-C2C | nvidia100_2 |
The NVLink-C2C interconnect provides extremely high bandwidth between CPU and GPU (or between two CPU modules), enabling a unified memory model where the CPU and GPU share the same physical memory pool. This eliminates the need for explicit data transfers between host and device.
The Grace-Hopper configuration is particularly well-suited for workloads that benefit from large, coherently shared memory between CPU and GPU, such as graph analytics, sparse linear algebra, and large-model inference.
Access¶
NVIDIA nodes are available to all FTP users. See the FTP overview for how to request access.
Using NVIDIA Nodes¶
Connect to the ARM login node and submit jobs to the desired partition:
Use sbatch or salloc to access a partition, e.g.:
or, for interactive use:
Software for NVIDIA¶
NVIDIA Grace nodes support CUDA (on Grace-Hopper), OpenMP, and OpenACC for GPU offload. The standard NVIDIA HPC SDK and CUDA toolkit are available. For Grace-Grace nodes, standard ARM-compatible compilers (GCC, LLVM, NVHPC) can be used.