Computing Hardware

The CRC provides different hardware types to target different computing use cases. These hardware profiles are grouped together under a common cluster name and are further divided into partitions to highlight differences in the architecture or usage modes.

Below, you will find the hardware specifications for each cluster and the partitions that compose it: 

MPI Cluster

The MPI nodes are for tightly-coupled codes that are parallelized using the Message Passing Interface (MPI) and benefit from low-latency communication through an Infiniband (HDR200) or Omni-Path (OPA) network. Your job must request a minimum of 2 nodes.

HTC Cluster

These nodes are designed for High Throughput Computing workflows such as gene sequence analysis, neuroimaging data processing, and other data-intensive analytics.

SMP Cluster

The SMP nodes are appropriate for programs that are parallelized using the shared memory framework. These nodes are similar to your laptop but with more memory and more CPU cores. To request a particular feature (such as an Intel host CPU), add the following directive to your job script:

 #SBATCH --constraint=intel

Multiple features can be requested by providing a comma-separated list (without intervening spaces):

 #SBATCH --constraint=amd,genoa

GPU Cluster

The GPU nodes are targeted for applications specifically written to take advantage of the inherent parallelism and massive amounts of cores in the architecture. We name the partitions after the GPU type along with a suffix as needed to indicate usage mode. The partition parameters are described below.

• Partition: l40s. This partition is appropriate for AI, simulations, 3D modeling workloads that require up to 4x gpus on a single node and rely on single or mixed precision operations (Note: This partition does not support double precision - FP64).
• Partition: a100. This is the default partition in the gpu cluster and is appropriate for workflows that require up to 4x gpus on a single node. To request a particular feature (such as an Intel host CPU), add the following directive to your job script:

 #SBATCH --constraint=intel

Multiple features can be specified in a comma-separated string.

• Partition: a100_multi. This partition supports multi-node GPU workflows. Your job must request a minimum of 2 nodes and 4 GPUs on each node.
• Partition: a100_nvlink. This partition supports multi-GPU computation on an Nvidia HGX platform with 8x A100 that are tightly coupled through an NVLink switch. To request a particular feature (such as an A100 with 80GB of GPU memory), add the the following directive to your job script: 

#SBATCH --constraint=80g

• Partition: gtx1080. Older gaming GPUs with 11GB of memory
• Partition: v100. Tesla V100 GPUs with 32GB of HBM2 memory
• Partition: power9. Four nodes of IBM Power System AC922: dual-socket Power9 (16C, 2.7GHz base, 3.3GHz turbo) with a direct NVLink to 4x V100 GPUs. Code must be compiled for the Power9 platform in order to work.

Portal: Viz Nodes

Users should open the Web URL in a browser to access a Linux Desktop. The Web URL will resolve to a backend hostname in a round-robin fashion to help balance usage.

Portal: Login Nodes

Users should use the hostname as the remote host target of an ssh session. The hostname will resolve to a backend hostname in a round-robin fashion to help balance usage.