Compute Hardware Overview: Difference between revisions
Phase 1 § 1 P1.1.3 / Phase 2 P2.1: new page — compute node types by partition, with TODO markers for current per-node specs (via create-page on MediaWiki MCP Server) |
mNo edit summary |
||
| Line 8: | Line 8: | ||
<!-- TODO: confirm and fill in — number of CPU nodes, cores per node, memory per node, CPU model. The cluster is heterogeneous, so there may be more than one CPU node type (e.g. standard and high-memory "fat" nodes). --> | <!-- TODO: confirm and fill in — number of CPU nodes, cores per node, memory per node, CPU model. The cluster is heterogeneous, so there may be more than one CPU node type (e.g. standard and high-memory "fat" nodes). --> | ||
We have 3 generations in use right now: | |||
* gen2 : 29 nodes with Intel CPUs, 32 cores and 375G RAM. 2 nodes with Intel CPUs, 64 cores and 1.5T RAM | |||
* gen3: 73 nodes with AMD CPUs, 128 cores and 1T RAM. 2 nodes with AMD CPUs, 96 cores and 4T RAM | |||
* gen4: 3 nodes with AMD CPUSs, 192 cores and 2.2T RAM | |||
== GPU nodes == | == GPU nodes == | ||
Revision as of 12:14, 9 July 2026
Anunna's compute nodes are grouped into partitions by the kind of hardware they provide. This page summarises what each node type offers; for how to request a particular node or feature in a job, see Choosing a node (constraints).
CPU nodes
Standard compute nodes for jobs that do not need a GPU. These make up the main partition and handle the bulk of the cluster's work.
We have 3 generations in use right now:
- gen2 : 29 nodes with Intel CPUs, 32 cores and 375G RAM. 2 nodes with Intel CPUs, 64 cores and 1.5T RAM
- gen3: 73 nodes with AMD CPUs, 128 cores and 1T RAM. 2 nodes with AMD CPUs, 96 cores and 4T RAM
- gen4: 3 nodes with AMD CPUSs, 192 cores and 2.2T RAM
GPU nodes
Nodes equipped with GPUs, for accelerated workloads such as deep learning. Request a GPU with --gres=gpu:<n> and the appropriate partition (see Choosing a node (constraints)).
NVIDIA GPUs
Available in the gpu partition. The cluster has several NVIDIA GPU models — A100, A6000, and V100. The scheduler hands out A100s first, then A6000s, then V100s; the price per GPU-hour is the same for all of them. As a rough guide, the A100 (80 GB) is about twice as fast as the A6000 (48 GB) or V100 (16 GB), depending on whether your workload can use the extra memory. Constrain a job to a specific model with, for example, --constraint='nvidia&A100'.
AMD GPUs
Available in the gpu_amd partition. Requested the same way as NVIDIA GPUs, using the gpu_amd partition.