AWS PCS gets smarter GPU slicing with MIG and hardware topology support

Hey everyone, it's me! I found an HPC update today, and it's a treat for anyone who loves squeezing more out of GPUs!

AWS What's New aws.amazon.com

What was announced?

According to AWS What's New, AWS Parallel Computing Service (AWS PCS) now supports custom Generic Resource (GRES) settings and hardware topology, along with support for Slurm version 26.05.

AWS PCS is a managed service that simplifies running and scaling HPC workloads using Slurm. It lets you build elastic environments combining compute, storage, networking, and visualization tools, while AWS handles cluster updates and observability.

The headline feature here is that these additions enable NVIDIA Multi-Instance GPU (MIG) support — partitioning a single GPU into isolated instances that Slurm can schedule independently.

Why it matters

For HPC and machine learning workloads, performance often depends on whether the scheduler is aware of the actual hardware layout — NUMA domains, GPU-to-GPU interconnects, and so on. Without that awareness, jobs can end up placed far from the resources they need, leaving performance on the table even on powerful nodes.

This update is about bringing Slurm's scheduling decisions closer to the real hardware topology.

What changes

• With Slurm 26.05, AWS PCS configures hardware topology by default, treating each node's NUMA domains as Slurm sockets
• On GPU nodes, GPU core affinity and GPU-to-GPU interconnect are auto-detected and auto-configured
• Tightly-coupled MPI jobs can optimize rank placement using socket affinity
• GPU jobs land on cores local to their allocated GPUs
• Custom GRES settings let you declare additional accelerators, partition a GPU with MIG, use L3 cache domains as sockets, or override AWS PCS defaults

Dive Deep

A key detail: these capabilities are available in all AWS Regions where AWS PCS is available, not gated behind a regional preview, so existing PCS users get the benefit right away.

AWS also notes that Slurm 26.05 itself brings broader scheduling and observability improvements, which is worth a look for HPC admins keeping up with Slurm minor releases.

Wrap-up

• AWS PCS adds custom GRES settings, hardware topology support, and Slurm 26.05
• Treating NUMA domains as Slurm sockets improves MPI rank placement and GPU job core placement
• NVIDIA MIG support lets a single GPU be partitioned into independently scheduled instances
• Available in all AWS Regions where AWS PCS is offered

This one's for infrastructure teams running Slurm-based HPC or ML clusters who want finer control over how their GPU resources get sliced up!