Our team measured NUMA Affinity's impact on a DeepSeek V3 pretraining workload running…
This is a dev post classified by AGK Intelligence as Hosting & infra (a tutorial), kept by the Dev Radar because it carries real work, not commentary.
Our team measured NUMA Affinity's impact on a DeepSeek V3 pretraining workload running with Megatron Bridge on 192 NVIDIA B300 GPUs in Managed Service for Kubernetes on Nebius AI Cloud. Exclusive NUMA binding reduced mean training step time by 1.11% and increased model throughput per GPU by 1.13%. In this post, we’ll explain how NUMA-aware process affinity works, compare it with the controls Slurm users already have, and walk through the measurement.
Posted by Nebius (43.1k followers) 3 h ago · 145 likes · 18.3k views · view the original post on X. Kept by the Dev Radar as Hosting & infra.
More dev work like this
- we just reached >2000 active paying users for https://boat.dev's cloud VMs! — @AniC_dev
- Checkout devvm - near instant VMs for every project folder. — @CodingGarden
- OpenAI, Anthropic, Cursor and Cognition all let customers bring their own sandbox. — @guilleflorvs
- Frontier AI data lab @SnorkelAI uses Modal Sandboxes and Modal Inference to power… — @modal
- Today’s class will be technical — @TheCloudSE
- Introducing Aspen: an open-source Git server built for agents. — @reasonmachines
- nvidia put out NVKE, managed kubernetes for your own gpu boxes. preview catch: api auth… — @ShubhamMal72313
- My lxd thin container on gb10 keeps crashing on pytorch start. 🫠 Turns out I limited… — @qubitium
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.8k posts from 5k X accounts over the last 21 days, 9.5k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 17:41 UTC. Full method.