Multi-GPU and NUMA Configs Significantly Impact Training Speed
AsamiKai · reddit · 2026-07-14
This post shares performance observations during LoRA training on a specific hardware setup: the author used a 5975WX CPU, quad-channel memory, and two 3080 20G GPUs, with no NVLink or ResizeBar, but communicating via NCCL over PCIe.
Key Factors Affecting Speed
- gloo / nccl
- trainbatchsize and gradientaccumulationsteps
- NUMA node setting: The author considers this the most impactful factor
Benchmark Results
- Setting NUMA nodes to 4 resulted in a training speed of 25s/it
- Setting NUMA nodes to 0 boosted the speed to 12s/it
The author explains that on this machine, NCCL routing data over PCIe combined with cross-NUMA CPU communication introduces significant latency, often leaving GPUs idling while waiting for data. Consequently, in high-latency communication scenarios, GPU training speed is largely insensitive to power limits; slightly lowering GPU power might actually be more cost-effective.
More from Infra
- NVIDIA publishes Vera CPU architecture details before AMD’s AI event — ryanshrout · 2026-07-22
- oMLX 0.5.2 adds Mac menu-bar stats, low-bit decode kernels, and faster downloads — awnihannun · 2026-07-22
- Strangeworks launches Aura to turn enterprise ops into production optimization systems — whurley · 2026-07-22
- Graph workload 854.graph500 enters SPEC CPU 2026 as a new CPU benchmark — Prof_DavidBader · 2026-07-22
- HilbertRaum open-sources a fully local AI chat and document analysis app for private use — Vladowski · 2026-07-22
- Hybrid and local inference are emerging as a response to AI energy and token costs — dmitry140 · 2026-07-22