Meta tests AMD's Helios AI servers; software tuning yields 1.5x training speedup
shashib · x · 2026-10-27
AMD announced on October 27 that Meta has installed the first of its new Helios AI server cabinets — refrigerator-size racks of chips acting as one machine — for testing ahead of wider rollout. Meta already runs revenue-critical work on AMD silicon: Instinct accelerators train models and rank the feed posts and ads that drive most of its revenue, while millions of EPYC server chips power the Muse AI assistant. Meta's engineers made AI training up to 1.5x faster by tuning software for AMD hardware, and AMD agreed to supply 6 GW of AI chips. Meta's newest models also ran on AMD hardware, including Ryzen AI PCs, from day one.
Related event: AMD Deepens Meta Partnership as Helios AI Servers Go Live(2 posts)→
More from Infra
- Agents built a custom inference engine: AMD MI350X serves GLM 5.3 FP8 37% faster than NVIDIA B200 — yisongyue · 2026-10-28
- Anthropic's S-1 admits its data centers rely on natural gas, exposing it to fuel costs — TobyWalsh · 2026-10-28
- Kimi 3.1 on Vera Rubin costs just 0.025 cents per million tokens, per TCO math — bookwormengr · 2026-10-28
- Hyperbolic raises $30M Series A1 to make fragmented GPU compute easier to use and trade — qinzytech · 2026-10-28
- Tailscale SSH: Investigation and remediation update — figmert · 2026-10-28
- Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs — zacharynado · 2026-10-28