Nebius benchmark: 8 small GPU replicas yield 42.5% more Cosmos video per hour
shashib · x · 2026-10-28
Nebius ran NVIDIA's Cosmos 3 Super world model four ways on one 8-GPU server: one large replica, or splits of 2/4/8 replicas. Splitting into eight 1-GPU copies produced 42.5% more video per hour at 30% lower server cost per second, but with 5.5x longer average wait per clip; the single large replica was fastest at 68.5s average. The takeaway: GPU layout determines whether you optimize latency or throughput — and cost per clip — which matters for robotics and self-driving teams generating training videos at scale.
More from Infra
- Hyperbolic raises $30M Series A1 to make fragmented GPU compute easier to use and trade — qinzytech · 2026-10-28
- Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs — zacharynado · 2026-10-28
- GPU capacity bust more likely than a bank run, argues AI compute cycle thread — bookwormengr · 2026-10-28
- Analyst: AI constraint map now tracks 50 components as everything hits bottlenecks — BenBajarin · 2026-10-28
- AXA Uses NVIDIA Earth-2 to Simulate 1,000+ Cyclone Seasons for Storm Risk Modeling — NVIDIAAI · 2026-10-28
- 2026 wafer fab equipment forecast raised to $153B as the cycle lengthens — BenBajarin · 2026-10-28