Darius Baruo
Feb 18, 2026 18:31
NVIDIA and Nebius benchmarks present GPU fractioning achieves 86% person capability on 0.5 GPU allocation, enabling 3x extra concurrent customers for combined AI workloads.
NVIDIA’s Run:ai platform can ship 77% of full GPU throughput utilizing simply half the {hardware} allocation, in accordance with joint benchmarking with cloud supplier Nebius launched February 18. The outcomes display that enterprises operating massive language mannequin inference can dramatically broaden capability with out proportional GPU funding.
The exams, performed on clusters with 64 NVIDIA H100 NVL GPUs and 32 NVIDIA HGX B200 GPUs, confirmed fractional GPU scheduling reaching near-linear efficiency scaling throughout 0.5, 0.25, and 0.125 allocations.
Exhausting Numbers from Manufacturing Testing
At 0.5 GPU allocation, the system supported 8,768 concurrent customers whereas sustaining time-to-first-token beneath one second—86% of the ten,200 customers supported at full allocation. Token era hit 152,694 tokens per second, in comparison with 198,680 at full capability.
Smaller fashions pushed these positive factors additional. Phi-4-Mini operating on 0.25 GPU fractions dealt with 72% extra concurrent customers than full-GPU deployment, reaching roughly 450,000 tokens per second with P95 latency beneath 300 milliseconds on 32 GPUs.
The combined workload state of affairs proved most hanging. Working Llama 3.1 8B, Phi-4 Mini, and Qwen-Embeddings concurrently on fractional allocations tripled whole concurrent system customers in comparison with single-model deployment. Mixed throughput exceeded 350,000 tokens per second at full scale with no cross-model interference.
Why This Issues for GPU Economics
Conventional Kubernetes schedulers allocate entire GPUs to particular person fashions, leaving substantial capability stranded. The benchmarks famous that even Qwen3-14B, the biggest mannequin examined at 14 billion parameters, occupies solely 35% of an H100 NVL’s 80GB capability.
Run:ai’s scheduler eliminates this waste by means of dynamic reminiscence allocation. Customers specify necessities straight; the system handles useful resource distribution with out preconfiguration. Reminiscence isolation occurs at runtime whereas compute cycles distribute pretty amongst lively processes.
This timing coincides with broader business strikes towards GPU partitioning. SoftBank and AMD introduced validation testing on February 16 for comparable fractioning capabilities on AMD Intuition GPUs, the place single GPUs can break up into as much as eight logical units.
Autoscaling With out Latency Spikes
Nebius examined automated scaling with Llama 3.1 8B configured so as to add GPUs when concurrent customers exceeded 50. Replicas scaled from 1 to 16 with clear ramp-up, steady utilization throughout pod warm-up, and negligible HTTP errors.
The sensible implication: enterprises can run a number of inference fashions on present GPU stock, scale dynamically throughout peak demand, and reclaim idle capability throughout off-hours for different workloads. For organizations dealing with mounted GPU budgets, fractioning transforms capability planning from {hardware} procurement into software program configuration.
Run:ai v2.24 is accessible now. NVIDIA plans to debate the Nebius implementation at GTC 2026.
Picture supply: Shutterstock
