The University of Chicago's Research Computing Center measured NVIDIA's B300 running a 70B Llama-3 model at 24% to 27% lower throughput than the previous-generation B200.

(preprint, arXiv 2610.41919)

Concept illustration: a hand at the right holds a black, round-cornered screen sideways, with a black strip along the screen's bottom lettered "UNIVERSITY OF CHICAGO" and, on the surface above it, a beige plaque at the centre lettered "NVIDIA" beside a green emblem. Left of the plaque stands a dark grey server rack lettered "B200" with four speech bubbles containing lines above it, right of the plaque stands a brick-red rack lettered "B300" with three such bubbles above it, and a stack of lined paper sits on the outer side of each rack. Printed on the image: 24% to 27% B300 throughput below B200 on 70B, Chicago test.

A new GPU generation isn't guaranteed to beat its predecessor on every workload, though only this one team has measured it and software versions may sway the result. Mohsen Zand, Trung Dac Nguyen and H. Birali Runesha published on October 30, comparing the two GPUs on two independent eight-GPU clusters (arXiv 2610.41919). The authors also found that two B300 servers of the same model, each running a different development build of vLLM, differed by 2.1 times in time to first token, the wait from sending a request to seeing the first output token. vLLM is widely used inference software. Outside the 70B language model, the gaps are larger. The B200 is 16 times as fast as the B300 on double-precision computation in the materials simulation code Quantum ESPRESSO. It is 8 times as fast on 3D vision transformer training and up to 5.6 times as fast on convolutional image inference. On the workloads where it trails, the authors observed the B300 drawing only 36% to 53% of its 1,100-watt power limit. On the image model ViT-Large, where the B300 wins, both GPUs reach 95% to 97% of their power limits. The abstract doesn't attribute each shortfall to a specific cause. How we checked →

Source and checks

This item ran in the brief of Nov 2, 2026, alongside 4 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "The University of Chicago's Research Computing Center measured NVIDIA's B300 running a 70B Llama-3 model at 24% to 27% lower throughput than the previous-generation B200.." https://secondsource.io/en/research/2026-11-02-1/. Accessed access date.