Researchers benchmark NVIDIA B200 vs B300 Blackwell

Researchers evaluate NVIDIA B200 and B300 Blackwell Ultra GPUs across clusters spanning computer-vision inference and training, LLM inference, and production HPC codes. They find that the B300 trails the B200 in several workloads, such as convolutional inference and double-precision Quantum ESPRESSO, because it draws a lower percentage of its power envelope on those tasks. The authors resolve these performance deficits and offer per-workload procurement guidance based on board power usage and kernel costs. Open article →