Benchmark figures released at the Hot Chips conference indicate OpenAI's custom inference processor delivers up to 1.9 times more work per watt than Nvidia GB300 systems.
