gpt-6IMAGE
NVIDIA and OpenAI discuss Astra self-optimizing on Rubin
At AI Infra Summit, Ian Buck and Sachin Katti described Astra running on Vera Rubin with a 3x out-of-box throughput gain, then acting as an agent to optimize its own inference for another 2x over 72 hours.
Clark Tang
@_clarktangArchived 2026.09.16Originally posted 2026.09.16
Clark Tang @_clarktangOriginally posted 2026.09.16
Ian Buck (NVDA) and Sachin Katti (OpenAI) at AI Infra Summit: Astra on Vera Rubin: Recursive Optimization - OpenAI took a Blackwell-optimized Astra checkpoint, ran it on Rubin w/ minor modifications - Out of box: 3x throughput improvement - Then ran Astra as agent to optimize its own inference on Rubin chips - Over 72 hours, the model produced an additional 2x throughput improvement - Work that previously took weeks / mos completed over weekend
x.com/_clarktang/status/2099976574296846801Archived 2026.09.16 10:17