OpenAI tests homegrown AI chips

TL;DR

OpenAI is testing 'Jalapeño,' its first family of homegrown AI chips.
The chips are designed for inference, aiming to increase computing capacity, lower costs, and reduce reliance on Nvidia.
Developed with core design by OpenAI and expertise from Broadcom, the chips are intended for more efficient use and better performance per watt.
Initial testing involves tasks like answering Codex queries, with enhanced thermal performance reported.
Commercial use is expected by the end of the year with partners like Microsoft, with substantial volume anticipated next year.
OpenAI aims to power 10 gigawatts of compute using custom chips by 2029.
This move aligns with industry trends where major AI companies design their own silicon.
OpenAI is considering expanding the use of these chips to training, in addition to inference.

Continuer la lecture de l'article original