OpenAI said it improved chip power, performance and area metrics by 14% in less than five months in collaboration with AMD. The company estimates that this amounts to roughly a year of additional progress on top of AMD engineers’ years of work. Some of the changes have already been incorporated into a design sent to manufacturing. The same method also helped develop Jalapeño, OpenAI’s first in-house inference chip.
OpenAI accelerated AMD chip optimization by a year in five months
From GPUs to custom chips
OpenAI tested the approach on AMD’s flagship GPU. The work focused on improving three metrics together: power consumption, performance and die area, collectively known as PPA. AMD has already incorporated some of the changes into a design it sent to manufacturing.
The method was also applied to Jalapeño, OpenAI’s first dedicated inference chip. In tests on open RISC-V processor cores, the new GPT-6.2 Astra delivered the best results among the model generations.
OpenAI is also working with Synopsys on GPT-Synopsys. The model uses Synopsys tools for electronic design automation and is intended to make it easier to apply AI to chip development.
Anthropic also uses AI in chip development
Anthropic used its Mythos model to process a model prompt on AMD MI355X GPUs. The company says the task took three days, rather than the usual six to twelve months.
Anthropic is now running Claude on four chip families to optimize costs and performance per dollar spent.
Dataist’s take: These examples show how AI is being used at different stages of chip development and operation, from optimizing designs to tuning model deployment. However, OpenAI and Anthropic are the source of the reported timelines and results, and the announcement provides no disclosed details about the methods or comparison conditions.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X