OpenAI's models did in under 5 months the kind of chip optimization work that normally takes about a year, even on AMD designs that engineers had refined for years.
- Working with AMD, OpenAI improved PPA by up to 14% on AMD's flagship GPU IP in under 5 months
- Some of the changes are already in the design AMD sent to manufacturing
- The same method was used for Jalapeño, OpenAI's first in-house inference chip
- OpenAI is building GPT-Synopsys with Synopsys for AI-assisted chip design
- Anthropic says Mythos handled in 3 days work that usually takes 6 to 12 months on AMD MI355X GPUs
- All figures come from the companies themselves
Table of Contents
OpenAI's reported results on AMD GPUs
According to OpenAI, its models working alongside AMD improved PPA (power, performance, and area) by up to 14% on AMD's flagship GPU IP. The whole process took under 5 months, a fairly short window by normal chip development standards.
What makes the figure notable is that it didn't come from a rough design. OpenAI says the gain is equivalent to roughly 1 year of additional progress on top of designs that AMD's engineers had optimized for many years. In other words, the models found more room to squeeze in areas that had already been worked over carefully. OpenAI also says AMD has put some of the changes into the design it sent to manufacturing, so the result goes beyond a lab experiment.
Applied to in-house chips and design tools
The same method also helped with Jalapeño, OpenAI's first in-house inference chip. In tests with an open-source RISC-V core, the new GPT-6.2 Astra model delivered the best results compared with earlier generations. That suggests this ability improves with each model generation rather than being a one-off stroke of luck.
OpenAI is also working with Synopsys on GPT-Synopsys, a model that uses Synopsys' EDA tools to bring AI-assisted chip design to more users. Since EDA is software that nearly every chipmaker has to use, this approach, if it succeeds, won't serve OpenAI alone.
Anthropic is using AI for hardware too
Anthropic has a similar example. It says its Mythos model handled the prompt processing work for a model running on AMD MI355X GPUs in just 3 days, a job that normally takes 6 to 12 months. A gap that large is hard to ignore, though the figure comes from Anthropic itself.
Anthropic currently runs Claude on 4 different chip families to optimize cost and performance per dollar spent. If porting a model to new hardware really shrinks from months to a few days, using several chip types at once becomes far less burdensome than before.
The most valuable part of this story isn't the 14% figure itself but that it was achieved on a GPU design AMD's engineers had optimized for years, where every percentage point is hard to come by. The "one year of progress" and "3 days instead of 6 to 12 months" claims both come from OpenAI and Anthropic themselves, so treat them as a direction worth watching rather than a verified conclusion. For readers in Vietnam, there is no clear immediate impact, but if chip design cycles really do get shorter, GPUs and AI chips could arrive faster over the long term. The sensible approach for now is to wait and see whether AMD ships a product based on the design it sent to manufacturing, with concrete specs, and judge it then.
Source: The Decoder (AI)
Reviewed by: Hero Association








