OpenAI issued a technical report titled "Advancing Chip Design with OpenAI Models." It talks about how models help design chips, including the chip used to run the model.
Several specific results: First, the self-developed AI inference chip Jalapeño, the model participated in optimizing the circuit design; second, in cooperation with AMD, on the key IP module of its flagship GPU, the three comprehensive indicators of power consumption, performance, and area have improved by up to 14%, and some changes have been incorporated into the design that AMD recently sent for tape-out; third, it covers the next generation CPU and open source RISC-V processor.
14% is not a small number in the chip industry. It often takes a mature team several months to achieve single-digit PPA improvements on the same process.
What's more worth mentioning are the changes in the links. In the past, when people talked about AI helping engineers, they mostly wrote verification scripts, checked documents, and added code. In this report, the model has extended to the level of circuit and microarchitecture optimization, which is where the actual shape of the chip is determined.
The statement "AI designs its own brain" sounds like science fiction, but in practice it is a very simple cycle: better chips train better models, and better models then optimize the chips. Whoever turns this cycle around first will see his cost curve go down faster.
Of course, the report was issued by OpenAI itself, and 14% is the "highest". The specific module and the baseline for comparison cannot be verified externally. AMD is willing to put the changes into the tape-out design, which is a relatively strong endorsement.