AI core manufacturing can really be done. A year's work can be completed in 5 months: AMD has already sent it to the factory for tape-out.

📅 2026-11-21

Abstract:

OpenAI released its latest technical report on November 18, which disclosed a very interesting development: their large model has been deeply involved in AMD's chip design process, and has achieved a PPA improvement of up to 14% in the circuit design of this chip factory's flagship GPU.

And some of the optimizations generated by AI have been incorporated into the chip design that has recently been sent to the factory for manufacturing. According to OpenAI, this cooperation is equivalent to compressing about a year of engineering optimization work into less than 5 months.


Different from common AI software code writing, this time the optimization target is hardware description language. OpenAI's model directly rewrites part of AMD's SystemVerilog RTL code.

This is the layer that defines how the circuit completes calculation, control and data transmission. After the changes are completed, Synopsys' EDA tools are used for functional verification to ensure that the circuit behavior has not changed.

What is rare is that these optimizations are based on designs that AMD engineers have polished for many years. They are benchmarked against highly optimized benchmarks and still dig out up to 14% of PPA space.

In terms of implementation progress, AMD has integrated some changes into the chip design that has recently been sent to the factory for manufacturing, and other optimizations have also been incorporated into the mass production design files for future products, and the scope is not limited to GPUs - AMD has also used the designs produced by the model to verify single-digit percentage PPA improvements on key modules of its next-generation flagship CPU.

However, it should be emphasized that 14% refers to the income at the PPA level, which does not mean that the actual performance of the GPU is directly improved by 14%. OpenAI has not broken down the respective contribution proportions of power consumption, performance, and area. AMD has not confirmed which products will use these optimizations.

Considering that RDNA5, CDNA5 and Zen6 are all next-generation architectures that AMD is ready to launch, the outside world will inevitably associate them with them, but there is currently no official statement.

The bigger point about the whole thing is the closed loop: AI models are participating in the hardware designed to run on them. How much actual benefits these optimizations can ultimately bring to AMD's commercial GPUs and CPUs depends on subsequent product performance.

However, the circuit changes generated by AI have moved from laboratory benchmark testing to design files sent to the factory for manufacturing. This step is landmark.


Related tags

Related articles

Comments

0/500
Captcha (click to refresh)
No comments yet