Skip to main content

AI Completes One Year of Chip Design Work in Five Months: AMD Has Already Sent Designs for Fabrication

On November 18, OpenAI released its latest technical report, revealing an interesting development: their large language models have deeply participated in AMD's chip design process, achieving up to a 14% PPA improvement in the circuit design of the chipmaker's flagship GPU.

Moreover, some AI-generated optimizations have already been incorporated into chip designs recently sent for manufacturing. According to OpenAI, this collaboration compressed approximately one year of engineering optimization work into less than five months.

AI造芯一年的活5个月干完:AMD已经拿去送厂流片

Unlike common AI-assisted software coding, this optimization targeted hardware description languages. OpenAI's model directly rewrote parts of AMD's SystemVerilog RTL code.

This refers to the layer that defines how circuits perform computation, control, and data transmission. After modifications, Synopsys EDA tools were used for functional verification to ensure circuit behavior remained unchanged.

Notably, these optimizations were extracted from designs that AMD engineers had already refined over many years. Even when benchmarked against highly optimized baselines, they still uncovered up to 14% PPA headroom.

In terms of implementation, AMD has already integrated some changes into chip designs recently sent for manufacturing. The remaining optimizations have been included in production design files for future products, and the scope extends beyond GPUs—AMD also used AI-generated designs to verify single-digit percentage PPA improvements on key modules of its next-generation flagship CPU.

AI造芯一年的活5个月干完:AMD已经拿去送厂流片

It is important to emphasize that the 14% figure refers to PPA gains, not a direct 14% increase in actual GPU performance. OpenAI did not break down the respective contributions to power, performance, and area, nor has AMD confirmed which specific products will utilize these optimizations.

Given that RDNA5, CDNA5, and Zen6 are AMD's upcoming next-generation architectures, it's natural for outsiders to make connections, but there is currently no official statement.

The bigger highlight of this entire affair lies in the closed loop: AI models are now participating in designing the very hardware that runs them. How much actual benefit these optimizations will bring to AMD's commercial GPUs and CPUs remains to be seen in subsequent product performance.

However, the fact that AI-generated circuit modifications have moved from laboratory benchmarks to design files sent for manufacturing is a milestone with significant symbolic meaning.