Newshot

OpenAI releases latest chip design results: up to 14% PPA improvement on AMD flagship GPUs, up to 13% area reduction for in-house Jalapeño chip

AI summary

OpenAI published a technical report titled "Advancing chip design with OpenAI models," detailing how its AI models have been used to assist chip design, including applications to its in-house inference chip Jalapeño and to AMD's next-generation GPU and CPU design optimization, with testing on open-source RISC-V processor cores. OpenAI says models brought up to 14% PPA improvement on AMD's flagship GPU IP and reduced Jalapeño's vector compute unit standard cell area by about 13% and matrix compute unit area by about 10%, with some AMD optimizations already incorporated into designs recently sent to manufacturing.

Translation(model translation, not the source's own text)

OpenAI Releases Latest Chip Design Results: Up to 14% PPA Improvement on AMD's Flagship GPU, Up to 13% Area Reduction on In-House Chip Jalapeño

IT Home, November 19 — OpenAI released its latest technical report, "Advancing chip design with OpenAI models," on November 18 local time, disclosing the latest progress in using AI models to assist chip design.

According to the report, OpenAI has applied the related technology to its in-house inference chip Jalapeño, as well as to design optimization for AMD's next-generation GPU and CPU, and has evaluated the chip design capabilities of models across different generations through open-source RISC-V processor tests.

OpenAI said that as early as early 2025, engineers began using GPT-4o to assist in designing Jalapeño. Even the model at that time was already able to propose substantive improvements in chip performance, power, and area (PPA).

In Jalapeño's vector compute unit, the AI model reduced hardware resource usage by optimizing the floating-point adder, multiplier, conversion functions, reduction logic, and special function units.

In the matrix compute unit, the AI model successfully improved floating-point operations and low-precision dot-product circuits. Among these, one multiplier redesign reduced post-synthesis standard cell area by about 15% while shortening operation latency from two clock cycles to one clock cycle; another adder design also reduced standard cell area by about 15% while maintaining a two-clock-cycle latency.

OpenAI estimates that the modifications proposed by the model reduced the standard cell area of Jalapeño's vector compute unit by about 13% and the matrix compute unit by about 10%. These optimizations have been validated on actual silicon.

In its collaboration with AMD, OpenAI models also participated in GPU and CPU chip design optimization. OpenAI said that on AMD's flagship GPU IP, compared with a highly optimized baseline design, the models delivered up to 14% PPA improvement. The related models further optimized circuit structures by rewriting the SystemVerilog RTL code describing the chip circuits. OpenAI said this is equivalent to compressing about a year of design optimization progress into less than 5 months.

AMD has already incorporated some of the optimization results into chip designs recently sent to manufacturing, while other modifications were added to production design files for use in future products. In addition, AMD also validated single-digit percentage PPA improvements from the models in key modules of its next-generation flagship CPU.

To evaluate the ability of AI models across generations to independently complete chip design tasks, OpenAI also selected two open-source 32-bit RISC-V processor cores for testing: PicoRV32, which targets FPGAs, and Ibex, which serves as the embedded processor for the OpenTitan security chip.

In the tests, OpenAI provided the models with the original hardware description code and asked them to improve processor throughput, while checking design correctness through simulation. For designs that passed the checks, researchers also calculated instructions per second by combining instructions per clock cycle under hidden test workloads and the maximum clock frequency estimated by Yosys. If a design failed the correctness check, or if the chip area exceeded 16 times that of the original design, it was scored zero.

Test results showed that under a per-run budget of about 112,000 tokens, GPT-6.2 Astra achieved the highest average scores on both the PicoRV32 and Ibex processors. When optimizing based on the existing PicoRV32 core, GPT-6.2 Astra's score under a similar budget was about twice that of GPT-6 Astra.

OpenAI also conducted another test: instead of asking the models to optimize an existing processor, it directly provided written specifications and code templates and let the models design RISC-V processor cores from scratch, deciding on their own the dataflow and pipeline stage partitioning.

In this task, all five models participating in the test scored higher under a run budget of about 112,000 tokens than they did when optimizing existing designs. IT Home noted that the entire evaluation covered two design tasks, two processor cores, and different token budgets, comprising a total of 9,881 scored runs.

However, OpenAI also pointed out that when redesigning a processor from specifications, the models may omit features in the original chip such as performance counters and extra instructions, and these features offer limited help for the throughput objective in the tests, so this may partly explain the score improvement. If the models were required to retain these features, further limit circuit area, and check timing after physical synthesis, the test results would be closer to real-world mass-production chip design scenarios.

Regarding follow-up plans, OpenAI announced a collaboration with electronic design automation (EDA) software vendor Synopsys to train a new model called GPT-Synopsys, enabling it to use Synopsys' EDA tools like a professional chip designer to complete chip design and verification work.

OpenAI said this research demonstrates that AI models can not only explore new processor architectures, but have also been able to propose further optimization schemes based on complex mass-production designs such as Jalapeño and AMD's next-generation flagship GPU and CPU.

Related Articles

  • AMD Releases 26.17 Linux Radeon Software Upgrade: Adds Support for High-Bandwidth HDMI and More
  • ByteDance Reportedly in Talks with Microsoft for $500 Million Azure Contract, Drawing Strong Displeasure from OpenAI
  • AMD Releases 26.11.2 Adrenalin Driver: Adds ROCm 10.2 Support, Optimized for Phantom Blade Zero
  • 马斯克 "Blasts" 奥尔特曼: SpaceX Expected to Receive No Revenue from OpenAI
  • AMD Launches Instinct MI450X Graphics Accelerator: Same Core Specs as MI455X, Memory Reduced to 288GB HBM4
  • AMD Updates Unreal Engine FSR Plugin, Expands FSR Super Resolution 4 Support

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗