ManufacturingOpinion

Lisa Su’s AMD Puts OpenAI to Work on Its Next GPU

OpenAI said its models rewrote AMD’s already-optimized GPU circuits, delivering roughly a year of additional progress in under five months, with some changes included in a design sent for manufacturing.

Answer
AMD Takes OpenAI’s Chip Design Changes to Manufacturing

The brief

  1. AMD's collaboration with OpenAI has led to a reported 14% improvement in power, performance, and area (PPA) for its flagship GPU designs, achieving a year's worth of engineering progress in under five months.
  2. OpenAI's models rewrote SystemVerilog register-transfer level (RTL) code for AMD's circuits, enhancing calculations, control operations, and data movement while maintaining circuit behavior.
  3. OpenAI's custom inference chip, Jalapeño, has undergone physical-chip validation, with model-proposed redesigns resulting in significant area and latency reductions in its matrix unit.

AMD chief Lisa Su has a reason to smile. After her engineers spent years optimizing flagship graphics processing unit (GPU) circuitry, OpenAI’s models found further improvements, and some have already made it into a design sent for manufacturing.

The brief

  1. AMD's collaboration with OpenAI has led to a reported 14% improvement in power, performance, and area (PPA) for its flagship GPU designs, achieving a year's worth of engineering progress in under five months.
  2. OpenAI's models rewrote SystemVerilog register-transfer level (RTL) code for AMD's circuits, enhancing calculations, control operations, and data movement while maintaining circuit behavior.
  3. OpenAI's custom inference chip, Jalapeño, has undergone physical-chip validation, with model-proposed redesigns resulting in significant area and latency reductions in its matrix unit.

“Proud of these fantastic results - @OpenAI + @AMD engineers pushing the frontiers of chip design together!” Su wrote on X.

OpenAI reported up to 14% improvement in power, performance and area (PPA) relative to heavily optimized baselines on AMD’s flagship GPU intellectual property. The company described the work as roughly a year of engineering progress achieved in less than five months.

OpenAI President and Co-founder Greg Brockman said some results were included in a design that had recently been taped out, the point when a completed design is released for manufacturing. Other changes have been added to AMD’s production design files for future versions.

Rewriting the Circuits

The models worked on SystemVerilog register-transfer level (RTL) code, which describes how a chip’s circuits operate. They rewrote circuits responsible for calculations, control operations and data movement.

Synopsys electronic design automation (EDA) tools checked that each rewrite preserved the circuit block’s behavior and evaluated its power, performance and area trade-offs.

“We’ve observed the model reason about RTL at the level of an expert microarchitect,” OpenAI said.

The work also extended to processors. According to OpenAI, AMD verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship central processing unit (CPU), using model-written designs.

Broadcom Helps Bring OpenAI’s Own Chip to Silicon

OpenAI reported physical-chip validation for a separate project, Jalapeño, its first custom inference chip. OpenAI designed the architecture around its model workloads, with Broadcom providing silicon implementation and networking technology, and Celestica contributing board, rack and system integration expertise.

Image
OpenAI, led by Sam Altman, and Broadcom, led by Hock Tan, unveiled Jalapeño, OpenAI’s first custom chip designed for large language model inference.

In Jalapeño’s matrix unit, one model-proposed multiplier redesign used about 15% less cell area in synthesis and reduced latency from two clock cycles to one. A separate adder redesign cut cell area by about 15% while retaining two-cycle latency.

Across the matrix unit, OpenAI estimated that model-proposed changes reduced standard-cell area by about 10%. The company said the optimizations were verified during silicon bring-up, when engineers test newly manufactured chips.

The AMD results have reached the manufacturing-design stage, while the Jalapeño changes have been tested on physical silicon.

AI Designing the Chips That Run It

OpenAI CEO Sam Altman has pointed to chip design as a task suited to AI. Discussing the work needed to produce the files sent for manufacturing on the Cheeky Pint podcast, he said, “That is a great problem for AI.”

The AMD work illustrates the potential to find further improvements in circuits experienced engineers have already optimized. Smaller circuits can free up chip area, while lower latency can shorten the time needed to complete an operation. The production impact depends on which changes survive the full design and manufacturing process.

AMD is also working to reduce the resources needed to check designs. In a PrimeTime timing test, a server with 256 GB of dynamic random-access memory (DRAM) and 256 GB of flash-backed capacity retained 95% of the performance of a 512 GB all-DRAM setup. That work addressed memory requirements for verification, alongside the circuit improvements reported with OpenAI.

NVIDIA Remains Part of OpenAI’s Compute Strategy

OpenAI’s custom silicon and AMD work sit alongside its NVIDIA relationship. In a letter of intent announced last year, OpenAI and NVIDIA outlined plans to deploy at least 10 gigawatts of NVIDIA systems. They also said they would coordinate OpenAI’s models and infrastructure software with NVIDIA’s hardware and software roadmaps.

“Everything starts with compute,” Altman said when that partnership was announced. The agreement focused on supplying and coordinating infrastructure for training and running models.

The AMD results add another role for OpenAI in that supply chain. Its models are helping improve the hardware designs on which future computing capacity depends.

Bringing the Approach to Other Designers

The AMD work follows a multi-year agreement announced by OpenAI and Synopsys in September. The companies are developing GPT-Synopsys, a specialized model trained to operate Synopsys’ chip-design tools.

The planned service will bundle computing capacity, the model and software licenses. Agents will run tools, interpret results and iterate on designs for engineer review. The agreement includes joint commercialization and revenue sharing.

Synopsys said customer design data will be protected through encryption and configurable access, retention and audit controls, and excluded from model training.

“We’re using our most advanced technology to improve the systems that power AI,” Brockman said when the partnership was announced.

Last updated

Written by Amit Raja Naik

Join the discussion

0 answers

People who run these systems answer here. Your verified role is shown, never your employer unless you choose.

Sign in with LinkedIn to answer

Read next

MachineCon USA · from AIM Media House

Where enterprise AI is decided

An invitation-led day for the CIOs, CDOs and Chief AI Officers who own enterprise AI strategy.

  • New York5th edition · 583 Park Avenue
  • Dallas1st edition

Register for MachineCon USASee MachineCon USA →

  • Keynote on the main stage, MachineCon New York 2025
  • AI150 award winners on stage, MachineCon New York 2025
  • At the tables, MachineCon New York 2025
  • Networking, MachineCon New York 2025