OpenAI says its latest AI models are becoming capable of making expert-level improvements to semiconductor designs, reporting up to 14% improvements in power, performance and area (PPA) for AMD’s flagship GPU intellectual property. The company also revealed that AI-assisted redesigns reduced the standard-cell area of key computing components in its own Jalapeño inference chip by approximately 13% and 10%, with the improvements validated in silicon.

In a technical report published on November 18, 2026, titled Advancing Chip Design with OpenAI Models, the company detailed how successive generations of its models have helped engineers rewrite semiconductor circuit designs, optimize processor architecture and improve hardware efficiency. OpenAI says some AMD optimizations have already been incorporated into a design submitted for manufacturing, while its latest model, GPT-6.2 Astra, achieved the highest average scores in evaluations involving two open-source RISC-V processor designs.

Key Takeaways

  • AMD GPU improvement: OpenAI models achieved up to 14% PPA improvement against heavily optimized baselines on AMD’s flagship GPU IP.
  • Jalapeño optimization: AI-assisted changes reduced standard-cell area by approximately 13% in the warp engine and 10% in the matrix unit.
  • Silicon validation: OpenAI says the Jalapeño improvements were validated during silicon bring-up.
  • Faster engineering: OpenAI describes the AMD optimization progress as roughly one year of work completed in less than five months.
  • Manufacturing progress: AMD has incorporated some model-generated changes into a design sent for manufacturing.
  • RISC-V results: GPT-6.2 Astra achieved the highest average scores across OpenAI’s two processor-design evaluations.
  • Synopsys partnership: OpenAI is working with Synopsys to train a model that can use professional electronic design automation tools.

OpenAI Models Deliver Up to 14% PPA Improvement for AMD GPUs

One of the most important findings in OpenAI’s report concerns its collaboration with AMD.

The companies used OpenAI models to identify improvements in AMD’s flagship GPU intellectual property. The models reportedly achieved up to 14% improvement across power, performance and area compared with heavily optimized baseline designs.

PPA is a central measurement in semiconductor engineering because it captures three competing design objectives:

  • Power: The electricity required to operate a chip or circuit.
  • Performance: How quickly the hardware can execute operations.
  • Area: The amount of physical silicon occupied by the circuit.

Improving these characteristics can help semiconductor companies build more efficient processors, reduce manufacturing costs or increase performance within existing hardware constraints.

OpenAI says its models achieved these improvements by rewriting SystemVerilog RTL, the code engineers use to describe the circuits that make up a semiconductor.

AMD Has Incorporated Some AI-Generated Changes Into Manufacturing Designs

The significance of the AMD results extends beyond laboratory experiments.

OpenAI says AMD has already incorporated some of the model-generated optimizations into a recent design submitted for manufacturing.

Other modifications have been added to AMD’s production design files for potential use in future products.

This is important because semiconductor design improvements must pass extensive verification before they can be used in commercial hardware.

An AI-generated circuit cannot simply be accepted because it looks more efficient. Engineers must establish that it preserves the original design’s behaviour and meets timing, power and physical implementation requirements.

OpenAI says it used Synopsys tools to verify the rewritten circuits and evaluate their PPA trade-offs.

The company also reported single-digit percentage PPA improvements across key blocks of AMD’s upcoming next-generation flagship CPU using model-written designs.

OpenAI did not identify the exact AMD GPU or CPU product involved in the announcement.

AI Reduces the Area of OpenAI’s Jalapeño Chip Components

OpenAI also detailed improvements made to its own custom inference processor, Jalapeño.

The company began using GPT-4o to assist with Jalapeño’s design in early 2025. Even that earlier model identified opportunities to improve the chip’s computing components.

Subsequent model generations helped engineers redesign arithmetic circuits, optimize data processing and reduce hardware resource requirements.

The latest report says model-proposed changes reduced standard-cell area by approximately 13% in the warp engine and 10% in the matrix unit.

Jalapeño componentReported standard-cell area reduction
Warp engineApproximately 13%
Matrix unitApproximately 10%

These improvements were validated during Jalapeño’s silicon bring-up, according to OpenAI.

Standard-cell area measures the physical space occupied by the basic logic elements used to implement a chip’s digital circuits.

Reducing this area can create opportunities to fit more computing capability into a fixed silicon footprint or improve manufacturing economics.

AI Redesigns Improve Arithmetic Circuits

OpenAI highlighted several examples of how its models improved Jalapeño’s internal computing logic.

In one multiplier redesign, the model proposed a circuit that reduced synthesized standard-cell area by approximately 15% while reducing latency from two clock cycles to one.

Another adder redesign also reduced standard-cell area by approximately 15% while retaining its two-cycle latency.

The improvements were achieved through alternative logic implementations, more efficient hardware sharing and changes to how computations were distributed across pipeline stages.

These are important engineering decisions because arithmetic circuits are fundamental building blocks of modern processors and AI accelerators.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.