AI-Written Circuit Changes Entered an AMD Design Sent for Manufacturing
Artificial Intelligence

AI-Written Circuit Changes Entered an AMD Design Sent for Manufacturing

OpenAI says its models achieved up to 14% power, performance and area gains on selected AMD GPU blocks, with some RTL changes entering a design sent for manufacturing. Jalapeño changes were validated in silicon, but the AMD figures remain block-level company reports rather than measured whole-chip results.

NewTqnia Artificial Intelligence Desk Updated 4 min read
AI-Written Circuit Changes Entered an AMD Design Sent for Manufacturing

OpenAI says its models rewrote production-level chip circuitry and delivered power, performance and area improvements of up to 14% on heavily optimized AMD GPU blocks. AMD placed some proposed changes into a design already sent for manufacturing, moving AI-assisted hardware design beyond an open benchmark and into a real production flow.

The 30-second summary

  • What happened? AI models proposed SystemVerilog RTL rewrites, while conventional Synopsys tools checked functional equivalence and estimated power, timing and area.
  • Why does it matter? Some changes entered AMD production files, and earlier model-proposed optimizations in OpenAI's Jalapeño inference chip were validated during silicon bring-up.
  • What is the catch? The reported gains apply to selected blocks, not complete chips; AMD has not published independent block details, and manufacturing inclusion does not yet prove final silicon performance.
Key number: OpenAI reports up to a 14% PPA improvement against heavily optimized AMD GPU baselines, but does not disclose how power, performance and area contributed to that maximum.

Models changed the circuit description, not the physics checks

Modern chip teams describe digital circuits in register-transfer-level code such as SystemVerilog. Engineers and electronic design automation tools then transform that logic into gates, place and route them, verify timing and power, and eventually prepare a design for manufacturing.

In the AMD work, OpenAI's models rewrote RTL for blocks that calculate, control operations or move data. Synopsys tools checked that each candidate preserved the original behaviour and evaluated its power, performance and area, commonly grouped as PPA. This division is important: the model searched design alternatives, but deterministic engineering tools remained the gatekeepers for functional correctness and physical estimates.

OpenAI says AMD observed gains of up to 14% relative to already optimized GPU intellectual property and single-digit PPA improvements across key blocks of a forthcoming flagship CPU. Some GPU changes were included in a recent design sent for manufacturing, while others entered production design files for later versions. The companies have not identified the exact products or blocks.

Jalapeño provides the silicon evidence

The stronger physical validation comes from OpenAI's own Jalapeño inference processor. The company says model-proposed changes reduced standard-cell area by about 13% in its vector unit and 10% in its matrix unit, then survived silicon bring-up. Individual multiplier and adder redesigns used about 15% less synthesized cell area, with one multiplier also cutting latency from two clock cycles to one.

Those are block-level results rather than a claim that the entire processor became 10% or 13% smaller. Still, validation on manufactured silicon matters because synthesis estimates can miss problems that appear only after layout, fabrication and operation.

Open benchmarks show both progress and shortcuts

OpenAI also ran 9,881 scored attempts on the PicoRV32 and Ibex open-source RISC-V cores. Models optimized existing RTL or designed new cores from written requirements. GPT-6.2 Astra had the strongest average result at comparable token budgets, and specification-to-RTL tasks often scored better than editing mature designs.

The company identifies a major limitation: a model starting from requirements could omit performance counters or instructions that were present in the original core but did not help the hidden throughput workloads. The benchmark allowed designs up to 16 times the original area and estimated clock frequency before full physical synthesis. That makes it a capability test, not evidence that an automatically designed core is ready for a commercial product.

Before we overstate the result

The headline 14% is company-reported and lacks enough public detail for independent reproduction. It combines three competing design objectives, so it should not be interpreted as a 14% improvement in finished-chip speed. Tape-out also marks the start of physical validation, not its completion. The AMD result will become stronger when the manufactured device demonstrates the expected gains under disclosed workloads and power conditions.

AI-assisted electronic design is not new. EDA companies already use machine learning for placement, routing and optimization. The more specific advance here is a language model proposing low-level RTL changes on production IP while remaining inside established equivalence, synthesis and sign-off workflows.

Where the workflow could matter

The result suggests models can broaden the search around circuits that expert teams have already optimized, particularly when every candidate is checked by trusted tools. Synopsys and OpenAI are now developing GPT-Synopsys to operate EDA software directly, with engineers reviewing the resulting designs.

This creates a feedback loop between AI software and its hardware. NewTqnia recently covered a new transistor platform that implemented complete logic blocks and software that made specialized AI chips easier to use. The AMD work addresses another layer: optimizing the circuit description before fabrication. The next decisive evidence will come from manufactured AMD silicon, where block-level estimates must survive a complete physical system.

Verified topics and entities

Sources and citations4 sources

Published by

N

NewTqnia Artificial Intelligence Desk

An institutional editorial team within NewTqnia

A new version of NewTqnia is ready.