📅 November 18, 2026 • OpenAI
OpenAI reports that its AI models delivered up to 14% PPA improvement — power, performance and area — on AMD's flagship GPU IP relative to heavily optimized baselines. Per OpenAI, the result corresponds to roughly a year of design progress achieved in less than 5 months. The claim covers the chip-design layer — EDA and RTL — rather than GPU benchmark results. The company published the technical details on November 18, 2026, including a look inside its own Jalapeño chip program.
The method works on RTL code. The models rewrote SystemVerilog RTL, the hardware description code from which chip circuits are built, targeting calculations, control operations and data movement. Every rewrite was checked by Synopsys EDA tools to confirm that the block's behavior stayed unchanged. Because verification ran on each candidate change, engineers could accept or reject edits block by block. OpenAI says the model reasons about RTL at the level of an expert microarchitect, and the company is working with Synopsys to bring frontier models into the EDA toolset engineers already use.
Adoption at AMD moved fast. Per OpenAI, AMD has already included some of the model-proposed changes in a recent design sent for manufacturing and added others to its production design files for future versions. AMD also confirmed single-digit gains in key parts of its next flagship CPU, according to OpenAI. The same approach ran inside Jalapeño, OpenAI's first custom inference chip, where model-assisted design started in early 2025 with GPT-o3 mini.
The Jalapeño numbers are silicon-proven. Per OpenAI, model-proposed changes reduced standard-cell area by about 13% in the vector unit and about 10% in the matrix unit, and the optimizations were confirmed during the chip's bring-up. One multiplier redesign used about 15% less cell area while cutting latency from two clock cycles to one, per OpenAI, and a separate adder redesign cut cell area by about 15% while keeping two-cycle latency. The team also evaluated successive model generations on two open-source processor designs, per OpenAI.
In practice the loop compresses design iteration. AI proposes RTL rewrites, Synopsys EDA tools verify that behavior is preserved, and AMD folds the verified edits into tape-outs and production files — a cycle that OpenAI says previously took about a year and now runs in under 5 months with models operating at microarchitect level. For context, the work lands ahead of the first 1 GW deployment of AMD Instinct MI450-series GPUs for OpenAI, scheduled to begin in H2 2026 per OpenAI's announcement. The headline figures remain vendor-reported, with Synopsys tooling cited as the verification layer inside the flow.
Sources: openai.com the-decoder.com thenextweb.com magica.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)