OpenAI reported on November 18, 2026, that its models helped improve power, performance and area (PPA) on unnamed AMD flagship GPU intellectual property by up to 14% against heavily optimized baselines. The company also described work on its Jalapeño inference chip and results from RISC-V benchmarks. OpenAI’s technical report presents these as distinct design and evaluation efforts.

PPA combines power, performance and area, so the 14% figure describes a composite design result. It is a maximum for the reported GPU-IP work, not a uniform speed figure for AMD GPUs. OpenAI says some changes went into a design sent for manufacturing and others into production design files for future versions. It also reports single-digit PPA gains across key blocks of an upcoming next-generation flagship AMD CPU, without naming the processor or giving an exact percentage.

For chip-design results, keep the metric attached to the design and baseline it describes; a PPA gain and a benchmark score answer different questions.

What changed in Jalapeño?

OpenAI reports that model-proposed changes reduced standard-cell area by about 13% in Jalapeño’s vector unit and about 10% in its matrix unit, and says those changes were silicon-proven. Standard-cell area measures the area occupied by the basic logic cells in a design; it is not the same as the overall chip’s physical size.

The company also describes two individual matrix-unit redesigns. A multiplier used about 15% less cell area in synthesis and reduced latency from two clock cycles to one. A separate adder redesign used about 15% less area while retaining its two-cycle latency. Those examples describe individual circuits, not the aggregate area figures for the full vector and matrix units.

Jalapeño is OpenAI’s custom inference chip. For background on that separate project, see NeoTeo’s earlier coverage of OpenAI’s Jalapeño inference chip.

How did Synopsys tools check AMD’s design changes?

An abstract left-to-right diagram links a code-like signal to an EDA analysis stage, then to checked circuit logic.

OpenAI says Synopsys electronic design automation (EDA) tools checked whether model-generated rewrites of AMD’s SystemVerilog register-transfer-level (RTL) code preserved each block’s behavior, and assessed trade-offs in power, performance and area. RTL describes a digital circuit’s behavior and structure before it is fabricated.

That account describes a tool-assisted checking process: model-proposed code changes were evaluated for behavior and design trade-offs. OpenAI also says it is working with Synopsys on GPT-Synopsys, a model intended to use Synopsys EDA tools for chip design and verification.

What do OpenAI’s RISC-V scores measure?

OpenAI evaluated models on two open-source 32-bit RISC-V processor cores, PicoRV32 and Ibex, across existing-core optimization and specification-to-core design tasks. The evaluation totalled 9,881 scored runs.

For existing-core optimization, designs first had to pass simulation checks. A failed design, or one exceeding 16 times the original core’s area, received a zero. For passing designs, OpenAI combined the geometric mean of instructions per cycle across hidden workloads with maximum clock frequency estimated by Yosys. It normalized each core’s best observed throughput to 1.0 and averaged run scores, including zeros. The result is a normalized benchmark score for these tasks, not an absolute production-chip throughput measurement.

At comparable budgets of about 112,000 tokens per run, OpenAI reports that GPT-6.2 Astra had the highest average score on both cores. GPT-6.1 Sol also scored higher than GPT-5.6 Sol on both. On the same approximate budget, all five evaluated models scored higher on specification-to-core design than on optimizing existing cores.

OpenAI says that difference may partly reflect the freedom to omit features in the specification-based task. For example, a generated core could leave out performance counters or additional instructions that contribute little to the benchmark’s throughput objective. Retaining such features, tightening area limits and checking timing after physical synthesis would make the task more like production design, according to the company.