explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR — what changed on November 18
  • What is Jalapeño, and what did the models change?
  • How did OpenAI work with AMD on GPU and CPU RTL?
  • What are the open RISC-V benchmarks?
  • How does this connect to GPT-Synopsys?
  • What developers are saying
  • How this compares to earlier AI-in-EDA promises
  • How AMD fits the wider custom-silicon race
  • What this means for AI-assisted hardware teams
  • What to do this week
  • Related reading
← Back to blog

explainx / blog

OpenAI Chip Design: Jalapeño PPA Gains and 14% AMD RTL Wins With Astra

OpenAI, AMD, Chip Design, GPT-6.2 Astra, Hardware

Part of OpenAI and ChatGPT

OpenAI says GPT-6.2 Astra cut Jalapeño vector and matrix area ~13% and ~10%, and helped AMD ship up to 14% PPA gains on flagship GPU RTL in under five months.

Nov 18, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
OpenAI Chip Design: Jalapeño PPA Gains and 14% AMD RTL Wins With Astra

OpenAI published a chip-design results post on November 18, 2026 saying its models already changed silicon-bound RTL on AMD flagship graphics intellectual property and on OpenAI's own Jalapeño inference chip, with up to 14% power-performance-area (PPA) improvement on heavily optimized AMD GPU blocks in under five months. The company frames that window as roughly a year of human-paced progress, citing work verified with Synopsys tools and continuing the GPT-Synopsys partnership we covered in October.

If you build agents or hardware, the post matters for two reasons: it names which model generation helped (through GPT-6.2 Astra on public RISC-V evals), and it shows production-adjacent edits, not only lab demos. Source: OpenAI, Advancing chip design with our models.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

TL;DR — what changed on November 18

table · 3 cols
TopicWhat OpenAI claimsWhat to verify yourself
Jalapeño vector unit~13% less standard-cell area from model proposalsInternal chip; no public RTL
Jalapeño matrix unit~10% area savings; some multipliers 15% smaller with latency changesSame
AMD flagship GPU IPUp to 14% PPA vs optimized baseline; RTL rewrites in SystemVerilogAMD has not posted a separate release note at publish time
AMD next-gen CPU blocksSingle-digit % PPA gains on key blocks with model-written designsPartner verification only in OpenAI post
RISC-V evalsGPT-6.2 Astra leads on PicoRV32 and Ibex at ~112k tokens/runRe-run sim plus Yosys on your own toolchain
GPT-SynopsysStill in flight; evaluators broadeningNo new SKU named

What is Jalapeño, and what did the models change?

OpenAI first described Jalapeño in an August 2026 post on using frontier models to accelerate its first custom inference chip. The November 18 follow-up goes deeper on compute blocks:

  • In the vector unit, OpenAI says early GPT-4o-era help already found nontrivial PPA wins across floating-point adders, multipliers, conversion logic, reductions, and special-function hardware, including sharing hardware across pipeline stages.
  • In the matrix unit, it credits savings on adders, multipliers, and low-precision dot products. One multiplier redesign used about 15% less cell area in synthesis while cutting latency from two cycles to one; another adder saved about 15% area at two-cycle latency.

Across Jalapeño, OpenAI estimates ~13% standard-cell area reduction in the vector block and ~10% in the matrix block, and says those optimizations were silicon-proven during bring-up. That last phrase is the difference between a pretty synthesis report and something that survived lab validation — still OpenAI's characterization until independent teardowns appear.

For context on why inference silicon keeps showing up in OpenAI's roadmap, see our coverage of physical engineering with Astra and Codex and Samsung custom chip talks.

How did OpenAI work with AMD on GPU and CPU RTL?

OpenAI describes a direct AMD partnership on PPA, not just public benchmarks:

  1. Models rewrote SystemVerilog RTL for circuits that calculate, control, or move data.
  2. Synopsys flows checked that behavior matched before and after each rewrite.
  3. OpenAI says AMD already merged some changes into a design sent for manufacturing and added others to production design files for future versions.
  4. On next-generation flagship CPU blocks, OpenAI reports single-digit percentage PPA gains using model-written designs, again per AMD verification in the post.

The headline 14% PPA figure is against heavily optimized baselines, which matters: vendors do not ship green-field RTL to flagship GPUs. The claim is that models found additional wins on top of expert human optimization, in less than five months.

Anthropic executives have made parallel noise about using AI for chip work to cut cost per dollar of compute, a theme outlets echoed when relaying this story. OpenAI's post is the primary sourced detail for AMD specifically; treat competitor quotes as context, not confirmation of AMD's numbers.

What are the open RISC-V benchmarks?

To track model-to-model progress, OpenAI ran two 32-bit RISC-V cores:

  • PicoRV32 — compact FPGA-oriented core.
  • Ibex (within OpenTitan) — security-oriented embedded core.

Task A — optimize existing RTL: Each model received the original hardware description and was asked to raise throughput (instructions per second), balancing cycle count and frequency. Designs that failed checks or exceeded 16× the original area scored zero. Passing runs used hidden workloads, geometric-mean IPC, and Yosys max frequency estimates.

Task B — design from a specification: Models received written requirements and connection templates, choosing microarchitecture and pipeline staging under the same correctness constraints.

OpenAI reports 9,881 scored runs across tasks, cores, and token budgets. At about 112,000 tokens per run:

  • GPT-6.2 Astra had the highest mean score on both cores for optimization.
  • Design-from-spec beat optimize-existing for all five models tested on both cores — a useful warning that "improve this legacy RTL" may be harder than green-field work with trimmed features.
  • OpenAI acknowledges that allowing models to drop low-value features (like some performance counters) may inflate spec-design scores versus strict production constraints.

Builders who want a comparable harness should read GPT-6.2 Astra launch benchmarks for API access, then treat Yosys numbers as directional until you add placement, timing closure, and sign-off steps your foundry requires.

How does this connect to GPT-Synopsys?

The November post's "what's next" section points back to GPT-Synopsys: a specialized model trained to operate Synopsys EDA tools the way expert engineers do, hosted on OpenAI infrastructure with Synopsys.ai and Autopilot.

September's announcement stressed no customer tape-out data in training, encryption, and retention controls — contractual claims, not a substitute for your own legal review. Today's post adds empirical PPA outcomes on Jalapeño and AMD as evidence that the general models already move RTL, while GPT-Synopsys targets native tool use inside customer flows.

If your team runs coding agents today, the split mirrors what we keep writing about agent harnesses: the model proposes edits; the harness must enforce simulation, lint, equivalence, and human sign-off gates.

What developers are saying

The Hacker News thread on the announcement was still thin at publish time (single-digit points), so there is no substantive community benchmark thread to cite yet. Watch for hands-on replication on PicoRV32 forks and skepticism about whether 14% PPA survives full physical synthesis — a common pattern when AI RTL posts land without open replication artifacts.

How this compares to earlier AI-in-EDA promises

Vendor roadmaps have promised copilot-style EDA for years. What is different in the November 18 post is the combination of named partner silicon (Jalapeño and AMD IP already in manufacturing pipelines) and reproducible public cores (PicoRV32 and Ibex) with thousands of scored runs. That is closer to how the software world treats SWE-bench than to a single demo waveform.

Still, the gap to your tape-out is large. OpenAI's own caveats note that production checks would require tighter area budgets, mandatory feature parity, and post-synthesis timing. Those are exactly the steps human teams argue about in sign-off meetings. Models may accelerate iteration, but someone with fab accountability still owns the final #ifdef.

Chip supply chain diagram evoking AMD and OpenAI silicon collaboration on PPA optimizationChip supply chain diagram evoking AMD and OpenAI silicon collaboration on PPA optimization

OpenAI's post ties Jalapeño block savings to inference silicon already in bring-up — not a hypothetical RTL playground.

The parallel NVIDIA Nemotron for Science line of work shows the same pattern in another domain: publish eval harnesses, partner on hard problems, and keep the messiest production data inside the lab. Chip design is simply where the cost of a wrong gate is measured in mask sets instead of bad citations.

How AMD fits the wider custom-silicon race

AMD is not the only hyperscaler-adjacent vendor racing custom AI silicon. Microsoft Maia, Google TPU generations, and Amazon Trainium each depend on squeezing PPA out of blocks that already had senior architects. If model-assisted RTL becomes routine, the competitive question shifts from "can we hire enough verification engineers" to "can we audit model edits fast enough to keep annual cadence."

None of that diminishes human microarchitects. OpenAI describes the model reasoning at expert microarchitect level on rewrites — a claim about capability on selected blocks, not a job replacement memo. Teams should plan for human review of every AI diff, much like GitHub's AI findings pipeline still needs humans to bless outbound security reports.

What this means for AI-assisted hardware teams

  1. Treat RTL proposals as untrusted until equivalence proofs run. Synopsys verification in the post is the right class of guardrail; your CI should rerun it on every model output.
  2. Separate inference-chip optimization from your product RTL. Jalapeño lessons may not transfer to your power domain or safety requirements.
  3. Budget tokens like compute. OpenAI standardized on ~112k tokens per eval run; long-horizon agent loops cost real money on frontier APIs.
  4. Watch AMD and NVIDIA supply narratives. PPA wins feed the same debate as NVIDIA Rubin throughput claims and TSMC upfront cash rules — faster silicon per watt changes who can afford frontier training.

What to do this week

  • Chip teams: Ask EDA vendors what logged proof bundles they require before accepting model-generated RTL.
  • Agent builders: If you wire models to EDA CLIs, add read-only sandboxes first; see external sandboxes for coding agents for the pattern.
  • Investors and PMs: Do not extrapolate 14% GPU PPA to entire SoCs; OpenAI names IP blocks, not full tape-outs.

Serverless inference stack evoking custom accelerators like OpenAI JalapeñoServerless inference stack evoking custom accelerators like OpenAI Jalapeño

Related reading

  • GPT-Synopsys partnership explainer
  • OpenAI physical engineering with Astra and Codex
  • GPT-6.2 Astra launch and pricing
  • PCB routing agents in KiCad
  • NVIDIA Q3 outlook and China data-center halt
  • What is an agent harness?

Figures and timelines follow OpenAI's November 18, 2026 post unless another source is named inline.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 14, 2026

192GB Local AI Without a $9,500 Mac: Hardware Paths and Real Trade-offs

You need about 192GB of fast memory to run quantized DeepSeek-class models with any prompt headroom. This guide maps the non-Mac options Martin Weinberg demos on YouTube and cross-checks them against Lucebox and Framework public benchmarks.

Nov 10, 2026

OpenAI ModRetro Chromatic Codex Edition: Build Game Boy Games With Codex

OpenAI put the ModRetro Chromatic Codex Edition on Supply Co. for $200 with one flashable cartridge. The Codex desktop plugin turns prompts into GB Studio projects, browser emulators, and cartridge ROMs you can play offline on the handheld.

Nov 3, 2026

OpenAI Shows GPT-6 Astra and Codex Doing CAD, PCB Design and Simulation: BenchCAD 2.0, Mazda and Chip Ganassi

On November 3, OpenAI described how GPT-6 Astra and Codex help engineers with CAD, circuit boards and simulation. It reports a 23-fold rise in hardware-team token use, a new benchmark score and partnerships with Mazda and Chip Ganassi Racing.