🔥 Trending on HN

Can AI improve chip designs? OpenAI reports gains from Jalapeño and AMD

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
PPA(P-P-A)

A combined number for power use, speed, and circuit area.

RISC-V(risk five)

An openly shared design for a computer's calculating part.

RTL

Code that describes how circuits work inside a chip.

What happened

OpenAI, the company that builds AI models, published a report on November 18, 2026. It says its models improved designs for chips that run AI. The work covers Jalapeño, OpenAI's custom inference chip, and AMD, a semiconductor company. It also describes tests on two open-source RISC-V processor designs.

The report drew 161 points and 102 comments on Hacker News. Those numbers show community attention. They do not verify the company's claims.

What changed

OpenAI says it began using GPT-4o on Jalapeño in early 2025. In its vector unit, model-suggested changes cut standard-cell area by about 13%. Changes in its matrix unit cut area by about 10%. One multiplier used about 15% less cell area. Its latency fell from two clock cycles to one. OpenAI says these changes were proven during Jalapeño's silicon bring-up.

In work with AMD, the model rewrote SystemVerilog RTL, the code that describes circuits. OpenAI reports up to a 14% PPA improvement over heavily optimized baselines on flagship GPU IP. PPA combines power, performance, and area. OpenAI calls the result roughly one year of progress in less than five months. AMD reportedly put some changes into a design sent for manufacturing. It added others to future production files. AMD also verified single-digit PPA gains in key blocks of a future flagship CPU, according to the report.

How was it tested?

OpenAI tested two 32-bit open-source RISC-V cores: PicoRV32 and Ibex. It asked models to improve throughput, meaning instructions completed per second. It ran simulations and measured performance on workloads hidden from the models. A design scored zero if it failed checks or exceeded 16 times the original area.

At comparable budgets of about 112,000 tokens per run, GPT-6.2 Astra had the highest average score on both cores. GPT-6.1 Sol also beat GPT-5.6 Sol. The full evaluation contained 9,881 scored runs. Models scored higher when designing from written requirements than when improving an existing core. OpenAI notes a reason: the new designs could omit features that mattered little to the test workloads.

Why it matters

Chip design has many trade-offs. A small circuit rewrite can change power use, speed, and area at once. If models can search through such alternatives and pass checks, engineers may examine more options faster. This is more than asking an AI to write ordinary software. It suggests models can work with circuit descriptions and design tools at an expert microarchitect's level.

The workflow matters as much as the model. OpenAI says it used Synopsys tools to verify behavior and measure trade-offs. It is partnering with Synopsys to train GPT-Synopsys to use electronic design automation tools like expert designers.

What is confirmed, and what is uncertain

The report gives concrete results for one custom chip effort, AMD collaborations, and two test cores. It does not establish that every chip will improve by the same percentages. The benchmark used selected cores, workloads, budgets, and constraints. OpenAI also says physical synthesis timing and stricter area limits would make the test more representative of production design. These limits matter because a design can look good in a simulation yet fail another requirement.

The numbers also come from OpenAI's own announcement. AMD is named as a collaborator and verifier in that account. Independent replication would provide a different kind of evidence.

What to watch next

The next questions are practical. Do the changes sent toward manufacturing produce lasting gains? Can the approach keep correctness while meeting power, performance, area, and physical-timing limits? Will Synopsys tools make this a repeatable part of chip engineering? And will outside evaluators reproduce the RISC-V results?

The Hacker News response can point to questions the community cares about. It remains separate from the engineering evidence. The important test is whether these model-written changes survive real production checks and help designers repeatedly.

💬 What the HN debate says about AI chip design

The thread splits between readers who see the announcement as pre-ANT-IPO marketing and readers who see public communication as part of competition. Technically, commenters are more optimistic about RTL optimization than about the physical design, verification, data, and customer-IP problems of a complete ASIC.

  • On motivation, some commenters suspect daily publicity and an ANT-IPO narrative, while others argue that communicating progress is itself part of competing. These are opinions about intent, not evidence of technical success.
  • The measurement language draws strong criticism. Commenters note that the mean-normalized-reward metric has no visible equation or units and that the graphs do not directly show performance, power, or area. The article's reported AMD-verified single-digit PPA improvement also leaves unclear how much came directly from the model. Those numbers are company-reported claims, not independent measurements by the commenters.
  • The optimistic case is that RTL is structured digital-circuit description, and timing or power simulators can provide a feedback loop for generating and improving candidates. This is a commenter prediction, not an independently established benchmark result.
  • The skeptical case is that ASIC design is not just writing RTL. Placement, routing, floorplanning, analog and RF effects, parasitics, interference, DRC, and manufacturing validation matter. Much high-quality data is proprietary, while public process-design kits may be old or too limited for a useful training environment.
  • Users self-report that AI has helped write TCL, automate EDA flows, and supervise long simulations. One self-report also says that final DRC fixes still require hand editing and that a single prompt is not yet producing a competitive chip.
  • Another user self-reports that verification and validation, rather than writing RTL, dominate the schedule. Even if specification-to-RTL time approached zero, the number of chips produced per year might not increase much; this is an Amdahl's-law objection, not a measured result.
  • Partnerships with AMD and other chip companies raise concerns about how customer RTL and intellectual property are isolated from retention, training, or later competitive use. Without contract and operational details, the comment thread cannot establish either misuse or safety.

initial digest at 152 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

OpenAI says its AI can help make chips smaller and faster

📰 Full story: Can AI improve chip designs? OpenAI reports gains from Jalapeño and AMD

The company tested its models on real chip work and two open designs.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
PPA(P-P-A)

A number for power use, speed, and area.

RISC-V(risk five)

An openly shared chip design.

Hacker News

A website where technology communities discuss news.

💡 The gist

  • OpenAI says its AI improved parts of its Jalapeño chip.
  • Work with AMD showed up to 14% better PPA.
  • Hacker News discussed the report, but attention is not proof.

What happened?

OpenAI, the company that makes AI models, published its report on November 18, 2026. It says models can change code that describes chip circuits. Jalapeño is OpenAI's custom chip for AI inference. AMD is a semiconductor company. OpenAI says some Jalapeño areas became about 13% and 10% smaller. One multiplier became about 15% smaller. Its delay fell from two clock cycles to one.

PPA means power, performance, and area. A good chip balances all three. A smaller circuit can use less space. A faster circuit can finish work sooner. But changing one part can hurt another part. That is why engineers run many checks.

How did it test the idea?

OpenAI also used PicoRV32 and Ibex. These are two open RISC-V core designs. The models tried to improve their throughput. Tests used hidden workloads. Designs failed if checks broke or area grew above 16 times. OpenAI says the study included 9,881 scored runs. Around 112,000 tokens per run, GPT-6.2 Astra had the best average scores for both cores.

Models did better when starting from written requirements. OpenAI says this may be because they could leave out features. Those features helped less on the test workloads.

Why does it matter?

Chip experts spend time comparing circuit choices. An AI that finds more choices could help them move faster. It still needs tools to test every change. OpenAI says it used Synopsys tools. It is also training GPT-Synopsys with Synopsys.

What remains unclear?

These results do not cover every chip. They came from selected designs and rules. OpenAI made the announcement, so its claims need independent checks. AMD reportedly used some changes in designs headed toward manufacturing. We still need to see their results in products.

Hacker News gave the post 161 points and 102 comments. That shows interest from the community. It does not show that the results are correct. Next, watch for outside tests and repeated gains in real production designs.

💬 Can AI really make better chips?

HN commenters disagree about whether the announcement is mainly advertising or normal competitive communication. They are more confident about AI helping with parts of chip design than about AI making a complete ASIC by itself.

  • The results are hard to judge. The reward score is not explained clearly, and the reported single-digit PPA improvement does not show how much the model contributed. It is a company report, not an independent user measurement.
  • Some commenters think AI can improve RTL, a written description of digital circuits, by using simulators to check timing and power. Others say ASICs also require physical placement, routing, analog work, and careful testing.
  • Users self-report that AI helps with TCL and EDA automation. They also report that final DRC fixes and verification still need people, so a single prompt does not yet produce a proven competitive chip.
  • Commenters also worry about how customer RTL and IP are protected when chip companies work with AI providers. The thread does not provide enough evidence to settle that question.

initial digest at 152 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

OpenAI asked AI to make a chip better

📰 Full story: Can AI improve chip designs? OpenAI reports gains from Jalapeño and AMD

AI helped change some tiny paths inside a computer part.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
chip

A tiny computer part that does counting.

Jalapeño(ha-la-PEH-nyo)

The name of OpenAI's chip.

Hacker News

A website where people talk about technology.

What happened?

OpenAI, an AI company, asked its AI for help. Its report says so. A chip is a tiny part that does counting. Jalapeño is OpenAI's chip for AI work. The AI changed parts of its design. It also tried AMD's design. AMD is a company that makes chips.

Some parts became about 13% smaller. Other parts became about 10% smaller. One part became about 15% smaller. It finished in two clock beats, then one. These are numbers from OpenAI's report.

Hacker News is a website where people talk about technology. The story got 161 points and 102 comments. That means many people noticed and talked about it. It does not mean the story is true.

People still need to test the chip. The same result may not happen everywhere. The AI is helping people check designs. It is not making every chip alone.

💬 Is AI a chip-making helper?

Some people think the announcement is advertising. Others think companies must explain their work when they compete.

  • The score is hard to understand. The reported single-digit PPA gain does not tell us how much the AI really did, and commenters did not independently measure it.
  • AI may help write digital circuit instructions and run design tools. A real chip also needs physical placement, wiring, and careful testing, and good examples are scarce.
  • Users self-report that AI helps with some automation, but people still do final checks and worry about protecting secret designs.

initial digest at 152 comments (revision 2). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources