Tech Featured
  • Home
  • Apple
  • Apps
  • Automobiles
  • Gadgets
  • Gaming
  • Reviews
  • Tech Tips
No Result
View All Result
  • Home
  • Apple
  • Apps
  • Automobiles
  • Gadgets
  • Gaming
  • Reviews
  • Tech Tips
No Result
View All Result
Tech Featured
No Result
View All Result
Home AI AI News

OpenAI’s Models Just Redesigned AMD’s Chips, and Compressed a Year of Work Into Five Months

Nov 20, 2026
in AI News
Reading Time: 4 mins read
0
OpenAI’s Models Just Redesigned AMD’s Chips, and Compressed a Year of Work Into Five Months

OpenAI’s models are now designing the chips that will train OpenAI’s next models. In a research post published November 18, the company said its AI systems rewrote circuit designs for its own Jalapeño inference chip and for AMD’s flagship GPU hardware, delivering up to 14 percent gains in power, performance and area. The claim that matters most is the pace: OpenAI says the AMD work compressed roughly a year of chip engineering progress into less than five months.

It is the clearest public example yet of AI speeding up AI research, and it arrives in the same week OpenAI reported that its coding agents now finish 46 percent of two to four hour research tasks on their own, up from 23 percent in March. The loop is starting to close: better models build better chips, and better chips train better models.

What OpenAI’s models actually changed

Chip design is a good fit for AI because so much of it is rewriting code. A chip’s circuits are described in SystemVerilog RTL, a hardware description language, and OpenAI’s models rewrote that code directly, then used Synopsys verification tools to confirm each block still behaved correctly after the change.

On Jalapeño, OpenAI’s first custom inference chip, the work started in early 2025 with GPT-o3 mini. The model proposed changes that cut standard cell area by about 13 percent in the vector unit and 10 percent in the matrix unit, and those savings were proven in real silicon during bring up. One multiplier redesign used about 15 percent less cell area and cut latency from two clock cycles to one. Compared with human baseline designs, OpenAI reported improvements of 56 percent for a BF16 multiplier, 21 percent for an FP4 dot product block, 10 percent for an FP32 accumulator, and an 8 percent area reduction for the SIMD unit.

The AMD work went further. OpenAI says its models produced up to 14 percent combined power, performance and area gains on AMD’s flagship GPU intellectual property, which the company describes as already heavily optimized. AMD has already included some of those changes in a design sent for manufacturing and added others to production design files for future versions. AMD also verified single digit percentage gains across key blocks of its next generation flagship CPU.

GPT-6.2 Astra led the benchmark that measures chip design skill

To compare model generations, OpenAI ran 9,881 scored runs across two open source 32 bit RISC-V processor cores, PicoRV32 and the Ibex core used in OpenTitan’s security chip. Each model was given about 112,000 tokens per run to improve the cores. GPT-6.2 Astra, released November 17, recorded the highest average score on both cores, roughly doubling GPT-6 Astra’s score on PicoRV32.

OpenAI was open about a limitation in that test. When models designed a core from a written specification rather than improving an existing one, they scored higher partly because they could leave out features the original had, like performance counters and extra instructions. The numbers show capability improving fast, but they are not a clean measure of production design skill.

The same honesty applies to the headline 14 percent figure. OpenAI did not break it down into separate power, performance and area deltas, and it did not name the AMD products involved. Greg Brockman, OpenAI’s president, called the results unprecedented in a post on X, but the five month acceleration figure rests on OpenAI’s own account, without a public stage by stage breakdown or independent confirmation from AMD.

GPT-Synopsys is the next step

OpenAI is now working with Synopsys to train a specialized model, GPT-Synopsys, that operates the company’s electronic design automation tools directly, the way expert chip designers do. Today’s models rewrite RTL code and hand it to verification tools. A model that drives the tools itself could run the full design loop, from specification to verified layout, with far less human involvement.

The partnership sits inside a much larger OpenAI and AMD relationship. On October 6, 2025, the two companies announced a strategic deal committing OpenAI to 6 gigawatts of AMD compute, starting with 1 gigawatt of MI450 GPUs in the second half of 2026, alongside warrants for up to 160 million AMD shares. A follow up agreement on November 16 added 457 megawatts of TensorWave capacity on AMD MI455X GPUs expected in the first half of 2027.

OpenAI’s second generation custom AI accelerator is heading toward tape out, and a third generation is already in development. The company is not just buying chips from AMD and Google, it is building its own silicon, and its models are now part of that design team. For more on the buying side of that strategy, see our report on OpenAI and Anthropic buying 26 million Google TPUs.

The safety question OpenAI is not dodging

There is a reason this story matters beyond chip benchmarks. AI improving AI is one of three frontier capability areas OpenAI tracks in its Preparedness Framework, alongside biological and chemical risks and cybersecurity, because a system that can improve itself could accelerate in ways nobody planned for.

GPT-6.2 Astra is the first model OpenAI has released externally at its High rating for AI self improvement, and the company says it has met the threshold in its Responsible Scaling Policy for AI accelerated AI research, based on expected future models rather than a named current one. An independent evaluation by METR, which spent about eleven weeks inside the company, estimated a 1.9x acceleration in research speed, attributing roughly 80 percent of it to AI.

That is the number to watch. If coding agents keep doubling research output every few months, the chip gains announced this week are a preview, not a peak. The models that designed a 14 percent better GPU block will train on that better GPU block, and the next generation will start from a higher floor.

For now, the practical takeaway is narrower. AMD gets better silicon into manufacturing sooner, OpenAI gets cheaper and faster training hardware, and the rest of the industry gets a public benchmark, 9,881 scored runs on open RISC-V cores, for measuring how good models are at the job. The next GPT-Synopsys results will show whether the loop can run without a human in the middle.

Read our launch coverage of GPT-6.2 Astra and Sol for the model behind these numbers, and Nvidia’s latest earnings for the other side of the AI hardware race.

Previous Post

iPhone Duo vs Galaxy Z Fold8: Which Foldable Wins, and What the $100 Gap Really Buys

Emily Cooper

Emily Cooper

Emily Cooper writes about artificial intelligence, gadgets, and emerging technology. She follows new AI model releases, benchmarks, and developer tools closely, and focuses on what they actually mean for readers.

For any queries, you can reach us at [email protected]

Latest Posts

  • OpenAI’s Models Just Redesigned AMD’s Chips, and Compressed a Year of Work Into Five Months
  • iPhone Duo vs Galaxy Z Fold8: Which Foldable Wins, and What the $100 Gap Really Buys
  • Gemini 4 Neon Is Google’s New Coding Model: First on the Coding Agent Index at $3 per Million Tokens
  • Anthropic’s $81.3 Billion IPO Is the Largest Ever: $83 a Share, a $1.5 Trillion Valuation, and Trading Starts Today
  • A Hacked Website Tool Put Malware on the BBC, Zoom and CNN: Update Your iPhone and Check Your Cards
  • Guest Posts
  • Advertise
  • Contact Us
  • Privacy Policy

© 2026 JNews - Premium WordPress news & magazine theme by Jegtheme.

No Result
View All Result
  • Home
  • Apple
  • Apps
  • Automobiles
  • Gadgets
  • Gaming
  • Reviews
  • Tech Tips

© 2026 JNews - Premium WordPress news & magazine theme by Jegtheme.