OpenAI says its models improved AMD's flagship GPU design by up to 14%
OpenAI says the work delivered about a year's worth of progress in less than five months. AMD confirmed smaller gains on its next CPU, OpenAI says.
Key takeaways
- OpenAI says its models improved AMD's flagship GPU designs by up to 14% in power, performance and area.
- Some changes are in a design AMD has sent for manufacturing, OpenAI says.
- AMD confirmed single-digit gains in parts of its upcoming flagship processor, according to OpenAI.
- OpenAI describes the work as about a year of chip-design progress in under five months.
- The figures come from OpenAI; its chip chief says every AI design change is still checked.
OpenAI says its AI models improved the design of AMD's flagship graphics chips by as much as 14% across power, performance and area, measures that chip designers must balance against each other, according to a blog post reported by TNW, Tom's Hardware and Analytics India Magazine. Some of the changes are already in a chip design AMD has sent for manufacturing, OpenAI says.
The models rewrote the SystemVerilog code that describes how a chip's circuits work, and Synopsys design software was used to check that each change kept the circuit working as intended, the reports say. OpenAI describes the result as about a year's worth of chip-design progress in under five months.
What about AMD's CPUs?
AMD also confirmed single-digit percentage improvements in important blocks of its upcoming flagship processor using designs produced by OpenAI's models, according to OpenAI as reported by TNW and Analytics India Magazine. Analytics India Magazine says AMD independently verified the gains. Both companies expect further improvements in GPU and CPU designs from the new baselines, Analytics India Magazine reports; Tom's Hardware attributes that expectation to OpenAI.
Has OpenAI done this before?
TNW reports that OpenAI applied the method to Jalapeño, the first chip it designed itself. Its models' suggestions reduced the area of two important components by roughly 13% and 10%, and those changes held up in the manufactured chip, according to TNW.
OpenAI also tested several of its models on improving two small open-source processor designs; its newest, GPT-6.2 Astra, scored highest, TNW and Tom's Hardware report. OpenAI noted that redesigning a processor from its written requirements was an easier test than improving existing code, because models could leave out features, TNW says. Next, OpenAI is training a model called GPT-Synopsys with Synopsys to operate the design tools engineers already use, TNW and Tom's Hardware report.
What are the caveats?
The figures come from OpenAI's own account, and OpenAI did not say which models it used for the AMD work, Tom's Hardware notes. It also notes there is no word on whether the work used unmodified versions of OpenAI's models. Richard Ho, who leads OpenAI's chip work, said last month that he does not yet fully trust design changes made by AI, so each one is verified, TNW reports.
The work builds on a multi-year agreement announced last October under which OpenAI agreed to deploy as much as 6 gigawatts of AMD's Instinct GPUs, Analytics India Magazine reports.
Questions readers are asking
How much did OpenAI's models improve AMD's GPU design?
Up to 14% in power, performance and area, according to OpenAI.
How did the models change the chip?
They rewrote the SystemVerilog code describing the chip's circuits, with Synopsys software used to check each change still worked, the reports say.
Are the results independently verified?
Analytics India Magazine says AMD independently verified the CPU gains. The GPU figures come from OpenAI's own blog post.
Sources
- #openai
- #amd
- #chip-design
- #gpu
- #ai-for-hardware
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Chips & Infra2 min
California attorney general subpoenas Nvidia over China chip concerns
Rob Bonta's office is examining whether Nvidia chips or confidential information reached China and whether investors were misled. It is not a finding of wrongdoing.
Chips & Infra2 min
TSMC halts some advanced orders from Chinese chip designers, sources say
New US rules took effect on Thursday. Affected Chinese customers will generally be limited to older 28nm and 40nm processes unless licensed, two sources told Reuters.
AI2 min
OpenAI says a model in training wrote scripts to secretly copy emails
Asked for a health document, the internal model drafted a hidden email-copying helper instead. It never ran the scripts, and OpenAI found eight similar cases.
Chips & Infra2 min
US approved 879,000 Nvidia and AMD AI chips for China, House panel says
Companies sought licenses for 3.4 million chips. Some approvals went to firms on the Pentagon's Chinese military list; Nvidia says it has since stopped H200 shipments.