❯ OpenAI and AMD report up to 14% improvement in flagship GPU circuitry, with some changes already taped out
A mature starting pointOpenAI says it worked with AMD for less than five months to identify improvements of up to 14% in portions of a flagship GPU design. AMD engineers had already spent years optimizing the underlying circuitry.
Where the model workedIn its technical report, OpenAI describes giving the model SystemVerilog code specifying how circuits behave. It rewrote blocks that compute, control operations or move data. Synopsys tools checked that each revision preserved the intended behavior and measured its power, timing and area trade-offs. AMD selected some changes for tape-out and added others to production design files for later versions.
Three competing objectivesChip designers balance power, performance and area, collectively known as PPA. The 14% figure applies to specific design work and cannot be extrapolated to the speed or efficiency of the whole GPU. OpenAI president Greg Brockman said some changes made it into a recent tape-out, the point at which a design is sent for fabrication.
What changed inside a multiplierOn Jalapeño, one redesigned multiplier returned its result in one clock cycle instead of two while using about 15% less standard-cell area in synthesis. A different adder kept its two-cycle latency and saved about 15% in area. Across multiple changes, the team estimated savings of roughly 13% in vector-unit standard-cell area and 10% in the matrix unit, with the optimizations subsequently checked during silicon bring-up.
A separate silicon resultOpenAI applied the same method to Jalapeño, its custom inference chip, and says it verified gains on physical silicon there. The milestone cited for AMD is tape-out. OpenAI is working with Synopsys to bring the capabilities into design and validation tools that engineers already use.
▮ SIGNALPutting model-suggested changes into a design sent for fabrication is a stronger test of engineering usefulness than a chip-design benchmark alone.