Nvidia works on Feynman GPU
Cette publication existe aussi en Français
The next generation of Nvidia GPU is to be called Feynman, slated to ship in 2028. This will follow the Rubin GPU that is set to ship next year after Blackwell, launched last year and now in full production.
The Vera Rubin combination will pair the Vera CPU with the Rubin GPU and is named after the physicist who discovered dark matter. The Vera CPU will have 88 custom dual thread ARM cores, expected to be based on the Neoverse V3 core. This will provide twice the performance of the current Grace CPU that has 72 Neoverse V2 cores that is used in the Grace Hopper (GH200) and Grace Blackwell (GB200 and GB300 chips) while still having a power envelope of 50W.
The Grace superchip with two processor running at 3.1GHz has a performance of 7.1TFLOPS in a power envelope of 500W with 960Mbytes of memory.
Architectural work on the Feynman GPU will be happening now, supported by AI-assisted tools from Synopsys and Siemens EDA.
“In the second half of 2025 we will easily transition to Blackwell Ultra, with 1.5x the inference and 1.5x the memory and 2x more networking bandwidth,” said Jensen Huang, CEO of Nvidia.
“But this transition is going to take years of planning, this isn’t like buying a laptop, which is why I am telling you now. We have to plan with the land and the power for data centres with engineering teams two to three years out which is why I show the roadmap.”
“One year out is named after Vera Rubin who discovered dark matter,” said Huang, “The CPU is brand new with twice the performance of Grace at just 50W with HBM4 memory.”
The Rubin Ultra GPU is planned for the second half of 2027 with NVLink interconnect to 576 GPUS in a 600kW rack. This will have 15exaFLOPS of dense Int4 inference with 14x the performance of the GB300 chips today.
The Feynman GPU, named after Richard Feynman, who won the Nobel prize in 1965 for quantum electrodynamics.
The Feynman GPU will also be paired with the Vera CPU for systems with the 8th generation NVLink interconnect at 7.2Tbit/s. This is expected to use the silicon photonics interconnect that Nvidia has been working on with TSMC.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
