Nvidia taps Intel for Feynman
- Nvidia is reportedly planning to use Intel Foundry for some 2028 “Feynman” AI GPU components, while keeping major compute silicon and other packaging work with Taiwan Semiconductor Manufacturing Co. - Reports tracing back to DigiTimes say Intel’s EMIB packaging would handle some chiplet links, while Taiwan Semiconductor’s CoWoS would still package the main compute modules on Feynman. - Nvidia has already mapped Feynman to 2028 after Rubin and Rubin Ultra, and advanced packaging has become a supply chokepoint for AI chips. (nvidia.com) (cnbc.com)
Nvidia is reportedly lining up Intel Foundry and Intel’s EMIB packaging for parts of its 2028 “Feynman” AI GPU program, while keeping much of the design tied to Taiwan Semiconductor Manufacturing Co., or TSMC. (notebookcheck.net) (techpowerup.com) The report, cited by Notebookcheck and TechPowerUp from DigiTimes, says Nvidia is looking at Intel 18A and 14A process technology for selected Feynman components rather than the main compute die. (notebookcheck.net) (techpowerup.com) The same reports say Intel’s Embedded Multi-die Interconnect Bridge, or EMIB, would be used for some chiplet connections inside the package, while TSMC’s Chip-on-Wafer-on-Substrate, or CoWoS, would still be used for other Feynman modules. (notebookcheck.net) (techpowerup.com) Packaging is the step that turns separate silicon pieces and memory stacks into one usable accelerator, and it has become one of the tightest constraints in the AI server market. CNBC reported on April 8 that Nvidia has reserved a majority of TSMC’s most advanced packaging capacity. (cnbc.com) That makes the Intel angle notable even if the report covers only part of the chip. A split build would give Nvidia another path for assembly and interconnect work as demand for high-bandwidth memory systems keeps rising. (cnbc.com) (notebookcheck.net) Nvidia has publicly put Feynman on its long-range data center roadmap for 2028. At GTC 2026, the company highlighted Vera Rubin for current systems and showed Feynman as the generation after Rubin Ultra. (nvidia.com) (nvidianews.nvidia.com) The reporting does not indicate that Nvidia is moving the heart of Feynman away from TSMC. TechPowerUp said the core silicon is still expected to be built on TSMC’s A16 process, with Intel handling only selected pieces. (techpowerup.com) Intel has been trying to win more outside foundry customers for both manufacturing and advanced packaging. CNBC reported this month that Intel is pitching its packaging services to customers including Amazon, Cisco, SpaceX and Tesla. (cnbc.com) Neither Nvidia nor Intel appears to have publicly confirmed the specific Feynman manufacturing split described in the DigiTimes-based reports. For now, the clearest takeaway is that the next big AI bottleneck may be less about chip design than about who can package the chips at scale. (notebookcheck.net) (cnbc.com)
