Rumor mill: The attraction of AI techniques for enterprise prospects is not slowing down anytime quickly, and tech corporations are in want of a rising variety of GPUs and devoted chips to correctly course of these ML algorithms. Nvidia is sourcing its AI chips from TSMC, however the Taiwanese foundry could possibly be unable to satisfy the US firm’s manufacturing expectations.
Nvidia might select Intel Foundry Services as a brand new accomplice for GPU manufacturing within the subsequent few months. In accordance with rumors from unnamed business sources, Nvidia is seeking extra manufacturing capability of 5,000 wafers monthly. This may translate to 300,000 H100 chips if the contract is solely for these enterprise AI accelerators.
TSMC’s foundry companies appear to be struggling to maintain up with Nvidia’s enterprise progress. The GPU firm wants a major enhance in manufacturing capability of the chip-on-wafer-on-substrate (CoWoS) course of, an interposer-based packaging expertise that provides excessive integration density between completely different chip parts.
Superior packaging processes for chip manufacturing are anticipated to be in high demand within the subsequent few years, and CoWoS alone is projected to extend by 130 % by the second half of 2024. TSMC is working to boost its CoWoS capability to twenty,000 wafers monthly by the tip of 2024, and Nvidia is more likely to safe a majority of this elevated capability for its AI and GPU chips.
Intel Foundry Providers would not supply the identical CoWoS packaging answer utilized by TSMC, though the corporate has an identical interposer-based superior packaging course of referred to as Foveros. Nvidia might want to take a look at and validate Intel’s expertise by way of a brand new qualification course of, and the ultimate product is more likely to have barely completely different traits in comparison with GPUs made by TSMC.
Nvidia might additionally select to make the most of Intel’s foundry service to fabricate a choose variety of merchandise, outsourcing solely part of its present portfolio. The unconfirmed rumors recommend that Nvidia will begin utilizing Intel foundry crops within the second quarter of 2024.
Regardless of searching for Intel’s help, analysts predict that Nvidia will proceed to depend on TSMC as its foremost manufacturing accomplice for years to come back. Intel’s manufacturing capabilities for superior packaging options embrace crops situated in Oregon and New Mexico, and the corporate is working to quickly broaden its foundry enterprise to draw extra prospects.