Industry sources from China claim that NVIDIA will leverage Intel’s advanced packaging technologies to improve the production capacity of its AI accelerator chips. UDN reports that a monthly production capacity of 5,000 pieces has been reserved. This accounts for roughly 10% of NVIDIA’s overall capacity. It’s important to note that Intel will (allegedly) only handle the packaging and not the production of the GPUs.
A while back, Intel celebrated the opening of Fab 9, its first high-volume factory dedicated to advanced semiconductor packaging solutions, including Foveros 3D packaging. Intel’s had 3D packaging since 2019 when Lakefield was introduced in limited quantities with DRAM stacked atop the base die and the compute chiplet. Foveros is also used in the Core Ultra “Meteor Lake” processors to stack the compute and tGPU tiles on the base die.
Intel and TSMC have been racing to increase their advanced packaging capacity. For TSMC, this is going to be one of the primary sources of revenue going forth. For Intel’s foundry business, it can be a leading source of third-party customer revenue. That said, TSMC will remain NVIDIA’s primary foundry partner, fabricating 90% of its Tensor Core GPUs.
If that weren’t enough, the Taiwanese foundry plans to drastically expand its packaging production capacity in 2024. By March, its monthly packaging capacity is expected to grow to 50K pieces, up from ~40K in December last year.
NVIDIA has been using a dual-foundry approach for several years now. It tapped into Samsung’s old but cheap 8nm LPU for the production of its GeForce RTX 30 series GPUs but maintained its contract with TSMC to produce the Tensor Core AI GPUs. With the RTX 40 family, the chipmaker has completely shifted to TSMC’s N5 process for better performance but lower markups. The decision to assign part of the GPU production to Intel can be seen as a strategy to maximize production and profits, at least till TSMC can consolidate its own.