[Hardware] Nvidia Taps TSMC's 5nm for RTX 40-Series GPU Architectures

[Agent47]

 

 

Nvidia plans to introduce two brand-new GPU architectures, Ada Lovelace and Hopper (named for American computer scientist Grace Hopper), next year. The new graphics processors will address virtually all segments of the market, and industry sources believe they will be made using TSMC's N5 process technology, reports DigiTimes. This remains a widely reported rumor, so as usual take it with a pinch of salt.

Right now, Nvidia uses Samsung's 8LPP fabrication process to make its Ampere-based GeForce RTX 30-series GPUs, as well as TSMC's N7 node to produce its high-end GA100 datacenter GPU for artificial intelligence (AI) and high-performance computing (HPC) applications. Such a dual-source strategy allows Nvidia to diversify its supply chain, but since neither if its foundries get the maximum orders, it cannot really get the maximum allocation. As a result, next year the company will exclusively use TSMC for all of its upcoming GPUs.
Next year Nvidia plans to introduce two GPU architectures: the Hopper architecture for datacenters, AI, and HPC applications as well as the Ada Lovelace architecture for gaming GeForce GPUs.

Nvidia’s Hopper H100 GPU is expected to be the company’s first multi-die compute GPU to maximise performance per socket. The multi-die GPU is projected to be paired with HBM memory and will use TSMC’s chip-on-wafer-on-substrate (CoWoS) packaging technology.

By contrast, Nvidia’s GeForce RTX 40-series Ada Lovelace GPUs are aimed at gamers and are likely to use traditional memory and FC-BGA packaging. DigiTimes claims that ASE Technology was tapped to test and package such GPUs.

What is completely unclear is when exactly Nvidia plans to introduce its next-generation GPUs. Based on rumors, the company intends to refresh its GeForce 30-series family once again early next year with new Ti and Super parts. Since it does not make sense to refresh a lineup if you plan to launch a brand-new series shortly, we may speculate that Ada Lovelace graphics cards will only be available sometime in late 2022.

Keeping in mind that Nvidia’s A100 compute GPU now has a strong rival in the form of AMD’s Instinct MI200-series accelerators, Nvidia is certainly interested in launching its Hopper architecture sooner rather than later. To that end, we will likely see Hopper arriving ahead of Lovelace.

Nvidia did not comment on the news story, as the company traditionally does not publicly talk about future products.