As compute demands for AI and HPC applications continue to skyrocket, chipmakers are seeking new ways to boost performance and efficiency. One innovative approach is to move beyond traditional chip packaging to wafer-scale integration. TSMC has been pioneering this with its System-on-Wafer (SoW) platform, and now the company is taking it to the next level - literally - by going 3D.
TSMC's InFO-SoW technology, introduced in 2020, allows entire silicon wafers to be used as the base "package" for massively parallel processor designs. Instead of dicing up a wafer into separate chips, InFO-SoW leaves the processor cores interconnected across the wafer for incredible throughput and bandwidth density.
Early adopters like Cerebras and Tesla have leveraged InFO-SoW for state-of-the-art AI processors and HPC accelerators. Tesla's Dojo system, for example, uses wafer-scale chips built on InFO-SoW that provide huge performance-per-watt gains versus traditional packaging approaches.
But as impressive as these monolithic processor wafers are, they have limitations. All logic has to use the same silicon process technology, and they rely solely on on-chip memory capacity. For future ultra-high performance AI workloads, even more integration will be required.
Enter TSMC's Chip-on-Wafer System-on-Wafer (CoW-SoW) platform. This revolutionary advance combines the company's InFO-SoW and System on Integrated Chips (SoIC) technologies to enable vertical stacking of additional logic or memory chips on top of a wafer-scale processor base.
With CoW-SoW, a wafer-scale InFO-SoW processor could be fabricated using an optimized logic process, while ultra high-bandwidth memory like HBM4 gets integrated via separate memory chips (likely processed with a specialty memory-optimized node). The combined SoW base and stacked chips get seamlessly interconnected using TSMC's packaging prowess.
Kevin Zhang, TSMC's VP of Business Development, explained the vision: "SoW is no longer fiction...By leveraging our advanced wafer level integration technology, we can provide customers an important path to continue growing computation capability and energy efficiency for their AI clusters and supercomputers."
The benefits could be game-changing. Architects will have a heterogenous foundation for building massively-parallel compute engines, with logic optimized for performance while tightly-coupled high-bandwidth memory ressources get stacked on top. This allows breaking the traditional sacrifices between logic and memory process optimization.
TSMC is already working on designs for CoW-SoW, with production targeted for 2027. If successful, we could see a new era of exascale-class AI and HPC processors enabled by advanced 3D system-on-wafer integration from the packaging masters at TSMC.
Comentarios