A few weeks ago we wrote about how Eliyan’s NuLink PHY could do away with silicon interposers and integrate everything into a single elegant package. How the socket could essentially become the motherboard.
At the recent 30th annual North America Technology Symposium, the Taiwan Semiconductor Manufacturing Company (TSMC) unveiled plans to build a version of its chip-on-wafer-on-substrate (CoWoS) packaging technology that could lead to system-in- packages (SiPs) are more than twice as large as the current largest.
“With System-on-Wafer, TSMC offers a revolutionary new option to enable a large array of dies on a 300mm wafer, delivering more computing power while taking up far less data center space and increasing performance per watt by orders of magnitude of size will be increased,” the company said.
An enormous amount of power
TSMC’s first SoW offering, a logic-only wafer based on Integrated Fan-Out (InFO) technology, is already in production.
A chip-on-wafer version using CoWoS technology is expected to arrive in 2027 and will enable the “integration of SoIC, HBM and other components to create a high-performance wafer-level system with computing power comparable to a server rack in a data center, or even an entire server.”
Report on the move, Tom’s hardware expands on this saying: “One of the designs TSMC envisions is based on four stacked SoICs, combined with twelve HBM4 memory stacks and additional I/O chips. Such a giant will certainly consume a huge amount of power – we are talking about thousands of watts and will require very advanced cooling technology. TSMC also expects such solutions to use a 120x120mm substrate.”
However, TSMC’s ambitious push to make giant chips pales in comparison to Cerebras Systems’ latest Wafer Scale Engine 3 (WSE-3), which is being called the “world’s fastest AI chip.” The WSE-3 features four trillion transistors and is twice as powerful as its predecessor, the WSE-2, while power consumption and price remain the same. This new chip, made on a 5nm TSMC process, offers stunning AI peak performance of 125 petaflops – equivalent to 62 Nvidia H100 GPUs.