Date: November 5, 2026
Time: 11:45 a.m. - 12:45 p.m.
Location: G01 Gates Hall
Title: Wafer Scale Computing: A New Hardware Foundation for High Performance AI
Speaker: Robert Schreiber, Cerebras Systems
Host: Giulia Guidi
Abstract: Hardware and use case drive changes in computer system design. VLSI and CMOS displaced mainframes and specialized scientific machines because of their superior cost and power efficiency, driven by the huge personal computing market. Now, AI has created an explosion of compute demand sufficient to sustain investment in and use of hardware that is better tailored to AI’s needs, which needs it shares with parts of computational science. Enter the wafer-scale computer.
A wafer-scale processor chip is made in much the same way as a CPU or GPU, printing a matrix of dies by a step and repeat photolithography processes. But rather than cutting the wafer into chips, dense wiring is laid down interconnecting the dies. The whole wafer remains intact, serving as a single “chip”, but with two orders of magnitude more transistors than a conventional CPU or GPU. Cerebras Systems was founded in 2016 in order to make wafer-scale chips, and today is the only manufacturer of commercially available wafer-scale chips and the computers that use them. AI is the principal user case.
Wafers have two big advantages over chips: performance and energy. The energy advantage stems from eliminating data movement across chip boundaries when accessing off-chip memory and when communicating with other chips. Communication bandwidth is far greater and latency is dramatically lower for on-wafer transfers. An AI datacenter is built with a fixed power capacity, hence better performance is possible only with better energy efficiency.
Driven by the new hardware capabilities and the driving use case in AI, the processor and the communication network architectures and the programming model can now change to achieve the best performance and efficiency. In the talk, I will highlight the hardware and its features, the issues around defect tolerance, power provisioning, and cooling, and the novel architecture of processors and memories as well as the design and use of the on-wafer network.
Bio: Rob Schreiber works at the frontier of research on parallel computing. At Caltech he designed and implemented the first high-fidelity supercomputer simulation of a fluid flow. At Stanford, he discovered the elimination tree, a fundamental tool in sparse matrix work. With John Gilbert and Cleve Moler he gave MATLAB sparse matrix capability. His co-invention with Van Loan, the compact WY representation, is fundamental in modern implementations of stable matrix factorizations. He was a designer of the High Performance Fortran language. At HP Labs, Rob led the development of a system for the synthesis of custom hardware accelerators and co-developed the widely used alternating least-squares algorithm for matrix completion problems and collaborative filtering. There he also helped to pioneer nanophotonic signaling in processors and networks as well as the HyperX network architecture for datacenters. He is an ACM Fellow, a SIAM Fellow, and was awarded the SIAM Career Prize in Supercomputing.