custom chips

This week, insights into datacenter technology reveal a shift in GPU dominance, with companies exploring alternatives. Cerebras, founded by Andrew Feldman, has developed chips the size of dinner plates that contain 44 gigabytes of ultra-fast SRAM, circumventing high-bandwidth memory supply constraints. In 2024, Cerebras launched an inference server that processes tokens faster than any GPU configuration, utilizing structured sparsity and partnering with AWS and AMD. The latest Cerebras chips, CS-4 systems, feature a modular architecture that doubles performance and bandwidth. Google is also moving towards custom silicon with its Tensor Processing Units (TPUs), while Marvell competes with Broadcom in providing IP for custom chips. Waymo announced a custom machine learning accelerator for its autonomous vehicles to enhance compute density and address limitations of previous FPGA use. On the software side, Microsoft is reintroducing the ability to move the taskbar in Windows 11, responding to user feedback after nearly five years. This change reflects Microsoft's focus on user experience amid competition from alternatives like macOS and Linux.