The BIGGEST CPU Ever! - Waferscale Explained

The video explains the concept of waferscale computing by examining Cerebras wafer scale engine 2, the largest chip currently in existence. It describes how traditional transistors shrinkage is offset by building an enormous single chip rather than many small chips, to reduce data interconnect latency and increase processing efficiency for AI and scientific workloads. The speaker notes that WSE-2 is built on a seven nanometer process and packs around 850,000 cores, spanning roughly 21.5 centimeters in length, which is more total area than multiple desktop CPUs combined. Because of its massive size and transistors, the device consumes about 15 kilowatts, implying that even a hypothetical PC integration would require substantial power infrastructure. The discussion covers the trade-offs between monolithic, wafer-scale designs and chiplet approaches, explaining that in some scenarios a single large chip can outperform a matrix of smaller chips with slower interconnects. The video also mentions memory constraints and the need for high-speed memory in such systems, comparing approaches like chiplets with large memory to a single monolithic wafer-scale design. Finally, the host connects these ideas to industry trends and echoes a continued interest in large-scale silicon for AI, while noting potential hurdles like manufacturing yield and cooling requirements. The segment wraps by highlighting the ongoing appeal of supersizing compute for data centers and AI research, and invites viewers to share ideas for future topics and subscribe for more content.

Topics · technology · computing · semiconductors · artificial_intelligence · data_centers