Cerebras has introduced the successor to their record-breaking Wafer Scale Engine. The newly re-engineered Wafer Scale Engine 2 has been redesigned for TSMC’s 7 nm manufacturing course of – a extreme enchancment over the unique’s 16 nm. That Cerebras has moved on to TSMC’s 7 nm for this big, wafer-sized accelerator is telling of the arrogance and state of yields on TSMC’s 7 nm – if the method wasn’t thought of to be secure and guaranteeing extremely good yields, I doubt such an effort would have been undertaken.The up to date Wafer Scale Engine 2 now packs a staggering 850,000 cores (in comparison with the unique’s 400,000); presents 40 GB of on-chip SRAM (up from 18 GB); 20 PetaBytes/s reminiscence bandwidth (up from 9 PB/s), and 220 PB/s interconnect cloth pace (up from 100 PB/s). These enhancements and will increase have been achieved with the identical energy envelope as the unique, set at a staggering 20 kW (system) and 15 kW (chip) energy consumption. These are enabled by the a lot greater transistor density of seven nm vs 16 nm, which permits 2.6 Trillion transistors on the Wafer Scale Engine 2 in comparison with a “paltry” 1.2 Trillion on the unique, 16 nm model of it, while occupying the identical 46,225 sq. millimeters (about 21.5 x 21.5 cm). Cerebras has one-upped itself, constructing upon what already needed to be probably the most spectacular feat of classical chip engineering.
