[ad_1]
Cerebras Programs has unveiled its Wafer Scale Engine 3 (WSE-3), dubbed the “quickest AI chip on the earth.”
The WSE-3, which powers the Cerebras CS-3 AI supercomputer, reportedly gives twice the efficiency of its predecessor, the WSE-2, on the identical energy consumption and worth.
The chip is able to coaching AI fashions with as much as 24 trillion parameters, a big leap from earlier fashions.
CS-3 supercomputer
The WSE-3 is constructed on a 5nm TSMC course of and options 44GB on-chip SRAM. It boasts 4 trillion transistors and 900,000 AI-optimized compute cores, delivering a peak AI efficiency of 125 petaflops – that’s the theoretical equal to about 62 Nvidia H100 GPUs.
The CS-3 supercomputer, powered by the WSE-3, is designed to coach next-generation AI fashions which can be 10 instances bigger than GPT-4 and Gemini. With a reminiscence system of as much as 1.2 petabytes, it could reportedly retailer 24 trillion parameter fashions in a single logical reminiscence house, simplifying coaching workflow and boosting developer productiveness.
Cerebras says its CS-3 supercomputer is optimized for each enterprise and hyperscale wants, and it gives superior energy effectivity and software program simplicity, requiring 97% much less code than GPUs for giant language fashions (LLMs).
Cerebras CEO and co-founder, Andrew Feldman, mentioned, “WSE-3 is the quickest AI chip on the earth, purpose-built for the newest cutting-edge AI work, from combination of consultants to 24 trillion parameter fashions. We’re thrilled for carry WSE-3 and CS-3 to market to assist resolve right this moment’s largest AI challenges.”
The corporate says it already has a backlog of orders for the CS-3 throughout enterprise, authorities, and worldwide clouds. The CS-3 can even play a big position within the strategic partnership between Cerebras and G42, which has already delivered 8 exaFLOPs of AI supercomputer efficiency by way of Condor Galaxy 1 and a couple of. A 3rd set up, Condor Galaxy 3, is presently beneath building and can be constructed with 64 CS-3 techniques, producing 8 exaFLOPs of AI compute.
Extra from TechRadar Professional
[ad_2]
Source link
