Cerebras Techniques has right this moment confirmed that its CS-1 computing system has achieved a brand new efficiency milestone, delivering speeds far past what present CPUs and GPUs are capable of obtain.
The CS-1, which is predicated on Cerebras’ WSE processor – the world’s largest pc chip, can now declare to be the “quickest AI pc in existence.”
In accordance with Cerebras and the Division of Vitality’s Nationwide Vitality Know-how Laboratory (NETL), the CS-1 is greater than 10,000 occasions quicker than main GPU opponents and 200 occasions quicker than the Joule Supercomputer – at the moment ranked 82nd on the record of the highest 500 supercomputers on this planet.
“Cerebras is proud to increase its work with NETL and produce extraordinary outcomes on one of many foundational workloads in scientific compute,” Andrew Feldman, co-founder and CEO at Cerebras Techniques, explained.
“This work opens the door to main breakthroughs in scientific computing efficiency, because the CS-1, with its wafer-scale engine, overcomes conventional obstacles to excessive achieved efficiency, enabling real-time and different use instances precluded by the failure of robust scaling. Due to the unconventional reminiscence and communication acceleration that wafer scale integration creates, we’ve been capable of go far past what is feasible in a discrete, single chip processor, be it a CPU or a GPU.”
The necessity for pace
The workload carried out to ship the brand new CS-1 benchmarks concerned a big, sparse, structured system of linear equations, the type used to mannequin numerous real-world situations, together with fluid dynamics and power effectivity.
Regardless of its spectacular statistics, nevertheless, there stay query marks over what kind of purposes it will likely be potential to run utilizing the wafer-scale expertise that underpins the CS-1. Pc modelling at fast pace actually appears to be potential and there may be a lot pleasure in regards to the potential future intersection of supercomputing and synthetic intelligence.