A Cerebras chip outpaced the Frontier supercomputer by 179 times, transforming a year’s computation into just two days!!!
Cerebras, based in Sunnyvale, California, is revolutionizing computing by creating wafer-scale chips that integrate numerous processors on a single large wafer, enhancing computational speed and efficiency. Unlike traditional GPUs that face interconnect and memory loading challenges, Cerebras’ chips enable rapid on-chip communication, significantly outperforming traditional supercomputers in specific tasks.
In molecular dynamics, their second-generation wafer-scale engine, WSE-2, outpaced the Frontier supercomputer by 179 times, transforming a year’s computation into just two days. This advancement, achieved with national laboratories, is crucial for simulating long-term material stability in extreme conditions, aiding the development of durable materials for high-stress environments.
Cerebras partnered with Neural Magic in AI to optimize large language models (LLMs). They maintained model accuracy by implementing sparsity, reducing model parameters to zeros, and retraining while cutting inference energy costs by two-thirds. This efficiency stems from Cerebras’ high memory bandwidth, enabling rapid, unstructured sparsity handling, a capability GPUs lack. These innovations highlight the potential of wafer-scale chips in diverse computational fields.
Link to article:
https://spectrum.ieee.org/amp/cerebras-wafer-scale-engine-2668479355
Credit: IEEE Spectrum