The CS-1 supercomputer, the original world’s largest chip, has reached a new performance milestone
A high-end modern computer chip may have billions of transistors. But the number of transistors in WSE exceeds 1 trillion. In addition, the 400,000 cores loaded by WSE can transmit 9,000 trillion bytes of data per second; in contrast, the current Intel i9–9900k chip configured in personal computers has only 8 cores, which can transmit 40 billion per second. Bytes of data.
Cerebras’ WSE (Wafer Scale Engine), a single chip integrates 400,000 computing cores (PE), on-chip 18GB of SRAM, pay attention to SRAM, similar to what we usually call Cache, this on-chip SRAM is better than most chips’ off-chip storage (DDR) are big. The bandwidth is 100Pb/s (1Pb=1000TB). This is basically one unit (3 orders of magnitude) higher than the relevant parameters of existing chips. The size of the processor is 215mm x 215mm.
CS-1 is a WSE processor based on Cerebras, and it can now be called the fastest AI computer. Because it is easy to install, quick to start, and integrates with existing AI models in TensorFlow and PyTorch, it can provide value on the same day.
CS-1 and WSE overcome the traditional obstacles that cannot achieve high performance, real-time and scalability. If this high-profile super AI chip enters commercialization, machine learning may enter a new level.