"To train a model, up to 16 of these chips could be ganged together over a high-speed interconnect. This was kind of important too, because unlike GPUs, which stor"
Cerebras Systems completed an IPO raising $5.55 billion and reaching a first-day valuation above $66 billion. The company was founded in 2015 and took more than a decade to reach this milestone, including two IPO attempts. Its early chips differed from typical GPU designs by using a wafer-sized Wafer-Scale Engine instead of cutting a wafer into smaller dies and stitching them together. The Wafer-Scale Engine measured 46,225 square mm and was built for AI training. It used a compute engine optimized for sparse matrix multiply-accumulate operations, leveraging neural network parameters that become zeros. This increased effective output from 2.65 to 26.5 16-bit petaFLOPS. Up to 16 chips could be connected for training via high-speed interconnects.
Read at theregister
Unable to calculate read time
Collection
[
|
...
]