The C-Transformer chip is fabricated on Samsung's 28nm process and has a die area of 20.25mm2. It operates at a maximum frequency of 200 MHz, consuming under 500mW and can achieve 3.41 TOPS. The chip's architecture features three main functional blocks: a Homogeneous DNN-Transformer / Spiking-transformer Core, an Output Spike Speculation Unit, and an Implicit Weight Generation Unit. Despite uncertainties about its performance, the chip is seen as a promising option for mobile computing.
Key takeaways:
- The Korea Advanced Institute of Science and Technology (KAIST) has unveiled their 'Complementary-Transformer' AI chip, claimed to be the world's first ultra-low power AI accelerator chip capable of large language model (LLM) processing.
- The C-Transformer chip is said to use 625 times less power and is 41x smaller than Nvidia's A100 Tensor Core GPU.
- The chip's architecture is characterized by three main functional feature blocks: a Homogeneous DNN-Transformer / Spiking-transformer Core, an Output Spike Speculation Unit, and an Implicit Weight Generation Unit with Extended Sign Compression.
- Despite uncertainties about the performance of the C-Transformer chip due to lack of direct comparisons with industry-standard AI accelerators, it is believed to be an attractive option for mobile computing.