Korean researchers power-shame Nvidia with new neural AI chip — claim 625 times less power draw, 41 times smaller

Key Points:

  • The new C-Transformer chip is claimed to be the world’s first ultra-low power AI accelerator chip capable of large language model (LLM) processing.
  • The C-Transformer chip uses significantly less power compared to Nvidia’s A100 Tensor Core GPU and features refined neuromorphic computing technology.
  • The architecture of the C-Transformer chip includes three main functional feature blocks focused on efficiency and energy consumption reduction.


Scientists from the Korea Advanced Institute of Science and Technology (KAIST) unveiled the groundbreaking ‘Complementary-Transformer’ AI chip at the 2024 International Solid-State Circuits Conference (ISSCC). This innovative C-Transformer chip is positioned as the world’s first ultra-low power AI accelerator chip developed for large language model (LLM) processing.


KAIST researchers boldly assert that the C-Transformer outperforms Nvidia’s A100 Tensor Core GPU by utilizing 625 times less power and boasting a size that is 41 times smaller. Fabricated by Samsung, the chip’s success is attributed to advanced neuromorphic computing technology.


Despite the lack of direct performance comparisons with existing GPUs, key specifications of the C-Transformer include fabrication on Samsung’s 28nm process, a die area of 20.25mm2, and a maximum frequency of 200 MHz with power consumption under 500mW. Notably, the C-Transformer achieves 3.41 TOPS, significantly slower than the Nvidia A100 PCIe card but with vastly reduced power consumption.


The chip’s architecture features three main functional blocks: the Homogeneous DNN-Transformer / Spiking-transformer Core (HDSC) with Hybrid Multiplication-Accumulation Unit (HMAU), Output Spike Speculation Unit (OSSU) to enhance processing efficiency, and Implicit Weight Generation Unit (IWGU) with Extended Sign Compression (ESC) for lowered energy consumption.


Notably, the C-Transformer chip offers a unique approach by incorporating neuromorphic processing, traditionally deemed unsuitable for LLMs due to accuracy concerns. The KAIST team claims to have enhanced the technology’s accuracy to align with deep neural networks (DNNs).


While empirical performance comparisons are currently lacking, the C-Transformer’s potential in mobile computing is undeniable, offering a compelling alternative in the field. The successful testing of GPT-2 on the Samsung test chip underscores the promising advancements made by the KAIST research team.


Despite the need for further performance validation in comparison to established AI accelerators, the KAIST C-Transformer chip stands out as a notable development in the AI hardware landscape, showcasing a distinct approach to low-power, high-efficiency processing for demanding language models.



Prompt Engineering Guides



©2024 The Horizon