Slim-Llama: An Energy-Efficient LLM ASIC Processor Supporting 3-Billion Parameters at Just 4.69mW

Slim-Llama, a new energy-efficient processor developed by KAIST researchers, optimizes large language model deployment using binary/ternary quantization and sparsity-aware techniques. The chip achieves 4.92 TOPS at 1.31 TOPS/W efficiency, supporting 3 billion parameter models with minimal latency and no external memory dependence, potentially revolutionizing AI hardware for edge devices and sustainable computing.