GSI Technology Gemini-II Processor Benchmarks

GSI Technology announced preliminary benchmark results for its Gemini-II Compute-in-Memory processor, showcasing a 3-second time-to-first-token (TTFT) performance for multimodal large language models at the edge, consuming approximately 30 watts. This performance is competitive with or superior to existing solutions from Qualcomm and Nvidia, which either had slower TTFT or significantly higher power consumption. Lee-Lean Shu, CEO of GSI Technology, highlighted the importance of these results for 'physical AI' markets such as drones and smart cities, where low latency and power efficiency are crucial. The company's proprietary compute-in-memory architecture aims to reduce data movement, a key factor in power consumption and latency. GSI Technology is also collaborating with partners like G2 Tech on system integration.