Their flagship products include the Gemini-I and Gemini-II APUs. Recent independent validation by Cornell University confirmed that the Gemini-I APU matches the performance of NVIDIA GPUs in AI workloads, notably retrieval-augmented generation (RAG), while using 98% less energy. This breakthrough has positioned GSI's technology as a potential game-changer in energy-efficient AI processing, with performance comparable to GPUs but dramatically lower power consumption. The Gemini-II APU, already released, offers approximately 10 times faster throughput and lower latency than Gemini-I, with further improvements in energy efficiency. This positions GSIT's products as highly competitive especially in edge AI, data centers, defense, and other areas where power efficiency is critical.

GSI Technology achieves its dramatic power reduction through its Compute-In-Memory (CIM) architectural design used in its Gemini and Plato AI chips—fundamentally different from GPU or CPU architectures. Rather than shuttling data constantly between separate memory and processing units (as GPUs and CPUs do), GSI’s Associative Processing Unit (APU) embeds computation directly into high-density SRAM memory arrays. This design allows simultaneous data comparison and processing in-place, eliminating the energy and time associated with memory-data transfers.