GSI Technology’s Compute-in-Memory APU Matches GPU AI Performance with 98% Lower Energy Use

Sunnyvale-based GSI Technology, Inc. announced that its Gemini-I Associative Processing Unit (APU) achieved GPU-class performance for large-scale AI applications while consuming over 98% less energy, according to a press release.

The findings, published by Cornell University researchers and presented at the ACM Micro ’25 conference, showed that GSI’s compute-in-memory (CIM) architecture matched the throughput of NVIDIA’s A6000 GPU on retrieval-augmented generation (RAG) workloads. The APU also performed retrieval tasks several times faster than standard CPUs, reducing total processing time by up to 80%.

The Cornell-led team benchmarked the Gemini-I APU across datasets ranging from 10GB to 200GB, highlighting its efficiency gains due to high-density and high-bandwidth memory integration. The study also introduced an analytical framework for optimizing general-purpose compute-in-memory devices.

GSI noted that its second-generation Gemini-II silicon offers roughly 10 times faster throughput and lower latency for memory-intensive AI workloads, while a future platform named Plato is expected to deliver greater compute power for embedded edge applications.