Particle.news
Download on the App Store

Cornell Study Finds GSI’s Memory-Centric APU Matches GPU Throughput, Slashing Power Use

An independent Cornell study presented at Micro ’25 validates GPU-class throughput at dramatically lower energy on retrieval-augmented generation workloads.

Overview

  • Benchmarking on 10–200GB RAG datasets showed GSI’s Gemini‑I APU delivering throughput comparable to Nvidia’s A6000 GPU.
  • Researchers reported more than 98% lower energy consumption versus a GPU across the tested corpora.
  • The APU cut retrieval bottlenecks seen on CPUs, reducing end-to-end processing time by up to 80%.
  • The work was published via ACM and introduced an analytical framework for general-purpose compute‑in‑memory devices.
  • GSIT shares spiked as much as nearly 200% Monday with further premarket gains Tuesday, while company claims for ~10x-faster Gemini‑II remain unverified by third-party production-scale deployments.