GSI Technology Eyes 2027 AI Production as Gemini-II Cuts Power 98% vs. NVIDIA

GSI Technology NASDAQ: GSIT is positioning its Gemini-II artificial intelligence processor as a production-ready hardware platform while it works to expand software tools and complete proof-of-concept programs that could lead to early production by the end of 2027, according to Vice President of Sales Didier Lasserre. Speaking at a Canaccord session following the company’s June-quarter report, Lasserre said GSI has progressed from introducing Gemini-II about 18 months ago to placing the hardware into production and securing third-party validation and government-backed engagements. Gemini-II is in production now as far as the hardware is concerned. He added that Cornell University compared a GSI board with an NVIDIA GPU in a retrieval-augmented generation, or RAG, test and found that GSI used 98% less power at comparable performance. Gemini-II’s compute-in-memory architecture is intended to reduce the power consumed by the repeated movement of data between processors and external memory in conventional CPU and GPU designs. The processor loads a model or database into on-chip memory and performs calculations within the memory bit line. That approach is particularly suited to search and high-performance computing applications, according to Lasserre. He said the company’s next-generation product, Plato, will target large language models, vision-language models and other edge applications by increasing the bandwidth for data entering the chip. Plato is expected to offer what Lasserre described as near data-center performance within a power budget of 2 watts to 10 watts. He said the company expects to complete the chip’s design and tape-out in the spring. While GSI develops its AI business, its long-standing SRAM business remains an important source of support for company operations. Lasserre said GSI has shipped more than 140 million SRAM devices over roughly 30 years. The products support AI-related infrastructure indirectly through customers including Cadence, whose emulation systems are used in chip design, and KYEC, which provides burn-in services for GPUs. GSI is currently working on two announced proof-of-concept programs for Gemini-II: a Department of Defense-funded drone surveillance project and a Phase I Smart City project with a municipality in Taiwan. In the drone surveillance program, Lasserre said the customer required a time-to-first-token result below three seconds and power consumption below 50 watts. He said GSI and drone partner G2 Tech demonstrated a result of 2.5 seconds at 30 watts in a June laboratory demonstration for the Defense Department. The Taiwan Smart City project is initially designed to analyze video from 20 cameras and identify events as they occur, rather than simply recording footage for later review.