d-Matrix Begins Full Production of Its In-Memory AI Inference Chip

AI chip challenger d-Matrix has entered full production of Corsair, its AI inference platform — and the design takes a notably different approach to the efficiency problem squeezing the industry.

Computing inside the memory

Corsair is built on an SRAM-based in-memory compute chiplet architecture. Instead of constantly shuttling data between separate memory and processing units — a major source of energy waste and latency in conventional designs — in-memory compute performs calculations much closer to where the data lives. For inference, where the same model is run over and over for users, that efficiency translates directly into lower cost per query.

Targeting the real recurring cost

Much of the AI chip conversation has fixated on training the largest models. But the recurring expense for anyone operating AI at scale is inference. By optimizing specifically for serving workloads, d-Matrix is going after the part of the bill that never stops growing — and pitching an alternative to general-purpose accelerators.

Why it matters

Reaching full production is the moment an ambitious architecture has to prove itself in the real world. If in-memory compute delivers meaningful gains in efficiency, it validates a different path forward for AI hardware — one focused on doing more with less power rather than simply building bigger. In a year defined by ballooning AI energy costs, that pitch has never been more timely.