D-Matrix, a Silicon Valley-based AI chip startup, said it will begin shipping its Corsair inference product this month for AI workloads that need low latency and lower power use.

Corsair is an SRAM-centric accelerator card with 2GB of on-chip SRAM per card. It is designed for memory-bandwidth-sensitive inference stages, not as a replacement for a general-purpose graphics processing unit.

In results cited by D-Matrix, Gimlet Labs tested a coding workflow using gpt-oss-120b with a 1.6 billion-parameter draft model, an 8,000-token input, and a 1,000-token output.

Gimlet said a mixed graphics processing unit and Corsair setup delivered 2x to 10x better interactivity at the same energy efficiency.

The benchmark did not disclose the exact graphics processing unit model and was based on Gimlet Labs’ own measured and modeled analysis, not independent third-party validation.

The results suggest Corsair is meant to work alongside Nvidia systems, offloading speculative decoding and other memory-bound steps.

D-Matrix also said it has commitments from unnamed hyperscalers, neoclouds, and frontier AI labs.

Founded in 2019, D-Matrix has raised about US$500 million at a roughly US$2 billion valuation, with Microsoft’s M12 among its investors.

🔗 Source: CNBC