NVIDIA Partners with d-Matrix for Hybrid AI Compute Infrastructure

AI chip giant NVIDIA has partnered with ASIC startup d-Matrix to develop a new type of hybrid compute infrastructure designed to support AI models. The collaboration aims to integrate hardware from both companies to create a more efficient solution.

d-Matrix announced that its Neocloud Parasail system, which pairs NVIDIA's Hopper and Blackwell GPUs with d-Matrix's Corsair ASICs, has achieved a tenfold increase in token generation rate. This performance improvement suggests significant advancements in AI inference tasks.

The Parasail system utilizes NVIDIA GPUs for the computationally intensive pre-filling tasks at the front end of AI inference workloads. Concurrently, d-Matrix's Corsair chips, featuring a Digital-In-Memory Computing (D-IMC) architecture and ample SRAM, are well-suited for the latency-sensitive decoding tasks at the back end of inference.

Sid Sheth, founder and CEO of d-Matrix, stated that Parasail validates their initial belief in a heterogeneous future for inference. He emphasized that GPUs and dedicated inference accelerators are complementary rather than competitive, and that Parasail's approach optimizes every stage of the workload.