d-Matrix
d-Matrix builds memory-centric AI inference accelerators and software for data centers, targeting low-latency, energy-efficient generative AI inference at rack scale.
Founded
2019d-Matrix builds memory-centric AI inference accelerators and software for data centers, targeting low-latency, energy-efficient generative AI inference at rack scale.
d-Matrix differentiates itself through a memory-centric compute architecture that moves inference directly into memory, reducing data movement and energy consumption compared with conventional GPU-based systems. Its Corsair platform is designed specifically for low-latency, high-throughput generative AI inference rather than general-purpose training workloads.
The company emphasizes rack-scale deployment built on open standards such as PCIe and Ethernet, heterogeneous pipelines that pair Corsair with GPUs for optimal prefill and decode phases, and a full-stack approach spanning silicon, networking, and software. Partnerships with NVIDIA, Arista, Broadcom, and Supermicro expand its ecosystem and deployment flexibility.
Source: d-matrix.ai
d-Matrix competes in a crowded AI accelerator market dominated by NVIDIA's entrenched GPU ecosystem, CUDA software stack, and established customer relationships. The company must prove sustained performance advantages for real-world workloads and overcome inertia among customers already committed to GPU infrastructure.
As a privately held hardware company, d-Matrix also faces capital intensity, supply-chain and manufacturing risk, and the challenge of scaling production to meet hyperscaler demand. Its success depends on broad adoption of heterogeneous inference architectures and continued ecosystem support from partners such as NVIDIA.
Source: d-matrix.ai
d-Matrix's flagship offering is the Corsair AI inference accelerator platform, built around digital in-memory compute to reduce latency and power consumption for generative AI inference in data centers. Corsair is delivered as PCIe cards, single servers, and rack-scale SquadRack reference configurations, scaling from fixed-memory workloads up to 100B-parameter models in a single rack.
The platform is complemented by JetStream networking accelerators for accelerator-to-accelerator communication and the Aviator software stack for inference optimization. d-Matrix also partners with NVIDIA to integrate next-generation Raptor XPUs into the MGX rackscale architecture using NVLink Fusion.
Source: d-matrix.ai