AI inference chipmaker d-Matrix announced today it will use NVIDIA NVLink Fusion to connect its next-generation Raptor XPUs to NVIDIA's AI infrastructure platform. The move places d-Matrix alongside a growing roster of ecosystem partners — including AWS, Arm, Intel, Fujitsu, SiFive, Alchip, Astera Labs, GUC, Marvell, MediaTek, Samsung, Cadence, Synopsys, Ayar Labs and Lightmatter — that are integrating custom silicon with NVIDIA's rack-scale architecture.

By adopting NVLink Fusion, d-Matrix gains access to NVIDIA's sixth-generation NVLink interconnect delivering 3 TB/s per XPU of all-to-all bandwidth, the MGX modular rack reference architecture, and the full AI factory software stack. The company said this provides a faster, lower-risk path from custom silicon to large-scale deployment without building rack-scale infrastructure from scratch. NVIDIA claims the integration offers 3x lower XPU-to-XPU latency than off-the-shelf Ethernet and 10x higher packet rates.

Confirmed

  • d-Matrix will connect Raptor XPUs via NVLink Fusion to NVIDIA NVLink scale-up networking and Spectrum-X scale-out Ethernet.
  • The integration includes NVIDIA Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs, and the MGX rack architecture.
  • NVLink Fusion supports all major CPU architectures — Arm, x86, and RISC-V — alongside NVIDIA GPUs.
  • NVIDIA claims 3x lower XPU-to-XPU latency than off-the-shelf Ethernet, 10x higher packet rates, and 3 TB/s per XPU all-to-all bandwidth via sixth-generation NVLink.
  • Architectural diagrams show the d-Matrix Raptor Compute Tray, NVIDIA NVLink Switch Tray, and d-Matrix MGX Rack as distinct hardware components integrated through NVLink Fusion.

Unknown

  • No timeline for Raptor XPU sampling, production, or general availability was disclosed.
  • Pricing, power envelope, and thermal specifications for the Raptor XPU remain unpublished.
  • Independent benchmarks comparing Raptor inference performance per watt or cost per token against NVIDIA GPUs or other custom XPUs are not available.
  • The scope of d-Matrix's supply-chain commitments and manufacturing partners for volume production has not been detailed.
  • Customer design wins or pilot deployments for Raptor-based systems have not been announced.

Our take

NVLink Fusion is becoming the default on-ramp for custom XPU vendors who want rack-scale credibility without reinventing the rack. For d-Matrix, the partnership converts a silicon innovation into a deployable product line — but the commercial test comes when hyperscalers evaluate Raptor's inference efficiency against incumbent GPUs and emerging alternatives like Groq LPX, both of which already run on NVIDIA's platform.

Series: 1. Nvidia Commits Up to $3 Billion to Lancium, Powering the Stargate AI Campus in Texas · 2. d-Matrix Adopts NVIDIA NVLink Fusion for Rack-Scale XPU Deployment · NVIDIA AI Factory

Sources