d-Matrix adopts Nvidia’s NVLink Fusion, plans 144-accelerator racks for AI inference
d-Matrix announced it will license Nvidia’s NVLink Fusion interconnect and use Nvidia’s MGX rack designs, targeting racks with up to 144 Raptor accelerators by next year.
AI-infrastructure company d-Matrix has joined a growing list of chip designers by licensing Nvidia’s NVLink Fusion interconnect and adopting Nvidia’s MGX rack-scale reference designs for its next generation of accelerators. Under the agreement, d-Matrix will incorporate NVLink Fusion into its forthcoming Raptor accelerators, which were revealed at the Hot Chips conference with 32 GB of 3D-stacked DRAM and a peak memory bandwidth of about 100 TB/s—roughly 4.5 times that of Nvidia’s Rubin GPU.
The firm plans to deliver rack systems that link up to 144 Raptor cards via a single all-to-all NVLink fabric by the end of next year, providing around 2.3 TB of memory capacity and 7.2 PB/s of aggregate bandwidth, sufficient for AI models exceeding four trillion parameters at 4-bit precision. d-Matrix’s smaller XPUs, about half the size of the Raptor cards, will still offer 16 GB of DRAM and 50 TB/s bandwidth, allowing flexible deployment either as standalone units or in heterogeneous configurations with GPUs for prompt processing. By aligning with Nvidia’s broader ecosystem—including Vera CPUs, NVSwitch, BlueField, ConnectX and SpectrumX—d-Matrix joins other firms such as Qualcomm, Arm, Marvell, MediaTek, Fujitsu and Amazon Web Services in Nvidia’s expanding licensing strategy.
Why it matters
The partnership could reshape AI hardware markets by giving customers high-bandwidth, low-latency inference systems without building their own interconnects.
In this story
