d-Matrix Plans to Connect Its Next AI Chip to NVIDIA’s Rack Architecture

d-Matrix plans to place its next inference chips inside NVIDIA’s rack-scale infrastructure, pairing a specialized accelerator with NVIDIA networking, rack designs and related hardware. The company has chosen NVIDIA NVLink Fusion for its next-generation Raptor XPUs, a move the companies say is intended to reduce the work and risk of turning custom silicon into a large deployment.

A chipmaker pursuing data-center-scale deployment must also solve for the machines around its processor: high-speed connections, rack design, power, cooling, software and supply-chain integration. NVIDIA says NVLink Fusion is built to let third-party XPU and CPU makers use those surrounding layers instead of developing a complete rack-scale platform from scratch.

The planned Raptor setup combines NVLink scale-up networking with Spectrum-X scale-out networking and NVIDIA’s MGX rack architecture. d-Matrix says the resulting systems can operate alongside NVIDIA GPU systems, including Vera Rubin NVL72, for disaggregated inference.

Demand for inference is soaring, but capital, time and energy remain finite.

Sid Sheth, cofounder and CEO of d-Matrix

The planned NVIDIA components

• NVLink and MGX form the scale-up interconnect and common rack foundation for Raptor systems.
• Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs and Spectrum-X Ethernet are also planned for integration.
• Astera Labs is working with d-Matrix on custom connectivity solutions for the system.

The arrangement is notable because NVIDIA is offering the infrastructure beneath the accelerator to another chipmaker. NVIDIA describes NVLink Fusion as support for custom third-party XPUs and CPUs within its rack-scale platform. The company lists d-Matrix among an ecosystem that includes AWS, Arm, Intel, Fujitsu, SiFive, Marvell, MediaTek, Samsung, Cadence, Synopsys, Ayar Labs and Lightmatter.

NVIDIA says sixth-generation NVLink can provide up to three times lower XPU-to-XPU latency than off-the-shelf Ethernet, 10 times higher packet rates and 3 TB/s of all-to-all bandwidth per XPU. Those are NVIDIA performance claims, not results from a deployed d-Matrix system. Systems pairing Raptor processors with NVIDIA’s rack-scale technology are expected to be available in 2027, according to reporting carried by Yahoo Finance.

The announcement does not establish how Raptor systems will perform in customer data centers or whether the promised deployment advantages will materialize. It does clarify the intended division of labor: d-Matrix supplies its inference accelerator, while NVIDIA supplies the interconnect and rack framework. For buyers, the appeal is the possibility of choosing specialized compute without adopting an entirely separate physical infrastructure.