Inference-chip company d-Matrix plans to connect its next-generation Raptor XPUs to Nvidia’s rack-scale infrastructure through NVLink Fusion. The integration covers NVLink scale-up connections, Spectrum-X networking and the MGX rack architecture, allowing d-Matrix processors to operate alongside parts of Nvidia’s broader AI platform.
For d-Matrix, the arrangement provides established rack designs, power and cooling systems, networking and a supply chain instead of requiring the company to build each layer itself. It also gives data-center operators a common physical design that can host GPUs, CPUs and specialized accelerators.
Nvidia says sixth-generation NVLink can provide 3 terabytes per second of all-to-all bandwidth per processor, three times lower processor-to-processor latency than off-the-shelf Ethernet and ten times higher packet rates. Those are platform specifications, not results from an independently tested d-Matrix deployment.
The companies say Raptor racks will also be able to work with Nvidia GPU systems for disaggregated inference, where different stages of an AI workload run on specialized hardware. d-Matrix still has to deliver and validate the integrated product; the announcement describes its planned architecture rather than a generally available system.