• d-Matrix will use Nvidia's NVLink Fusion technology to connect its next Raptor inference chips in rack-scale AI systems planned for 2027
  • The design keeps Nvidia networking and rack infrastructure around a different accelerator, giving d-Matrix an established route into larger AI systems

The fact

AI chip developer d-Matrix said on 10 September that its next Raptor processors will use Nvidia's NVLink Fusion technology as part of rack-scale systems planned for 2027. d-Matrix is developing Raptor for AI inference, the work a trained model does when it produces an answer or other output. The company expects to finish the chip's design by the end of 2026 before compatible systems follow next year.

The planned racks will combine d-Matrix processors with Nvidia's NVLink connections, Spectrum-X networking and MGX infrastructure. Astera Labs is also working with d-Matrix on high-speed connections between components. The agreement gives d-Matrix access to an established Nvidia system design rather than requiring it to build every part of the rack around its own chip. The companies have not yet published independent performance results from a finished Raptor system.

The assessment

Building a competitive AI chip is only part of the job. Customers also need servers, networking, and cooling that work with it. By using Nvidia's rack technology, d-Matrix can fit Raptor into infrastructure that is already familiar to many data-centre operators instead of asking them to adopt a completely new system around the chip.

That could make Raptor easier to test, but it does not prove the finished system will perform better. Customers will still want to know how quickly it handles real AI workloads, how much power it uses and how smoothly the hardware and software work together.

For BTW readers, the important step comes in 2027 when complete Raptor systems become available for testing. If d-Matrix can offer a different inference chip without forcing customers to rebuild the surrounding infrastructure, it may lower one of the practical barriers to trying a new AI processor.

What to watch

Watch for completion of the Raptor design, followed by the first working rack systems in 2027. Tests using real AI models should show response times, power use and performance under heavier demand, giving customers a clearer basis for comparing Raptor with existing inference systems.