NVIDIA D-Matrix Adopts NVIDIA NVLink Fusion to Power Rack-Scale XPU Deployments
NVIDIA AI inference chipmaker d-Matrix is taking a major step toward large-scale deployment of its next-generation Raptor XPUs by adopting NVIDIA NVLink Fusion. The technology will connect d-Matrix’s custom processors to NVIDIA’s broader AI infrastructure platform, giving the company a faster path to building and deploying rack-scale inference systems.
The move places d-Matrix among a growing group of silicon innovators adopting It’s rack-scale architecture to accelerate the development of specialized AI infrastructure. By combining its Raptor XPUs with NVIDIA NVLink, MGX rack architecture and advanced networking technologies, d-Matrix aims to reduce the complexity, cost and time involved in bringing custom AI silicon into production.
“With NVLink Fusion and MGX, we can integrate our Raptor XPUs into a broadly deployed, liquid-cooled architecture, giving customers a faster, lower-risk path to deploy and scale ultralow-latency inference,” said Sid Sheth, cofounder and CEO of d-Matrix.
Meeting the Growing Demand for AI Inference
AI inference is rapidly becoming one of the most important workloads in modern data centers. As AI models become larger and more capable, businesses need increasingly powerful infrastructure to process queries, generate responses and run AI applications at scale.
However, expanding inference infrastructure presents significant challenges. Data centers must balance computing performance with power consumption, cooling requirements, networking capacity and overall cost. At the same time, AI companies are looking for systems capable of delivering extremely low latency and high throughput.
d-Matrix has developed its Raptor XPUs specifically for AI inference workloads. The company’s adoption of NVLink Fusion is designed to help move those processors beyond the chip level and into complete, scalable AI factory infrastructure.
Rather than building an entirely separate infrastructure stack around its custom silicon, d-Matrix can leverage technologies and architectures already developed and validated by NVIDIA.
From Custom Silicon to Rack-Scale Systems
Developing an XPU is only one part of building an AI computing platform. Turning that processor into a production-ready, rack-scale system requires much more than silicon.
Companies must integrate high-speed interconnects, networking, power delivery and cooling while also developing a rack architecture capable of supporting large numbers of processors. Supply-chain management, system validation and certification add further complexity.
These requirements can significantly increase development time and investment.
NVLink Fusion is designed to address this challenge by allowing custom CPUs and XPUs to connect directly to NVIDIA’s infrastructure ecosystem. Instead of designing every component of an AI factory from the ground up, silicon innovators can use NVIDIA’s established networking, rack architecture and system technologies.
For d-Matrix, this means Raptor XPUs can take advantage of the NVIDIA MGX ecosystem, including validated rack designs, power and cooling infrastructure, and an established supply chain.
NVIDIA MGX Provides a Common Foundation
A key advantage of the approach is the use of a common rack architecture. Data centers can deploy infrastructure designed around NVIDIA MGX and potentially support different types of processors, including GPUs, CPUs and specialized XPUs.
This approach can help organizations avoid creating completely separate rack architectures for every processor platform.
For d-Matrix, NVLink Fusion provides a bridge between its custom inference hardware and It’s established AI infrastructure. The result is intended to give customers more flexibility while reducing the engineering effort required to deploy specialized inference systems.
The ability to integrate different compute architectures into a common infrastructure platform could become increasingly important as AI workloads diversify.
Connecting Raptor XPUs With NVIDIA Infrastructure
d-Matrix plans to use It NVLink to connect its Raptor XPUs within a high-bandwidth, low-latency scale-up domain. This architecture is designed to allow multiple processors to communicate efficiently while working together on demanding inference workloads.
The company also expects its systems to operate alongside It GPU-based platforms, including NVIDIA Vera Rubin NVL72 systems, enabling disaggregated inference environments.
This NVIDIA could allow data centers to combine specialized inference hardware with general-purpose AI compute, depending on the needs of individual workloads.
The flexibility is particularly important as AI infrastructure evolves beyond a one-size-fits-all model. Different AI models and applications can have very different requirements for latency, memory, throughput and energy efficiency.
Expanding the AI Factory With NVIDIA Technologies
Beyond NVLink and MGX, d-Matrix plans to incorporate several additional NVIDIA technologies into its infrastructure strategy.
These NVIDIA include NVIDIA Vera CPUs, NVIDIA ConnectX-9 SuperNICs, NVIDIA BlueField-4 DPUs and NVIDIA Spectrum-X Ethernet networking.
Together, these components are designed to create a complete infrastructure foundation around Raptor XPUs. High-speed networking can support communication between systems, while DPUs can help accelerate data-center infrastructure functions and networking workloads.
The integration of these technologies also gives d-Matrix access to a broader It ecosystem rather than relying exclusively on its own infrastructure designs.
For customers, the goal is a unified AI factory where specialized inference systems can coexist with NVIDIA-powered computing platforms.
NVLink Fusion Extends NVIDIA’s Open AI Platform
It describes its AI platform as both vertically integrated and horizontally open. While NVIDIA provides the underlying infrastructure, networking and software stack, NVLink Fusion expands that ecosystem to include custom processors from other silicon companies.
This approach allows companies such as d-Matrix to concentrate on their processor architecture and workload-specific innovations while relying on It technologies for the infrastructure surrounding those processors.
The NVIDIA strategy could be particularly valuable for specialized AI accelerators designed for specific workloads.
Instead of competing with the entire infrastructure stack, silicon innovators can focus on improving performance, efficiency and workload optimization while using established systems for deployment.
A More Flexible Future for AI Infrastructure
It’s broader AI factory platform includes systems such as NVIDIA Vera Rubin NVL72, Groq 3 LPX, the Vera CPU rack, Vera BlueField-4 STX storage and Spectrum-6 SPX Ethernet networking.
The goal is to create a flexible infrastructure environment capable of supporting a wide range of AI workloads, models and architectures.
NVLink Fusion adds another layer of flexibility by allowing customers to integrate specialized XPUs and CPUs into this broader ecosystem. Instead of forcing every workload onto a single type of processor, data centers can select the compute architecture best suited to each application.
For AI inference, that flexibility could become increasingly important as organizations seek to improve performance per watt while lowering the cost of processing each AI-generated token.
Accelerating the Path to Commercial Deployment
For d-Matrix, the partnership represents more than an interconnect technology upgrade. It provides a pathway for moving Raptor XPUs from custom silicon into large-scale AI infrastructure.
By using It’s networking, rack systems, cooling technologies, software ecosystem and supply-chain infrastructure, d-Matrix can potentially shorten development cycles and reduce the risks associated with deploying a new processor architecture at scale.
As demand for AI inference continues to grow, the ability to quickly deploy efficient, specialized computing systems will become increasingly important.
NVLink Fusion could therefore play a significant role in enabling a new generation of AI infrastructure in which GPUs, CPUs and specialized XPUs work within the same rack-scale environment.
For d-Matrix, connecting Raptor to It’s AI platform marks an important step toward that future. The company is combining its specialized inference technology with It’s established infrastructure ecosystem to create a more scalable and flexible approach to AI deployment—one designed to help customers move from custom silicon to production-ready AI factories more quickly.
Source Link: https://blogs.nvidia.com/