AI inference chip startup d-Matrixis the next generation Raptor NVIDIA's NVLink FusionThey announced that they would implement this. It's a partnership that allows their own chips to be connected directly to NVIDIA data center racks for joint use.
The final design of Raptor is scheduled for late 2026, with compatible server systems slated for release in 2027. Since the announced schedule is still in the planning stage, the timing of actual product validation and availability will need to be confirmed.
| What Is NVLink Fusion?
NVLink Fusion is It's a technology that enables high-speed connections between custom CPUs or accelerators and NVIDIA's rack systems. It provides connection standards and related technologies that enable the rapid exchange of large amounts of data between chips.
Through this, d-Matrix aims to reduce the burden of setting up new independent servers and integrate into existing NVIDIA infrastructure.
| Raptor focuses on inference
If training is the process of building a model, then inference is the process of running the completed model in an actual service. d-Matrix is designed for tasks that require quick responses, such as chatbots, coding assistants, and voice agents.
As service usage increases, As low latency and power efficiency become increasingly important, opportunities for dedicated inference chips are also growing.
| It's both competition and cooperation.
Although d-Matrix manufactures NVIDIA GPUs and other inference accelerators, it has adopted NVIDIA's connectivity ecosystem. The strategy is to integrate chips into NVIDIA systems that handle specific tasks, rather than directly replacing the entire platform.
From NVIDIA's perspective, it can also expand its influence within the ecosystem by integrating a variety of custom chips into its own racks.
| Astera Labs also supports connectivity
d-Matrix is also developing customized connectivity solutions with Astera Labs to enable fast data flow within the system. That's because even if the accelerator is high-performance, slow data transfer will still slow down the overall response time.
This is an example of how the competition in AI servers is expanding beyond individual chips to connectivity and memory design.
| Areas to be verified by 2027
Raptor is scheduled to complete its final design by the end of 2026, and the compatible rack is targeted for 2027. After that, We need to verify manufacturing yield, software compatibility, actual service latency, and power efficiency.
The financial terms of the partnership have not been disclosed.
| Platforms Remain Important Even in the Era of Custom Chips
This announcement indicates that the AI semiconductor market is The fact that it's not simply a matter of NVIDIA chips versus alternative chipsShow me. We may see more configurations where multiple accelerators share roles within a single rack.
d-Matrix Does it demonstrate its cost and speed advantages in real-world customer environments?, The next key point to watch is how widely NVLink Fusion-based systems are adopted.
References
Tags
#dMatrix #NVIDIA #NVLinkFusion #Raptor #AI Inference #AI Chip #Data Center #Semiconductors #Server Infrastructure #AsteraLabs #Semiconductor Infrastructure
