Nvidia's NVLink Fusion: A Game-Changer for Chip Architecture

At this year's Computex trade show in Taiwan, Nvidia has announced a significant development regarding its NVLink interface, which has traditionally been exclusive to its own computing accelerators. The NVLink Fusion initiative will now allow other manufacturers to integrate this high-speed communication protocol into their chips, marking a pivotal shift in the chip industry.

Jensen Huang, the CEO of Nvidia, emphasized the ongoing focus on artificial intelligence (AI) during his keynote address, highlighting the versatility and efficiency of the NVLink technology. This interface is crucial for linking Nvidia's computing accelerators and for facilitating rapid data exchange between accelerators and system processors. For example, NVLink connects Nvidia's Grace CPU with 72 ARM cores and the latest B300 AI accelerator, achieving an impressive cumulative data transfer rate of 1.8 TB/s thanks to 18 connections each capable of 100 GB/s.

The introduction of NVLink Fusion aims to broaden the ecosystem by allowing tailored server racks that combine Nvidia accelerators with custom processors from other companies. This flexibility also extends to Nvidia's ARM-based Grace CPU, which can now be paired with AI accelerators produced by various partners.

Several companies, including Alchip, AlsteraLabs, Marvell, and MediaTek, have already expressed interest in adopting NVLink Fusion in their application-specific integrated circuits (ASICs). In the CPU domain, Nvidia is collaborating with industry giants like Fujitsu and Qualcomm, fueling speculation that Qualcomm may soon unveil its own server processors.

For general consumer electronics, such as desktop computers and laptops, the NVLink technology remains less applicable at this stage.

Nvidia's New RTX Pro Servers

In addition to the NVLink announcement, Nvidia introduced its RTX Pro servers, which will feature the recently launched RTX Pro 6000 Blackwell graphics card. This server variant boasts 24,064 shader cores and 96 GB of GDDR7 RAM, making it suitable for tasks such as rendering and AI processing. However, this model does not include the NVLink interface, which means communication between the graphics cards will rely on the slower 16 PCI-Express 5.0 lanes.

To facilitate inter-server communication, Nvidia has equipped the RTX Pro servers with the ConnectX-8 SuperNIC network chip, which supports Infiniband connections with a bandwidth of 800 Gbps. This chip functions as a PCI-Express switch and provides 48 PCIe 6.0 lanes for the RTX Pro 6000 cards. Companies such as Asrock Rack, Dell, HPE, Lenovo, Mitac, and Supermicro are among the vendors that will offer these advanced server solutions.

Conclusion

The NVLink Fusion initiative represents a critical advancement for Nvidia and its partners, potentially reshaping the landscape of high-performance computing. By allowing other manufacturers to utilize this technology, Nvidia is fostering a more collaborative environment in chip design and development.