At Computex, Nvidia introduced NVLink Fusion, allowing non-Nvidia accelerators to utilize its high-speed interconnect technology. NVLink enables GPUs to work cooperatively, sharing compute and memory resources effectively. This fifth-generation technology supports 1.8 TB/s bandwidth and can connect multiple GPUs in one rack. The two configurations of NVLink Fusion focus on connecting custom CPUs with Nvidia GPUs or with other accelerators, though it necessitates Nvidiaâs involvement for integration. This initiative aims to potentially allow superchip-style compute assemblies while maintaining a layer of exclusivity.
NVLink Fusion opens Nvidiaâs high-speed interconnect to other semi-custom accelerator designs, enhancing connectivity and performance across diverse compute platforms.
The technology enables custom CPUs to connect seamlessly with Nvidia GPUs, boasting 14x higher bandwidth than PCIe 5.0, thus enhancing compute efficiency.
It leverages an interconnect that potentially integrates various CPU and GPU combinations from various manufacturers, but retains exclusivity by requiring Nvidiaâs involvement.
Nvidia remains cautious about the openness of NVLink Fusion, ensuring that only Nvidia CPUs can connect to non-Nvidia GPUs, thereby maintaining control over the ecosystem.
Collection
[
|
...
]