Today’s server GPUs are typically connected by the PCI Express (PCIe) bus, which provides a communication bandwidth of 12 GB per second. However, NVIDIA’s new NVLink ultrafast interconnect technology is set to replace PCIe, delivering 5 to 12 times more bandwidth. NVLink allows GPUs to communicate directly with each other, significantly accelerating large-scale simulations. Additionally, NVLink enables high-speed communication between GPUs and CPUs, allowing both processor types to access system memory at the same speed. This results in further benefits for accelerated computing, enhancing overall system performance.
Press enter or click to view image in full size
each node A100, 8 A100s are connected using NVlink and 2nd set of 8 A100s are connected through NVswitch
NVIDIA technology in our supercomputer: NVLink and its associated technology, NVSwitch.
NVLink: Interconnects up to eight GPUs in our DGX systems, providing 10 times the speed of traditional PCIe connections.
NVSwitch: Extends NVLink’s capabilities, allowing interconnection of up to 16 A100 GPUs, achieving up to 10 petaflops of compute power.
