In this video from the GPU Technology Conference, Steve Oberlin from Nvidia presents: Heterogeneous HPC, Architectural Optimization, and NVLink. NVlink will be the backbone of 150 Petaflop Coral supercomputers coming to ORNL and LLNL. As the world’s first high-speed GPU interconnect, NVLink will let data move between GPUs and CPUs 5-12 times faster than possible today.
The emergence of heterogeneous computing has demonstrated that the highest performance and efficiency can be achieved in a general way by tightly coupling compute engines optimized for latency-sensitive and throughput-oriented operations. This talk will explore heterogeneous node design and architecture and how NVLink, a new scalable node integration channel, enables uncompromising performance on the most demanding applications, using the next-generation DoE CORAL Summit and Sierra supercomputer systems as a case in point.
This video provides a quick overview of how NVLink works. In today’s servers, GPUs are connected by the PCIe, providing 12 GB/sec of communication bandwidth. NVLink interconnect makes high-speed communication between GPU and CPU possible, allowing both to access system memory at the same speed, further increasing the benefit of accelerated computing.