U.S. Government Awards NVIDIA $8.5M for Exascale Research
The U.S. Energy Department has awarded NVIDIA its fourth multi-million-dollar R&D contract to accelerate the development of next-generation heterogeneous supercomputers.
The Office of Science and the National Nuclear Security Administration (NNSA) awarded NVIDIA $8.5 million over two years to develop advanced interconnect architectures that enable "exascale" computing systems. These computers would perform a billion billion (a quintillion) floating point calculations per second.
This research, which is part of the NNSAs DesignForward program, will build on the network interface work initiated in Nvidia's DOE FastForward project. NVIDIA will focus on developing network architectures that support throughput-optimized processors, massive and configurable bandwidth, a high degree of network resiliency, and sophisticated routing to improve sustained performance at scale.
One of the goals of Nvidia's research will be to support very high message rates on the order of 500 million messages per second per compute node. This requires supporting very short messages, just 200Bytes/message at 100GB/s bandwidth. That would be a dramatic improvement from todays software processing rates, which are in the range of a few million messages per second on a large system with latency-optimized processing cores.
Nvidia will collaborate with partners to define open standards for a software API and a NIC-to-router link protocol to facilitate interoperability with multiple network fabrics and with other processors.
This research, which is part of the NNSAs DesignForward program, will build on the network interface work initiated in Nvidia's DOE FastForward project. NVIDIA will focus on developing network architectures that support throughput-optimized processors, massive and configurable bandwidth, a high degree of network resiliency, and sophisticated routing to improve sustained performance at scale.
One of the goals of Nvidia's research will be to support very high message rates on the order of 500 million messages per second per compute node. This requires supporting very short messages, just 200Bytes/message at 100GB/s bandwidth. That would be a dramatic improvement from todays software processing rates, which are in the range of a few million messages per second on a large system with latency-optimized processing cores.
Nvidia will collaborate with partners to define open standards for a software API and a NIC-to-router link protocol to facilitate interoperability with multiple network fabrics and with other processors.
