InfiniBand connections are used in supercomputers
As part of the SC20 supercomputer conference these days, Nvidia unveiled not only the A100 server graphics accelerator with 80GB of memory and the DGX Station A100 workgroup server , which has four such processors, but also the new generation of Nvidia Mellanox 400G InfiniBand technology.
New generation of Nvidia Mellanox 400G InfiniBand technology introduced
According to the manufacturer’s calculations, this is the seventh generation of the Mellanox InfiniBand. It is said to “guarantee ultra-low latency”, “double the bandwidth” and “add new network computing algorithms.” Implementing the latter, Nvidia says, is critical to AI challenges, allowing it to offload the computing load onto the network. Nvidia’s third-generation Mellanox SHARP compliant technology offloads and accelerates deep learning operations on the InfiniBand network, allegedly resulting in up to 32x faster AI tasks. The use of this feature does not require additional funds – the acceleration of scientific calculations is provided by the standard Nvidia Magnum IO software stack.
The Mellanox InfiniBand architecture edge switches deliver 51.2 Tbps aggregate bidirectional bandwidth and record over 66.5 billion pps throughput. The modular switches based on Mellanox InfiniBand will have an aggregate bidirectional bandwidth of 1.64 Pbps, five times that of their predecessors.
The Mellanox InfiniBand architecture is based on industry standards to ensure backward and future compatibility. Its solutions based on it should appear on the market in the second quarter of 2021.
