Most MPI parallel applications require low-latency and high-bandwidth compute network. Compute-storage converged solution becomes popular. Low-latency and high-bandwidth network also helps to improve the IOPS and I/O throughput of parallel storage system.
InfiniBand is coming!
|PCI-E||Theoretical Bandwidth||x||IB Encoding||x||PCI-E Encodin||=||Peak Bandwidth||Latency|
|QDR||PCI-E 2.0 x8||40Gb/s||x||8/10||x||8/10||=||3.2 GB/s||1.3 μs|
|FDR||PCI-E 3.0 x8||56Gb/s||x||64/66||x||128/130||=||6.68 GB/s||0.7 μs|
|EDR||PCI-E 3.0 x16||100Gb/s||x||64/66||x||128/130||=||11.93 GB/s||0.6 μs|