Re-imagine how data flows between computational elements of a supercomputer…
…not by arranging the transfer of a large data packet from a source processor to a destination processor but rather by the simultaneous transfer of shorter packets from multiple sources to multiple destinations.
For many applications, data communications are a more significant bottleneck than processor speeds. Nevertheless, the majority of the market is chasing multi-core processor and many-core accelerator technologies fueled by the TOP500 list and the race to Exascale.
Accelerators like GPUs or many-core co-processors are good for HPC. They enable a lot of computing through a lot of FLOPS. We believe you should examine your algorithms and decide which ones can be optimized with co-processors and which simply need better communication, not more cores. We provide a low-latency, congestion-free network for the Hard Problems.