2 ms·
(nit picking) 5000ns (aka 5us) seems ... too high ... for PCIe. Infiniband, connected over PCIe can be sub 600ns for latency[1] for 2 generation old hardware.
by hpcjoe 4y ago
(nit picking) 5000ns (aka 5us) seems ... too high ... for PCIe. Infiniband, connected over PCIe can be sub 600ns for latency[1] for 2 generation old hardware.
Still, that's around 1800 clock cycles for a 3GHz CPU. Which implies there is a floor, below which you want to keep computations local on-CPU versus communicating over PCIe. Whether it be to a remote system via RDMA, or to a PCIe connected accelerator.
[1] https://www.aravision.com/products/infiniband/pdf_mellanox_adapter/PB_ConnectX-5_VPI_Card.pdf https://www.aravision.com/products/infiniband/pdf_mellanox_a...
- dragontamer 4y agoI appreciate the hard data from a different PCIe system. I came to the 5000ns estimate because I tested some null-kernel a long time ago on a GPU. Basically a "hello world" of GPU kernels, and I tested to see how long it took for the kernel to be called. I hear that GPUs have sped up a bit since my test a few years ago, but I'd expect it to still be around 5000ns for a "CPU-GPU ping-pong".