3 ms·
Maybe for the competitive consumer graphics market, but I don't see them disrupting Nvidia's hold in the HPC world. With Nvidia's NVLink, you can achieve high b
by jclay 8y ago
Maybe for the competitive consumer graphics market, but I don't see them disrupting Nvidia's hold in the HPC world. With Nvidia's NVLink, you can achieve high bandwidth data transfer between graphics cards without having to pass through the PCIe to CPU. If you have large computations that require syncing data among multiple GPU's, NVLink is your most performant option by a large margin. That's not to mention how much further ahead CUDA is from Sycl, OpenCL, ROCm, etc. I certainly welcome the competition (and hopefully an open standard). In my experience, CUDA is ahead on developer tooling, performance and productivity features (Thrust, Unified Memory, C++17 support, etc).
- ct520 8y agoAMD's NVLink is infinity fabric (MI60, MI50) isnt it? with 200gb card to card transfer rate... Did you take that into account? I agree the toolset is light years ahead of what I seen from AMD, but I'm not much into that.
- jclay 8y agoAh, I hadn't seen that announcement. It looks like this was announced recently (November 2018) and it's still 100Gb/s slower throughput than the latest NVLink. Either way, it could certainly be interesting if they compete on price, but I cannot find any information yet on actually using it to sync with OpenCL / ROCm.