3 ms·
If you care about GFLOPS/Watt/$ then GPUs beat everything. Real requirements are usually more subtle, including things like compatibility with existing software
by wmf 5y ago
If you care about GFLOPS/Watt/$ then GPUs beat everything. Real requirements are usually more subtle, including things like compatibility with existing software, latency (single-thread performance), and management overhead per node.
- joe_the_user 5y agoThat is my impression, it would still be good be good have an at least semi/partly objective evaluation of all the parallel compute devices that one can see appearing. Also, to have a number of benchmarks involve memory, caching, etc.
- bch 5y agoI’m somewhat ignorant here, and indeed also niche (I’m a NetBSD user), but: > GPUs beat everything that’s workload dependant, and also subject to support, no? Doesn’t GPGPU basically mean CUDA currently, and therefore beholden to Nvidia support for your hardware/software platform?
- dragontamer 5y agoAMD GPUs are supported over their open-source ROCm builds. HIP (CUDA-clone) works with Vega and CDNA. It seems like RDNA (5xxx and 6xxx cards) are less supported, but reports are that OpenCL kinda-sorta works (RDNA cards have a very different architecture than Vega / CDNA)
- UncleEntity 5y agoNah, I was playing around with an AMD integrated gpu a while ago using openCL (I believe).
- Const-me 5y ago> Doesn’t GPGPU basically mean CUDA currently Depends. If the programmers take the path of least resistance, sure it gonna be CUDA. Personally, I made quite a few things on top of DirectCompute i.e. vendor agnostic.