2 ms·
I believe this is the paper in question! https://futhark-lang.org/publications/fhpc16.pdf https://futhark-lang.org/publications/fhpc16.pdf
by girvo 10y ago
I believe this is the paper in question!
https://futhark-lang.org/publications/fhpc16.pdf https://futhark-lang.org/publications/fhpc16.pdf
- BMarkmann 10y agoOh wow, hard numbers, thanks! That's very impressive, but strangely misleading in the graph towards the bottom that appears to clock TAIL->Futhark->OpenCL and Futhark->OpenCL exactly the same in more than half the cases (if you look at the raw numbers if appears that the second approach blows the first away in most).
- Athas 10y agoThat's because the speedup graph cuts away at a 100x speedup, but some of these (somewhat silly) benchmarks are compute-bound and hence gets these crazy 1000x speedups on the GPU.