but a friend of mine has done something similar and it's about 140 times faster than the current code (using MKL BLAS/VML). Not related to R but note that CUDA include a BLAS (cuBLAS) and that it has been binded in .Net. So you can quickly test it from say C# or F#. Here. In most case the perf. may be disapointing since the cost is in moving the data from main memory to on-board memory. You may have to re-organize your code for that...GPU - VPU anyone messing around with this?
-
Unfortunate Conflict Of Evidence
- Posts: 0
- Joined: Thu Jan 01, 2004 12:00 am
GPU - VPU anyone messing around with this?
Well... just tried to rewrite some blocked Gibbs sampler for a piece of MCMC, it's still buggy : writing good kernel is not easy but a friend of mine has done something similar and it's about 140 times faster than the current code (using MKL BLAS/VML). Not related to R but note that CUDA include a BLAS (cuBLAS) and that it has been binded in .Net. So you can quickly test it from say C# or F#. Here. In most case the perf. may be disapointing since the cost is in moving the data from main memory to on-board memory. You may have to re-organize your code for that...
but a friend of mine has done something similar and it's about 140 times faster than the current code (using MKL BLAS/VML). Not related to R but note that CUDA include a BLAS (cuBLAS) and that it has been binded in .Net. So you can quickly test it from say C# or F#. Here. In most case the perf. may be disapointing since the cost is in moving the data from main memory to on-board memory. You may have to re-organize your code for that...All your Bayes are belong to us