Timeline for Is there a constrained nonlinear optimization library like IPOPT that runs on GPUs?
Current License: CC BY-SA 3.0
8 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 22, 2019 at 10:36 | comment | added | Turbo | @GeoffOxberry Some links are not working. | |
| May 10, 2017 at 4:22 | comment | added | Andrew Hundt | cuSolver is NVidia's set of cuda solvers, and it does provide APIs like csrlsvchol which are based on Cholesky factorization docs.nvidia.com/cuda/cusolver | |
| Oct 7, 2015 at 18:14 | history | edited | Geoff Oxberry | CC BY-SA 3.0 |
added 95 characters in body
|
| Oct 7, 2015 at 15:28 | vote | accept | cauchi | ||
| Oct 7, 2015 at 15:24 | vote | accept | cauchi | ||
| Oct 7, 2015 at 15:28 | |||||
| Oct 7, 2015 at 2:12 | comment | added | Brian Borchers | For SDP, the matrix that has to be factored in each iteration is most often dense and the Cholesky factorization parallelizes well on a shared memory machine. The construction of this matrix is also fairly easy to parallelize. Thus it's possible to get reasonably good parallel speedups on a shared memory multiprocessor with an interior point method for SDP with numbers of processors up to about 64 (as far as I've gone.) | |
| Oct 7, 2015 at 2:09 | comment | added | Brian Borchers | Parallelizing IPM's for LP boils down to parallelizing the sparse Cholesky factorization. This is not a well saved problem, particularly on GPU's. Parallelizing simplex methods for LP is much harder and there's been very little useful work in that area- if you want to solve a large scale LP in parallel you'll probably want to use an IPM. | |
| Oct 6, 2015 at 23:33 | history | answered | Geoff Oxberry | CC BY-SA 3.0 |