Hi Peeps,
The new GPU-Tachyon code has now been merged into master - I.e. the trunk
now contains the GPU Tachyon code. Thus from now on if you are editing
PMEMD code
(particularly GPU code) please edit the master branch and NOT gpu-tachyon.
Scott you may want to make a new branch from master for the TI code and we
can merge it in later or you can work directly in the master - your choice.
GTX-Titan is now the gold standard for saved output now.
The following are the new features of this code:
1) 25 to 30% speed improvement for single GPU
2) Peer to Peer support for parallel - gives MUCH better multi GPU
performance (up to 200ns/day for JAC NVE on 2 GTX-Titans) BUT it very much
depends on your motherboard chipset, GPU slot locations etc. - This is
work in progress to make this simpler.
3) Monte Carlo Barostat Support (although currently broken, see below)
4) ScaledMD support
5) Extra points on multi GPUs (although currently broken, see below)
6) Jarzynski support
7) Fixes for NMR restraints on multi GPUs. (although currently issues -
see below)
8) Support for SM1.3 hardware is now deprecated.
9) A potential issue with inaccuracy in long NPT runs due to cumulative
roundoff error on both CPU and GPU (which has been there since before the
ark) has been fixed.
10) Improved skinnb error message for massively shrinking box sizes during
NPT.
11) aMD support for both raising minima and lowers barriers.
12) Multi-D REMD (Adrian and Jason please clarify support here?)
13) GBSA (via CPU)
14) H Mass Repartitioning Support (4fs timesteps 200ns+ on single GPU for
JAC NVE HMR - woo hoo! - take that ACEMD / Gromacs!).
15) SPSP and SPDP precision models now deprecated. DPFP replaces DPDP.
SPFP is now the default.
We have the following problems right now from the tests that need fixing
before release - volunteers please. I will create bugzilla reports for
these:
Serial (1 GTX-Titan)
--------------------
DPFP All tests pass EXCEPT the two MCBAR tests which are horribly broken
(bugzilla bug 246)and multi-angle restraint test where GPU output is stuck
at value of 0.400 (bugzilla bug 240)
SPFP All tests pass EXCEPT the two MCBAR tests which are horribly broken
(bugzilla bug 246) and multi-angle restraint test where GPU output is
stuck at value of 0.400 (bugzilla bug 240)
Parallel (2 GPU) - Peer to Peer NOT available
---------------------------------------------
In parallel right now on MSKCC machine (Scott I can give you a login if
need be). If CUDA_VISIBLE_DEVICES is unset, .OR. 0,1 .OR. 2,3 then one
gets:
cudaDeviceEnablePeerAccess all CUDA-capable devices are busy or unavailable
cudaDeviceEnablePeerAccess all CUDA-capable devices are busy or unavailable
Should this not default to NO peer to peer mode? - Bugzilla Bug 245
amd/rna_gb Run.gb.amd2 - Totally broken in parallel = EAMD_Boost is wrong
(Bugzilla Bug 247)
amd/rna_gb Run.gb.amd3 - Totally broken in parallel = EAMD_Boost is wrong
amd/dhfr_pme Run.pme.amd2 - Totally broken in parallel = EAMD_Boost is
always zero.
amd/dhfr_pme Run.pme.amd2 - Totally broken in parallel = EAMD_Boost is
wrong.
amd/gact_ips Run.ips.amd2 Same bug
amd/gact_ips Run.ips.amd3 Same bug
4096wat_oct Run.pure_wat_oct_NPT_NTT1 - Radically different on step 1
(Bugzilla Bug 247)
large_solute_count /Run.ntb2_ntt1 - Radically different on step 1 (same
bug???)
chamber/dhfr_cmap_pbc Run.dhfr_cmap_pbc_charmm_NPT.md - Radically
different on step 1 (same bug???)
tip4pew Run.tip4pew_box_npt - extra points still broken in parallel.
tip4pew Run.tip4pew_oct_npt - extra points still broken in parallel.
tip5p Run.tip5p_box_npt
tip5p Run.tip5p_oct_npt
mcbar tests - broken in parallel (but this is because they are broken in
serial too and because extra points are broken in parallel).
Remainder of tests pass fine.
REMD - PARALLEL
---------------
Not currently tested - somebody needs to test this with the NEW git master.
All the best
Ross
/\
\/
|\oss Walker
---------------------------------------------------------
| Associate Research Professor |
| San Diego Supercomputer Center |
| Adjunct Associate Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sat Jan 11 2014 - 22:00:03 PST
