Dear Amber,
I"m having some difficulty figuring out amber's GPU performance on 3
different machines in my lab. Something's not right.
NOTE the RTX 3060 TI is a LHR version.
By the specs the RTX3060TI should be about 1x to 2x faster than the
RTX2060 Super but it is gets about 1/2x the ns/day.
And the RTX4090 does not seem to get expected speeds regardless if I
compare to the 3060Ti or 2060 Super.
`
Here's what I've collected when running the exact same GB simulation,
with the same amber22 pmemd.cuda_SPFP executable, libraries, OS-version
etc.. on the three different computers.
The system has 47407 atoms and all simulations use
gbsa=3 :Compute surface area using a fast pairwise approximation
suitable for GPU machine.
The timings below are from amber's "inf" file the number multiprocs and
frequency are from amber's log file and the flops (single precision
flops) from wikipedia. Driver reported by nvidia-smi and libraries by ldd
GPU TIME #MultiProcs Freq Flops Driver libraries
CUDA Device Name: NVIDIA GeForce RTX 2060 SUPER 4.35 ns/day 34
1.65Ghz 6.7-7.2 535.113.01 CUDA Version: 12.2(11.7 libraries)
CUDA Device Name: NVIDIA GeForce RTX 3060 Ti 2.02 ns/day 38
1.66Ghz 13.7-16.2 545.29.06 CUDA Version: 12.3 (11.7 libraries)
CUDA Device Name: NVIDIA GeForce RTX 4090 7.38 ns/day 128
2.52Ghz 73.1-82.6 545.29.06 CUDA Version: 12.3 (11.8 libraries)
Let me know what more to post and I'll be happy to provide info.
Tom
--
***********************************
Thomas C. Bishop
Hazel Stewart Garner Associate Professor
Chemistry & Physics
Tel: 318-257-5209
https://dna.engr.latech.edu
Office Hrs: Winter 2023-24 Eng Annex Room 207
M-F: 1:30-3:30 or by appt
***********************************
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Feb 06 2024 - 11:30:02 PST
