recommendations for a fast multicore workstation

From: Jeff Comer (
Date: Tue Nov 24 2015 - 11:56:54 CST

Dear NAMD users,

I'm looking to buy a new multicore workstation for single-node NAMD
runs. I currently have 8-core/2-GPU and a 6-core/1-GPU workstations. I
show the specs and performance data for these below. I was wondering
if anyone had any recommendations on what might be the best way to
spend ~$3000–$5000. I'm currently using GeForce GPUs. Should I switch
to more expensive ones? Should I get more GPUs per core (e.g.
8-core/3-GPU) or fewer (10-core/2-GPU)? Should I get faster CPUs?

The tests below were done with NAMD_2.11b_Linux-x86_64-multicore-CUDA.
I get a 25–35% improvment moving from NAMD 2.10 to 2.11b. As you can
see, my 8-core/2-GPU machines run NAMD 90% faster than the
6-core/2-GPU machine, for a large (200K atom) system. Perhaps some of
this performance data will be useful for others.

CPU: Intel E5-1660v3 3.0GHz 8core Haswell
GPU: 2 x NVIDIA GTX 770 with 2GB GDDR5 memory
NAMD2.11b1 @oso 8 procs 2 gpus 192385 atoms ns/day=5.24
NAMD2.11b1 @oso 8 procs 1 gpus 192385 atoms ns/day=3.02
NAMD2.11b1 @oso 8 procs 0 gpus 192385 atoms ns/day=1.26
NAMD2.11b1 @oso 8 procs 2 gpus 10883 atoms ns/day=74.02

CPU: Intel E5-1650v2 3.5GHz 6core
GPU: 1 x NVIDIA GTX 770 with 2GB GDDR5 memory
NAMD2.11b1 @conejo 6 procs 1 gpus 192385 atoms ns/day=2.79
NAMD2.11b1 @conejo 6 procs 0 gpus 192385 atoms ns/day=0.67
NAMD2.11b1 @conejo 6 procs 1 gpus 10883 atoms ns/day=46.12

CPU: Intel Core i7-4820K 3.7GHz CPU
GPU: 1 x NVIDIA GTX 770 with 2GB GDDR5 memory
NAMD2.11b1 @mapache 4 procs 1 gpus 192385 atoms ns/day=2.41

For completeness, my NAMD configuration is shown below.


set temper 300

outputName $this
numSteps 4000
structure ../${sys}_NaCl.psf
coordinates ../${sys}_NaCl.pdb

# temperature control
langevin on
langevinDamping 1
langevinTemp $temper
langevinHydrogen off

# pressure control
useGroupPressure yes
useFlexibleCell yes
useConstantArea yes
langevinPiston on
langevinPistonTarget 1.01325
langevinPistonTemp $temper
langevinPistonDecay 300
langevinPistonPeriod 400

# parameters
parameters ../par_all36_prot.prm
parameters ../par_all36_GolP_special2.prm
parameters ../par_gold_111_prot_v0.prm
parameters ../toppar_water_ions_prot.str
paraTypeCharmm on
exclude scaled1-4
1-4scaling 1

switching on
switchDist 8
cutoff 9
pairListDist 11

# integration
rigidBonds all
timestep 2
nonBondedFreq 1
fullElectFrequency 2
stepsPerCycle 10

# output
binaryOutput yes
binaryRestart yes
wrapAll yes
wrapNearest yes
comMotion yes

outputEnergies 1000
outputPressure 1000
outputTiming 1000
xstFreq 4000
dcdFreq 4000
restartFreq 4000

# electrostatics
pme on
pmeGridSpacing 1.2

Jeffrey Comer, PhD
Assistant Professor
Institute of Computational Comparative Medicine
Nanotechnology Innovation Center of Kansas State
Kansas State University
Office: P-213 Mosier Hall
Phone: 785-532-6311

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:22:15 CST