Project

General

Profile

Feature #1227

double precision support in GPU acceleration

Added by Szilárd Páll over 6 years ago. Updated over 1 year ago.

Status:
Rejected
Priority:
Normal
Assignee:
Category:
core library
Target version:
-
Difficulty:
uncategorized
Close

Description

The CUDA GPU accelerated non-bonded kernels in 4.6 do not support double precision. The main reason for this is that
  1. some GPUs have very low double precision arithmetic throughput (e.g Kepler1 1/24-th of single precision);
  2. no native double precision atomic operations (can be emulated with CAS, see more here and here).

While 1. would require only warning the user that double precision is extremely slow (and perhaps switching to CPU kernels by default), 2. would need generating the extra set of kernels as well as some performance tweaking.

Most of this is more or less straightforward, but for now not considered of high priority.

History

#1 Updated by Mark Abraham over 3 years ago

  • Target version deleted (future)

#2 Updated by Erik Lindahl almost 2 years ago

Given how the double precision performance has developed on cheap GPUs, I see no reason to invest efforts in this (compared to the many other higher-priority features we want), so I would suggest we close it.

#3 Updated by Mark Abraham almost 2 years ago

Agreed. I have a hard time imagining when we would ever do this.

#4 Updated by Mark Abraham over 1 year ago

  • Status changed from New to Rejected

Also available in: Atom PDF