Optimal CPU Suggestion for RTX3060 or RTX3070?

User discussions
1

GROMACS version: 2020.4
GROMACS modification: No

Hi all, and happy new year! :)

After some gromacs gpgpu benchmarks I ran in a notebook (i5 8300H + pascal GTX1050M) went successful (test system consisted of two proteins in water with PME, ~238,000 atoms, performance jumped from 3.143ns/day to 7.899ns/day), I began to heavily consider the acquisition of a strong gpu at the new workstation I intend to buy.

A rapid search on the more up-to-date available GPU card models indicate that the RTX3060 and RTX3070 cards are among the best options nowadays.

However, a well chosen CPU (from either Intel or AMD) is obviously necessary to have a well-balanced system that would not create any bottlenecks.

As such: Which (minimum) CPU would be suggested to better pair with those GPUs, that would not lead to some sort of calculation bottlenecks?

Also, is there any other specification that I should look more carefully at the hardware configuration?

Thanks a lot in advance! :)

2

Go for the highest one with your available budget.
GTX 3070 is better and a processor of 8-12 core would be enough for gromacs.
See previous posts and gromacs benchmark paper of szilard pall

3

Hi @scinikhil

Thanks for the answer. I wasn’t aware of the updated benchmark paper (I had an older one in hand, which of course do not live up to today software and hardware improvementes). The latest version (http://cacs.usc.edu/education/cs596/Pall-GROMACS-GPU-JCP20.pdf) is a blessing, since it uses a very up-do-date GPU (RTX2060 super, which has a slighly lower cuda cores count than RTX3060) and CPU (Ryzen 3900X, which is one of the alternatives, side-by side with the more expensive Ryzen 5800X).

I’ll also search previous posts in the forum (I miss the mail list).

4

Hi,

Good to hear that you find the papers useful! Our most recent work is less extensive in terms of extensive benchmarks and review of hardware and mdrun settings, but feel free to ask if you have follow-up questions.

As I noted in previous posts, the sweet spot for CPU-GPU hardware balance tends to be quite low, but given the inherently heterogeneous design of GROMACS, using some features will require more CPU work than others. Hence it is always useful to consider the type of simulation and settings when comparing to published benchmarks.

Cheers,
Szilárd

1 Like