Suggestion about buy a system to run mcx and mcxlab

[Seyedmohammad Hosseini]

Hello Dr. Fang,

Greetings!

I am Mohammad Hosseini a PhD student working in Professor Andrew Berger's lab at the Institute of Optics University of Rochester. 

At the moment I'm running the mcx and mcxlab using a Linux-based university cluster. I need to buy a new system which enables me to run them on Windows. As I know the system should include an NVIDIA graphics card. I am considering 2 following options which both include NVIDIA® GeForce graphic card. Do these systems are right to run the simulation in Windows?

Inspiron 27 7720 (Intel) All in One Desktop: Desktop Computers | Dell USA

HP ENVY All-in-One 34-c1045t

Dell brand just has 1 option for graphic cards:

NVIDIA® GeForce MX550 2GB, GDDR6

While the HP brand has multiple options to select:

1. NVIDIA® GeForce® GTX 1650 (4 GB GDDR6 dedicated)

2. NVIDIA® GeForce RTX™ 3050 (4 GB GDDR6 dedicated)

3. NVIDIA® GeForce RTX™ 3060 (6 GB GDDR6X dedicated) with LHR Display Connectors: HDMI*1, DP*2

4. NVIDIA® GeForce RTX™ 3080 (8 GB GDDR6X dedicated) with LHR Display Connectors: HDMI*1, DP*2

Could you please help me with which of them is best to use for our simulation Also do you have any other suggestions in this regard?

Thank you for your time and consideration.

Mohammad

[Qianqian Fang]

which computer to buy depends on what your group intends to use this computer for - especially outside of the use of mcx simulations.

if the computer will be mostly used for mcx simulations, I don't recommend either of the options, because all-in-1 are designed for compactness/ergonomics, rather than performance.

to make the best use of mcx, my recommendations are the following

1. buy a (regular) ATX tower computer with ample interior space for adding dedicated/large GPUs, potentially multiple GPUs, for best performance, modern GPUs are getting big, and you must prepare for the new ones on the horizon (which is getting faster every generation)

2. buy intermediate to high-end consumer grade gaming GPUs (2080/3080/4080/4090) etc to get the maximum performance/per $, no need to use professional GPUs (A4000/A6000) or Tesla GPUs (V100/A100) GPUs as they are 1) expensive and 2) good at double precision and large memory, but mcx is not using either of those

3. do not use Windows, use Linux. Windows is very poor for development, takes tons of efforts for setting up environment for compilation, has poor compilers and toolchains (even you can build, it is not portable); plus, you can't use your GPU for both display and computing unless you set the TdrDelay in the registry. on the contrary, Linux has a slight learning curve at the start, but once you are familiar with it, installing/reinstalling gcc/make/cmake/cuda takes almost no effort. The workflow is highly portable, easily reproducible. Lastly, DO NOT USE a mac. Apple and NVIDIA had abandoned each other two year ago, and CUDA no longer works on Mac and Mac no longer supports nvidia GPUs.

if you do need mobility, buy a gaming laptop with a decent - 3060 or better GPU - to best use mcx's simulation power. Here is our benchmark page showing the huge difference between high-end (like 4090) and low-end GPUs (like 1065)

for my group, I currently run 20+ tower/rack servers, hosting an array of GPUs from different vendors for MCX/MCXCL/MMC development, testing and optimization. Nearly all of those were assembled by myself using the latest components bought at the time (clearly you can see I enjoy doing that) - you can get those components from Newegg, microcenter, or Amazon. This way, I can get the highest specs I can get for each component with a fraction of the cost comparing to buying from a branded product. Putting everything together is also a joy to me, check out my recent build

https://twitter.com/FangQ/status/1642907918239698945

https://twitter.com/FangQ/status/1666560985090629632

Qianqian

--
You received this message because you are subscribed to the Google Groups "mcx-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mcx-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mcx-users/f18cfc8c-7cc2-4a0d-a56a-8ac691e4ed6dn%40googlegroups.com.

[Qianqian Fang]

sorry, I forgot to include the mcx benchmark page link, here you go

http://mcx.space/gpubench/