Guidance for purchase of workstation and gpu

[Parmest Roy]

Hello everyone,

Our research group wishes to buy a workstation and a compatible gpu for running mumax3. We are trying to keep things under a budget of USD 1500. If you can recommend us good workstation + gpu combinations in this budget, it would be extremely helpful.

Thank you

Parmest

[Parmest Roy]

I'd be highly grateful if you could suggest something that works for you.

Thank you 

Parmest

[Felipe Garcia]

Hello Parmest,

For that price you can get the low end of Nvidia. Or one new card from the former RTX 30X0 family. For example, one single RTX 4080 can be more expensive than your whole budget. Of course, if one buys a nice workstation one likes to have more than one GPU, but that will exceed your budget.

A RTX 4060 Ti

https://www.amazon.com/Gigabyte-GV-N406TAERO-OC-8GD-Graphics-WINDFORCE/dp/B0C5C1N668/

or RTX 4060

https://www.amazon.com/Gigabyte-Graphics-WINDFORCE-GV-N4060AERO-OC-8GD/dp/B0C8K3DFD9/

This is an example, I don't work for Gigabyte, you can find more examples. The Nvidia range of GPUs is always changing and being more expensive does not necessarily faster. Ti versions are better than the same card with Ti.

One can take a look here of the comparison of the different families (of course it has to be Nvidia, AMD cards are not valid).

https://www.techpowerup.com/gpu-specs/geforce-rtx-4060-ti-8-gb.c3890

Best regards,

Felipe

--
You received this message because you are subscribed to the Google Groups "mumax2" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mumax2+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mumax2/8a877ed8-0ceb-4b2e-98c0-d4a1ee78c89bn%40googlegroups.com.

[Антон Луценко]

Hi all, 

Business-grade GPUs like Tesla P100, A100, V100, and workstation GPUs like A2000 are significantly cheaper and offer the same performance in simulations as gaming-grade GPUs. Our cluster uses the former three and my workstation has the latter. 

[Felipe Garcia]

Hi Anton Lutsenko,

I am very surprised to read that. The cost of those Tesla P100, A100, V100 in their time exceeded those of the gaming by a lot (they can reach 10000$ or beyond). The performance is not equal because mumax3 is optimized for single precision. If one takes a look at that graph, gaming GPU are superior in price and performance to the Tesla card of their time because Tesla are optimized for double precision but not for single

http://mumax.github.io/gpus.svg

These numbers are for millions of cells. For smaller systems any GPU will do the job.

If one considers for example the RTX 3090 (and sorry using wikipedia as source https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units) has

30 Tflops in single precision and 0.459 Tflops in double precision. The single precision of Tesla V100 is around 15 Tflops in single precision and around 7Tflops in double precision. For mumax3, one has to consider only the single precision. For a generic cluster double precision is also important. Therefore, looking at that list, one can find a compromise, which in my opinion is the low end of the gaming GPU. The third parameter to consider is the memory bandwith.

Unless you tell me that your provider has plenty of cheap Tesla (maybe he got a stock for a former bitcoin farm or miners), I would go for a reasonable RTX 3000 or RTX 4000 families. Of course, in the case you have them available for less than 1500 dollars then a Tesla is a good deal. The same applies to A2000 (8Tflops in single precision). It is a very good choice but then I can not find it for less than 600 euros. Of course that depends on the actual country.

Regards,

Felipe

To view this discussion on the web visit https://groups.google.com/d/msgid/mumax2/c0a2444b-3ac6-47a5-8a5d-f85628471ef3n%40googlegroups.com.