[AleDP]

Hello Everyone, I found this wonderful community after a couple of weeks of documentation review in the free time. Unfortunately my knowledge about hardware is non-professional and I am not really able to translate some information available on this forum to my specific problem.

I am a PhD student working inside a company and they are willing to dedicate 2000eur budget for a desktop configuration for CFD educational purposes (and some FEM, which is secondary). In other words, I will be using it to study cfd and eventually perform some simple modelling rather than a professional use, hence the "low" budget.

No used components can be acquired as consumer support and warranty on the workstation is required by our purchasing department (I'm the newbie so I don't want to be the one who asks for exceptions)(consider then not only the price of raw componets but a little surplus added by the assembler).

My line of thought was that it is pointless to try acquiring a server solution and even an AMD threadripper configuration, with my budget.
An example of configuration I am evaluating is:

cpu. R9 5900X
mb: Asus Tuf x570-plus gaming
ram: ddr4 corsair vengeanche rgb pro 3200 MHz CL16 4*16GB
gpu: Nvidia Geforce rtx 3060ti 8GB
hdd. samsung 970evo plus NVme 500GB
CoolerMaster MasterLiquid Lite 120 liquid cooling (for silence, efficiency and in the remote chance I will overclock the system?).

price:1850eur (excluded vat)

This configuration was developed by me considering that Ansys allows only few cores on the educational license and to add cores is expensive, same goes with HPC license that allows GPU acceleration: I looked for the highest single core performance cpu (given by clock and ipc).
I tried to maximise the ram to allow "bigger" models also in FEM analysis and also a relatively high speed (it is possible to have 3600MHz but only 4*8GB). I opted for a high end gpu to allow for other programs, such as nTopology, Inventor etc., to run smoothly and in general for the good price/performance.
My main doubt is that AMD supports only dual channel ram, so I am afraid that I won't be obtaining the expected performances with 6cores/channel, even at high ram bandwidth. Moreover, I have been reading about MKL issues with AMD Cpus which sounded a little like conspiracy theories but, as I am an amateur, they worry me.

I am conscious that my reasoning may be faulty in many points, expecially regarding the exclusion of server cpus, given also that, as a student, I could try to switch from ANSYS cfx/fluent to OpenFOAM, save licensing fees and, even if literature suggests that it requires finer meshing, obtain better overall performaces. Also, the budget was approved upon my request: I could have asked for more but opted for an "high end office pc" in order not to waste money that could be used for other research, on something that is, for now, educational. If you tell me that with a grand more I could acquire a machine satisfatory enough for professional use, I am all ears.

In conclusions, what do you think of my configuration? What are your suggestions? Thank you all in advance and Best Regards!

[pbasis]

Seems respectable, however a couple notes. With Fluent, a high-end GPU won't help much. Codes need to be specifically tailored for GPU acceleration, and Fluent isn't one of those code (in most instances). There are a few solvers within Fluent that are GPU optimized, but the performance gain will be marginal and you'd be better off with allocating your money towards a better CPU (search the hardware forum for "GPU Fluent" for more info).

I am not super familiar with the various offerings across the Ryzen line (maybe someone else can weigh in), but generally speaking memory bandwidth, memory speed, and cache size matter as much if not more than clock speed. Also when comparing options, bear in mind that all the cores on a socket will have to compete for shared resources (memory, cache, etc).

Also, I am not sure how Ansys handles academic licenses, but if it's just a lower cost variation of the standard pricing scheme, there are specific numbers of cores that are license-optimized. I'd double check that what I am telling you applies for academic licensing, but basically the way it works for standard licensing is that you need one "solver" license which allows you to run the program plus you get 4-cores for free along with the solver. Then "HPC Packs" allow you to use more cores, but they scale non-linearly based on the equation, (cores allowed)=2^(2*(HPC Packs)+1) + 4. Thus, 1 HPC gives you 12 cores, 2 HPC gives you 36 cores, 3 gives you 132 cores, etc. You can leverage the fact that the marginal cost of licensing an additional core is zero until you cross one of these thresholds.

[AleDP]

Thank you very much for this first indication. My biggest problem is that switching form Ryzen (dual channel ram) to something less limited, would go from 1.8 straight to 3k for a similiar configuration with threadripper.

Thank you also for explaining hpc licensing, i will keep that in mind. Let us then wait if someone has experience with Ryzen (or amd in general) and can suggest a better amd configuration (also with less powerful and chaper AMD graphics card?).

Best Regards,

Alessandro

[LaminarFlow]

If you want non-ECC RGB RAM for Ryzen I would definitly go with Micron E-Die (Crucial Ballistix). Samsung B-Die is also very popular for Ryzen but, if not changed, limited on 8 GB (SR) and 16 GB (DR). So you would need 4x DR modules which means higher load on the RAM controller and reduced effort of the high B-Die frequencies. E-Die is cheaper and the sweatspot between price and performance. In the beginning of 2020 Balistix 32 GB Modules came onto the market (it exists also a 32 GB DR ECC Kingston E-Die module). So I would take 2x32 GB Crucial Ballistix or the ECC pendant.

For cooling I would prefer an Arctic Liquid Freezer with more than one Fan (240, 280, 360).

The RTX3060ti is hard to get at the moment. Will you really profit from this performance? nTopology and Inventor are supporting OpenGL. So you will only profit from Cuda in ANSYS. Regarding to this test of flotus1 I am not sure if the RTX3060ti is worth the price: GPU acceleration in Ansys Fluent
If an OpenGL card satisfies your requirements I would go with AMD. A RX580 is cheaper, has as much VRAM as the RTX 3060 ti and is not that much worse in terms of double precision/FP64 performance (0,51 vs 0,39 TFLOPS). In Floating Point Performance/FP32 the difference is factor 3, what seems to be much at first glance. But you should also consider that the RTX3060ti has on the sheet more than twice the FP32 and FP64 performance than the fast and older Quadro RTX4000.

[AleDP]

Originally Posted by LaminarFlow

If you want non-ECC RGB RAM for Ryzen I would definitly go with Micron E-Die (Crucial Ballistix).

First, thank you for your reply.
I have been advised by a friend of mine to consider buying ECC ram as the system will be more stable. Do you have experience with stability issues' will the extra premium be worth it or should I try and maximise performances at all cost with gaming non-ecc components)
Thank you for the hint on the crucial 32GB, I'm taking a look at it.

Originally Posted by LaminarFlow

For cooling I would prefer an Arctic Liquid Freezer with more than one Fan (240, 280, 360).

So, you are confirming I was right to think about liquid cooling? I'll look the model you pointed out.

Thanks also for the dissertation about graphics card, I was really worried that the 3060ti would have been completely "wasted", I shall then allocate my money on better RAM modules and use an AMD gpu, which should also benefit from a better cpu gpu optimization by AMD?

Thank you again for the precious help.

Alessandro

[LaminarFlow]

ECC RAM prevents tilting of a single bit which can cause huge errors especially with fixpoint iterations, or even more dangerous errors you don't recognize. The probability of such errors is low, but preventing possible errors is still the best way. Since this machine is used in the industry/ a company I would also prefer ECC RAM. 'You can not reach the same RAM speed / latency with this kits, but the mentioned E-Die Dual Ranked 32 GB memory is not that expensive (in germany) and has still a decent performance: https://geizhals.de/kingston-server-...loc=at&hloc=de

What do you pay for this RAM in italy? What is the link/name of good italian websites to compare prices? If you choose another fast module (because Ryzen profits from RAM speed and latency) you have to take normal ECC RAM instead of reg ECC, because Ryzen/your mainboard does not support buffering.

Ryzen is easier to cool than the Intel competitors. Nevertheless cooling of a 5900x under heavy and longterm load can cause lower clockspeeds with aircoolers. So I would choose something like the Liquid Freezer. On the downside the manufacturing scatter / variance of AiO coolings are not that good, the cheap pumps cause a higher noise level compared to a good aircooler and there is a small risk of coolant leaking with aging, what can damage your hardware.

Ryzen 5000 in combination with AMD GPUs can slightly increase the performance. This will work in games and definitly with newer GPUs. But I do not think that you will profit from that now. So if you want a Nvidia GPU for CUDA, go for a cheap one (e.g. GTX 1650 Super with 0,138 TFLOPS FP64). But in terms of OpenGL Performance the RX580 is the better choice (doubled VRAM, almost factor 3 more FP64 performance).

[AleDP]

As we are a medium company, working with a small PC assembler, I am waiting to chat with him, and I will discuss using the info you provided. I was hoiping it would have been today but we had to postpone. I will keep you updated, thanks a lot for all the precious help!