[Fer_Arus]

Hi, I have been around for a while, reading this section of the forum mainly , but this is my first post.

Im a student working on a Formula Student team, very new on CFD. Im using Fluent to study a solution for a over heating brakes in our previous car, having trouble to do so with a i7-2760QM , 8GB RAM laptop. I could buy another 8GB but I dont know how much Id win. The thing is: It would improve from having more RAM (would be used for sure) and I guess for been able to use dual channel (didnt know this when I changed 2x2 GB for 1x8GB) but how much? I dont know if worth expend almost 70 e right now for another 8GB (Cost me 47e 18month ago) that would be unused in any other task.

I would like to start a post in which several options could be taken into account and several questions could be solved. Perhaps could be use as a guide for someone who want to build a low budget WS ( could be expanded to other budgets too).

My needs would be between 5M elements ( 90% of the time) and 25M (not the main purpose but has to be archivable ), external flow with heat transfer sometimes. But lets take a more general point of view as could be helpful to understand the hardware needs of CFD. The results could be a guide where anyone can decide what its best for his needs.

Whats important in CFD (this will be edited):

Memory Bandwidth : This seems to reduce the selection for a proper WS to socket 2011 and 2011-3 (4ch memory controller), but lets take into acount latest 2ch CPUs with fast RAM.

Number of cores: not always the higher the better but needed if the number of elements goes up. Any suggestion about what would be a correct ratio between number of cores and memory used for the mesh?

I will start with a few options and try to maintein the descriptions short.

CPUs/builds:

Socket 2011:

Dual e5-2670 with 64gb ddr3 1600 MhzRAM (server Intel MB)(16cores)(Memory BW?): Really cheap build (around 1000e) for what it can do, I guess there would be no problem at all with a big mesh, but how would stand vs other options in smaller problems? And why not, how it would perform on other non-CFD task, energy consuption etc.

I7-3930k (6 cores)(Memory BW?): CPU around 150e, but really expensive Motherboards (at least in comparison) Seems like a build would be no more than 100e cheaper than a dual xeon.Opions?

i7-4820k (4cores): Similar to previous situation in term of price.

i7-4960x(6cores): For the price of this used CPU I guess a 6800k is better option

Socket 2011-3:

i7-6800k(6cores)(MBW?): CPU for 415e and a motherboard arround 250e for what its seems one of the best options, plus it would accept fastest RAM. How could perform on a 20M mesh?

A build with 32gb of ram cost arround 1300e. In case a few more gb is needed for a few simulations per year, would get the job done?

Others CPUs?

2ch memory controler:

I7-7700k(4cores): Build with similar price to the dual xeon. My doubt here is how competive would be with 3000 or 3200 Mhz RAM vs dual xeon in a 5M elements problem, and if would tackle a 20M cells problem.

Ryzen 7 1700 (8cores) (MBW 47 GB/s with 3200MHz RAM): same price of above. Similar pros and cons but with 8 cores. Could we expect an advantage vs the above in larger problems?

Other Ryzen: the only option viable I see is the ryzen 5 1600 with a B350 motherboard, 6 cores and the cheaper option of all. Seems like Ryzen 1700x and 1800 goes too close to the i7-6800k to be better option for this particular use.


My intention is edit this post with your answer, I will try to visit it often. Any sugestions of any kind about the post would be listened and well received.


Cheers!!

[Santiago_cabello]

Hi Fernando
Possibly I can help you with second-hand machines (I had the same problem as you)
Where are you from?

[flotus1]

I dont know if worth expend almost 70 e right now for another 8GB (Cost me 47e 18month ago) that would be unused in any other task.

You can still buy 8GB DDR3 SODIMM for ~50€ new. Since you are specifically asking about used hardware options: buy used for even less.
It should reduce solution times with 4 cores active by ~30%.

I mainly see two options for CFD workstations with used parts and a focus on the price/performance ratio.
A) Dual Xeon E5-2670.
Your best option for number-crunching if you have enough parallel licenses. I don't see a real problem with smaller cases. The main disadvantage here is the comparatively poor single-core performance. So some workloads in pre- and postprocessing will take their time. Memory can be bought cheap for less than 25€/8GB.
B) Core-I7 5820k
The better allrounder. Can (and should) be overclocked. Slower than option A when it comes to solving in parallel but faster in pre- and postprocessing. Memory is quite expensive right now.
The I7-6800k is not necessarily the better option since it is harder to overclock and costs more.

I would no longer recommend older CPUs like the I7-3930k. You already mentioned the expensive motherboards (a part which I would rather not buy used if I can avoid it). And you won't get DDR3 memory that is as fast as DDR4 for the same price.

Memory Bandwidth : This seems to reduce the selection for a proper WS to socket 2011 and 2011-3 (4ch memory controller), but lets take into acount latest 2ch CPUs with fast RAM.

You already mentioned the key point here. CPUs with 4 cores and/or 2 memory channels make no proper CFD workstation but rather a normal PC. If the main purpose of the machine is CFD, I would avoid these processors.
It is not always memory bandwidth that limits CFD performance. It is the ratio between CPU performance and memory bandwidth. A quad-core CPU will not be severely bottle-necked by a dual-channel memory interface. But it will nevertheless be slower than a six-core processor with more cache and memory bandwidth.

Any suggestion about what would be a correct ratio between number of cores and memory used for the mesh?

There is no correct cell count/core ratio. The rule of thumb that scaling gets worse with less than 100k cells/core has been around a while now. It is not always true and should not prevent you from using more cores.

A build with 32gb of ram cost arround 1300e. In case a few more gb is needed for a few simulations per year, would get the job done?

You mean "out-of-core" execution with models that require more physical memory than the system has? No. I would never run a CFD job this way. It just takes too long. When it comes to memory the following rules apply in this exact order:
1: get enough memory
2: get fast memory
So if you have to choose better buy 64GB of DDR4-2400 than 32GB of ultra-expensive DDR4-3600.

[Fer_Arus]

Thanks for your answer flotus1, its a really helpful one.

I will get another 8gb for my laptop for the time being, seems like a fair improvement for what its cost.

CFD Workstation:

I am almost decided about the dual xeon, but Ill be using it as a PC too, so I dont know if it is the best option in general terms. Seems like a 5820k would be better choice.

5820k vs 6800k: I see a 12 euro difference, did you mean buying the 5820k used?

Ive seen few used bundles in ebay with 5930k and single xeons 26xx or 16xx (x99 chipset) with 32gb ram for around 600e, would you consider this option?

Thanks in advance.

[flotus1]

Yes, I implied a used I7-5820k for <=300€.
The I7-5930k is not a good option. It costs much more and the only real difference are 40 vs. 28 PCIe lanes which are basically useless in a normal workstation.
A "Xeon 26xx" can be virtually anything . You usually pay more than for a single-CPU with similar stock performance.
Be careful with Xeon 16xx. Up to version v3 they are unlocked. Intel decided to remove this feature for the v4 versions.