[Shayor]

Will not Ryzen 9 3900X(12 cores!!) and 2X16GB 3200 RAM be a better choice than Threadripper 1900X(8 cores) and 2X16 GB 3200 RAM?I am going to buy it next month.My simulations are pretty big and complex actually.Min 20 million cells with FSI and complicated boundary 3D Geometry(blood vessel) and pulsatile inlet conditions will be present.

[evcelica]

No, The Threadripper would be much better as it has 4 memory channels, where the Ryzen only has 2. You will bottleneck on memory bandwidth before you even get to use half those 12 cores.
CFD is all very sensitive memory bandwidth.
There are other AMD EPYC processors with 8 memory channels. Be sure that the one you pick can actually use all 8, some do not. Compare maximum memory bandwidth numbers.

[Shayor]

But the problem is Server Class and HEDT hardware is super expensive here in my country.If I overclock those memory modules won't it narrow the gap and be able to use atleast 8 cores effectively?I would want to use my PC for other works while only 8 cores are busy and 4 cores are still pretty much half-used for web browsing(work related),word,etc hence avoiding the PC becoming "motionless"....................Please help me out.I am appreciating your help a lot.

[flotus1]

Well, there is always aliexpress for those regions of the world where new hardware is prohibitively expensive, or not available at all.
My recommendation these days would be two Xeon E5-2678 v3. Those are 12-core CPUs with quad-channel DDR4 (and DDR3) memory controllers for around 100$.
Motherboards for these CPUs can be found with the search term "dual X99" on aliexpress. For example: https://www.aliexpress.com/item/4000983688769.html
Add 8x8GB or 8x16GB DDR4-2133 reg ECC to the mix, and the result is a pretty neat workstation at an affordable price. 8 DIMMs total is important here, that's the whole point of these builds for CFD and FEA.

Beware of shipping times though. If you are really in a hurry, try to search for listings that offer faster shipping methods to your country. And maybe search for one of those bundles with board+2CPU+RAM. This way, you decrease the risk of one component arriving too late.

[Shayor]

But other than Non-Brand Chinese ones the hardwares are unimaginably expensive.I would have no money at all to buy other parts if I buy only a Mobo and CPUs.That's what I found after following your advice sir and checking things out on Ali Express.

[flotus1]

I'm confused. How much did you intend to pay for a Ryzen 9 3900X alone?
Going by the regular listings I know, it is about the same as two of these Xeon CPUs, including a motherboard. And you have the benefit of cheaper memory, which you will need lots of.

[Simbelmynë]

Originally Posted by evcelica

No, The Threadripper would be much better as it has 4 memory channels, where the Ryzen only has 2. You will bottleneck on memory bandwidth before you even get to use half those 12 cores.
CFD is all very sensitive memory bandwidth.
There are other AMD EPYC processors with 8 memory channels. Be sure that the one you pick can actually use all 8, some do not. Compare maximum memory bandwidth numbers.

It is not much faster though. Assuming you can get 3200 MHz rank 2 memory to work on the 1900X then it is about 15% faster than a 3700X with rank 1 memory clocked at 3600 MHz (data from the OpenFOAM benchmark thread). This assumes you can use full bandwidth. For meshing and other tasks the 3700X would be faster (and 3900X would be faster still). Also, I suspect that Zen3 will bring some further improvements this year so I would not recommend going with a gen 1 Threadripper today. It is probably better to pick a 3600, a X570 motherboard and good memory (low latency high frequency) and prepare for a zen 3 replacement if you are on a budget.



Epyc on the other hand, well that is clearly a better choice, at least in terms of performance.

[Shayor]

@flotus1
Sir,I have checked them on Ali-Express and apart from super old supermicro cards with zero reviews of the seller the Asus or such companies having just one socket costs around 30000-36000 BDT while old Xeons about 8k BDT or if Octa core E5 2650 is considered then 4.5K approx.Then again these boards support quad channel not octa channel.So it would cost a lot more with all the risks involved in it not working and failing in which case I might have to drop out of college that year as not another penny can be spent and my thesis work and another independent paper publication dreams(already drafts are ready ,will start simulating with the new machine) would end.That's the problem I am facing here.

@Simbelmyne

Sir if I overclock the 3200MHz memory to 3600Mhz and adjust timings will the performance not reach (of the R9 3900X) very near to old 2.4-2.5 Ghz Xeons(old and weakened through use) and also can't I use all 12 cores and see improvements?I mean will those 12 cores amount to nothing if I am talking about 12-15 days simulation(12 hrs a day) time compared to an 8 core and 6 core?

[Simbelmynë]

Your question is impossible to answer. I'm not sure what CPU you wish to compare against. "Old and weakened through use" is not something that you will experience. Look at the benchmarking thread and then make your decision based on the price you are willing to pay. The general rules of thumb of CFD are (in order of importance)



1. What license do you have for your CFD software? This dictates what type of computer you purchase.



2. How large simulations will you run? You need sufficient amount of RAM to be able to run the case.


3. Memory bandwidth is almost always a bottleneck for CFD. This means you wish to have a computer with as much bandwidth as possible. Memory channels and memory speeds are most important. Memory timings can be important on some systems. Dual rank memory is always preferable to single rank memory since it gives about 10% higher performance in CFD loads.


4. You need sufficient number of cores to perform the calculations. This is dependent on the CPU architecture as well as CPU frequency. For a given architecture higher frequency is always better, but you cannot compare across architectures and draw the same conclusion. With modern CPUs you usually see a large drop in parallel scaling after two cores per memory channel. If you have a dual channel memory setup then simulations stop scaling well after 4 cores, but you can get some benefit up to about 6 cores. This means you should always have at least two cores per memory channel, but you should try to get three cores. Going higher may increase performance slightly. Note that higher core count usually decreases the frequency which leads to significantly worse performance in single threaded tasks.

[Shayor]

Sir,

1. I am going to use Code Saturne as CFD software,however I am going to use Ansys for checking my simulation in Code Saturne as new to me compared to ANSYS.

2.I mentioned it in my post at the beginning about it

3.I wanted to use dual rank CPUs(mainstream CPUs) not server grade ones as they are irritatingly expensive here and also very rare.

4.I wanted to use R9 3900x and as suggested by flotus1 here wanted to know how it stacked up against it( memory channels differeing a lot.......dual to quad) and I have mentioned it in my last comment sir.I justed wanted to know if I would be able to get a say 5% or so less say but 95 % of the performance if I went with Ryzen 9 3900x with 3200Mhz(overclocked to 3600 Mhz) instead of going dual 6 core Xeons(2.3 Ghz) and 1600 Mhz ddr3 memory setup with say 32GB of memory in each case for now.How much of a difference in performance would be there?

[Simbelmynë]

Originally Posted by Shayor

Sir,

1. I am going to use Code Saturne as CFD software,however I am going to use Ansys for checking my simulation in Code Saturne as new to me compared to ANSYS.

2.I mentioned it in my post at the beginning about it

3.I wanted to use dual rank CPUs(mainstream CPUs) not server grade ones as they are irritatingly expensive here and also very rare.

4.I wanted to use R9 3900x and as suggested by flotus1 here wanted to know how it stacked up against it( memory channels differeing a lot.......dual to quad) and I have mentioned it in my last comment sir.I justed wanted to know if I would be able to get a say 5% or so less say but 95 % of the performance if I went with Ryzen 9 3900x with 3200Mhz(overclocked to 3600 Mhz) instead of going dual 6 core Xeons(2.3 Ghz) and 1600 Mhz ddr3 memory setup with say 32GB of memory in each case for now.How much of a difference in performance would be there?

1. So if you are going to use ANSYS then you have an ANSYS license and this license dictates how many cores you can use.


3. I guess you mean dual channel CPUs? As flotus1 pointed out you can get used server CPUs quite cheap. If this is not an option then there are still better options than the 3900X.


4. It is impossible to answer since you have not said anything about which CPU model you are talking about. Look at the benchmark thread to get the best current information.


Finally after looking at your replies, I do not recommend you to do any overclocking. Overclocking is never guaranteed and good memory is also rather expensive.