[Nieuchwytny]

Hello, i had reading older topics but stlil have some questions about hardware. Im bulding low budget PC for small CFD simulation max. 5-7 mln elements in Fluent and in the future in Open foam.
Generally i need help with choose with processor/set have better performance in CFD to value. I know, that two sets have only dual memory chanells, but processors with quad channel are expensive in my country. And stlil these procs are great in daily life.

Amd Ryzen 1700x - 150 euro
Ram DDR4 32 GB 3200 mhz - 200 euro
Motherboard with good power section ~ 125 euro

or

Intel I7 8700k - 300 euro
Ram DDR4 32 GB 3200 mhz - 200 euro
Motherboard with good power section - 150 euro

So, AMD its cheapest, motherboard on X470 will be fit with new 7nm Zen 2 processor, but have worst performance per core.

If you will have suggestions which is better or have other propossition, it will be grat.

Sorry for some mistakes.
Cheers

[flotus1]

Based on your CPU choice, I conclude that you plan on buying used. In that case the I7-5820k is still a valid option. The CPU can be bought cheap, along with X99 boards that many people are trying to get rid of.

[Nieuchwytny]

I thought to buy new Ryzen or used 8700k.
I found used I7 5820K with MSI X99s SLI plus, cost +/- 250 euro, also found i7 6800K for 200 euro + Motherboard 150 euro

In your opinion, 5820K its best choice in that price?

[flotus1]

I think so, if you can find the I7-5820k plus motherboard for around 200€. A quick search on ebay yields the cheapest 5820k for 93€. Some bundles that include motherboard, CPU, RAM and cooler could also be a good deal.
I7-6800k is not really worth more money.
Keep in mind that you will need a graphics card with this build and with a Ryzen CPU, they don't have integrated grapics.

[Nieuchwytny]

Thanks for reply. I have my old gtx 970.

Ok. For quad channel procesor it's better to have a 4x16 GB ram or 8x8 gb?

[Simbelmynë]

Originally Posted by Nieuchwytny

Thanks for reply. I have my old gtx 970.

Ok. For quad channel procesor it's better to have a 4x16 GB ram or 8x8 gb?

For most cases 4x16 is a better option due to stability issues (check QVL of the motherboard). Try to get dual rank memory as well (much easier with 16 GB modules). Depending on the motherboard I would also suggest some degree of overclocking on the memory and CPU to really take advantage of the 5820 k.

[flotus1]

It doesn't make too much of a difference if you put in 8x8GB single-rank or 4x16GB dual-rank. For overclocking headroom (I7-5820k officially supports DDR4-2133, everything above that is overclocking), the amount of ranks per channel matters the most. For best results I would still choose 4x16GB over 8x8GB. It will still be less strain on the memory controller compared to fully populating all slots. And finding out before purchase if 8GB UDIMMs are single- or dual-rank can be difficult.
So far all 16GB UDIMMs are dual-rank. This won't change until 32GB UDIMMs become available.
Another benefit would be that you could expand memory later if you put in 4x16GB. At the expense of probably lower memory clock speed due to 4 ranks per channel. While Haswell-E officially only supports 64GB total, 128GB usually work with later bios versions.

Judging by your prerequisite of 5-7M cells, you would probably be fine with only 32GB total. I.e. 4x8GB.

[Xtreme]

If you suggests 32GB total , then why 4x8GB over 2x16GB ? Just curious

[flotus1]

Because the I7-5820k has a quad-channel memory controller. With only 2 DIMMs, you would only be using 2 of those channels.

[Nieuchwytny]

Hi again. I bought 5820k with Asus x99 Sabertooth, as you recommended. But have some ankward problem with generating mesh and running simulation.

I am running simulations using 6 cores with disabled HT, double precision on fluent 2019R1 with 3 milion elements (Poly-hexacore), and it is take a 20 hours to perform +/- 800 iterations. The same simulation was doing in academic workstation (i7-4790k, 4x8 ddr3), and it take a 10 hour to perform 1000 iterations.

Also have problem with generating mesh, it takes 20-40% more time for the same geometry.

PC spec:
i7 5820k
x99 Sabertooth
2x16 Gb DDR4
1 tb HDD

[Simbelmynë]

1. You did not opt to go with a quad channel memory setup as suggested in this thread.


2. Since the 4790k is faster I would also guess that you have placed the two memory sticks in a configuration that does not even give dual channel.


3. May also be thermal throttling of the CPU due to poor mounting or wrong frequency on the memory sticks.

[Nieuchwytny]

1. I know that, im looking for second ram kit. But actually cant find same model.

2. It's 100% dual channel. With XMP 2400 profile.

3. In 3 hours prime95 test, max. temperature was +/- 72 C. Stock clock.

IDK it will be helpful, but hard disc have very often full usage.

[Simbelmynë]

Then it seems thermals are OK and you should have higher bandwidth compared to the 4790k.


I assume you are on a Windows system. Excessive hard drive usage is never a good thing so this may indeed be your problem. 3 million cells should not fill your RAM so swap should not be an issue here. I also assume that you do not write to file at each time-step or similar.



Is the behavior the same when you are not running calculations, then it may be some other process.


Perhaps you can run some other benchmark to just compare the performance against other 5820k?

[flotus1]

First things first: your system should perform similar to a 4790k. Maybe a little slower thanks to lower stock speed. There are quite a few things that could have gone wrong, in no particular order:

1) Software versions.
The windows version you are running can have an impact on performance. Maybe it is not the same on both systems.
If Ansys marketing is anything to go by, the version number can have a huge impact on performance as well.
Bios/microcode version... all mitigations against hardware vulnerabilities combined can compound to a measurable impact on performance.

2) hardware other than the CPU.
You mentioned disk usage of 100%. This indicates that something is wrong.
Maybe the hard drive is on its last leg and needs to be replaced. Preferably by an SSD, they are cheap these days. You can use tools like CrystalDiskInfo to check
Maybe Windows was still running updates in the background. This would also take a chunk of CPU performance.
Or maybe you had some other tasks running that were using a lot of RAM, causing the system to swap. And eating lots of CPU performance at the same time. Task manager would be your first stop investigating this.

Anyway, you might want to run simpler benchmarks that allow you to compare your scores to other known good systems with similar hardware. Like e.g. Cinebench R15 or the memory benchmark built into Aida64.

[Simbelmynë]

Originally Posted by flotus1

First things first: your system should perform similar to a 4790k. Maybe a little slower thanks to lower stock speed.

I would guess the 5820k is much faster. Is it only single core calculations? Or do you mean due to the current DDR config?

[flotus1]

Yes, the config he bought and is currently running at stock speed.

[chrisro]

Hello for explicit simulations (quasi-static simulations and drop test at rectangular tubes) amd ryzen 9 3900x vs intel i9 9900k or i have to go for CPUs with quad memory channels?

GPU Quadro P1000 4 gb is ok for for these simulations?

Thank you in advance

[Echidna]

With a budget of around 600-700 Euros you can build a solid workstation with dual Xeon E5-2650V2 CPUs (16cores total) and something around 64GBs of DDR3 ECC RAM.
Plus you have the option to upgrade later with 2x E5-2690V2 (20cores) or even 2x E5-2697V2 (24 cores).



This is the path i am following until DDR4 ECC memory reaches reasonable price tags.

[flotus1]

As far as I am concerned, DDR4 memory prices are no longer a reason to stick with DDR3 platforms. We are now at around 4€/GB for UDIMM and below 5€/GB for RDIMM.

Neither of the two CPUs mentioned are a particularly good value for a CFD workstation.
For Intel, the I9-9700k offers basically the same performance at a much lower price. I would even consider the I5-9600k because scaling beyond 6 cores is usually not that great with these CPUs.
Same for AMDs AM4 platform. The CPUs with 6 or 8 cores provide a much better value.
Better spend the money you saved on the CPU for faster/more RAM.

If I had to buy in this price range right now, I would probably still choose Intels X99 platform.

A quadro P1000 would do the trick, but for a non-professional workstation any consumer graphics card with at least 4GB of VRAM will do the same for much less money,

[chrisro]

As far as I am concerned, DDR4 memory prices are no longer a reason to stick with DDR3 platforms. We are now at around 4€/GB for UDIMM and below 5€/GB for RDIMM.

Neither of the two CPUs mentioned are a particularly good value for a CFD workstation.
For Intel, the I9-9700k offers basically the same performance at a much lower price. I would even consider the I5-9600k because scaling beyond 6 cores is usually not that great with these CPUs.
Same for AMDs AM4 platform. The CPUs with 6 or 8 cores provide a much better value.
Better spend the money you saved on the CPU for faster/more RAM.

If I had to buy in this price range right now, I would probably still choose Intels X99 platform.

A quadro P1000 would do the trick, but for a non-professional workstation any consumer graphics card with at least 4GB of VRAM will do the same for much less money,

• I found 32 gb Ram at 230 euros-Corsair Vengeance RGB Pro 32GB(2X16) DDR4-3466MH

• Motherboard Gigabyte Z390 AORUS Pro 195euro

• The Intel I9-9990 at 440 euro

Do you have a specific suggestion around that prices ? I will have problem with dual memory channel CPU and 2x16 DDR4-3466MH?