[Kosmosisy]

Hello all.

I'm upgrading my system. What is the most important criteria about selecting ram. For example, 4x4gb 2400 mhz and 4x8gb 1600 mhz and ddr3/ddr4. In this case,is quantity or the frequency important case? I did not find any satisfactory results about ram comparison.

[flotus1]

If you are upgrading your system, DDR3 or DDR4 is not really the question. Use whatever is compatible with your CPU and motherboard.
The amount of RAM is determined by the size of the simulations you want to run. You have to make sure that your models fit into the RAM. More RAM than necessary does not increase the speed of your simulations. You can just run larger simulations with larger meshes and more complicated models. It is up to you to decide how much RAM you need. However, I would not recommend DIMMs smaller than 8GB for a decent workstation. The minimum number of DIMMs you need depends on the CPU/number of memory channels.
Choosing RAM frequency is simple. Use the fastest compatible modules you can afford. Simulation speed usually scales pretty good with RAM frequency and latency.

[Kosmosisy]

Originally Posted by flotus1

If you are upgrading your system, DDR3 or DDR4 is not really the question. Use whatever is compatible with your CPU and motherboard.
The amount of RAM is determined by the size of the simulations you want to run. You have to make sure that your models fit into the RAM. More RAM than necessary does not increase the speed of your simulations. You can just run larger simulations with larger meshes and more complicated models. It is up to you to decide how much RAM you need. However, I would not recommend DIMMs smaller than 8GB for a decent workstation. The minimum number of DIMMs you need depends on the CPU/number of memory channels.
Choosing RAM frequency is simple. Use the fastest compatible modules you can afford. Simulation speed usually scales pretty good with RAM frequency and latency.

Thank you for replying. May I ask another question? How motherboard important in cfd?

[flotus1]

The motherboard itself (provided all other components are the same) has little or no impact on performance for computations. However, you should not cheap out on this component especially if you are buying consumer-grade hardware. This can have a negative impact on system stability since cheap consumer boards are not designed to operate 24/7 for several years.

[Kosmosisy]

Originally Posted by flotus1

The motherboard itself (provided all other components are the same) has little or no impact on performance for computations. However, you should not cheap out on this component especially if you are buying consumer-grade hardware. This can have a negative impact on system stability since cheap consumer boards are not designed to operate 24/7 for several years.

There are two options in my mind. First, system with i7 5820k and the second option is system with i7 4930k. There are some benchmark tests. Old ddr3 system owns the 5820k even with lower ram frequencies.

>> http://www.anandtech.com/show/8426/t...5820k-tested/5 3D particle movement

>> http://www.overclockers.com/forums/s...20K-comparison AIDA64 tests

>> https://www.technopat.net/sosyal/kon...uclari.254064/ again aida64 test

For last link, first list(Bellek Okuma) means read, second list(Bellek Yazma) means write and then third(Bellek Kopyalama) means copy and the forth means latency.

Here is little background story which I pmed some other users;

I'm an undergrad student and doing my senior thesis. I'm simulating pem fuel cell in fluent. My system has 2x8 GB of ram and fx 8320 for the cpu. Generally, one simulation needs 2 or 3 days to convergence, so that necessarily I always have to go to the university and simulate the cases. However this is really tiring. I decided to upgrade the cpu and the ram. I'll sell my 2x8 GB of ram and buy 4x4 GB. I look the memory allocate table, it is way more lower than 16 gb.(Maximum element number was 2.1 m) For fuel cells a single case "generally" needs at least 20-25k iteration. Me and my advisor decided to make a good paper about my senior project and I want to publish it before my graduation. My advisor is really tough person, he wants at least six or seven grid indepence tests. Parallel process licensing isn't a problem.

For one iteration, how much mb of information flows through? I mean, for example lets say; in one iteration 100 gb/s of information flows through so there is no reason to have a higher frequencies of ram or higher memory bandwith.(For these tests) IF amount of information is 100 gb/s , the processor's clock speed will more significant? Or am I totaly wrong?

[flotus1]

Some of the better benchmark scores in your first link for the older processor are simply due to the fact that it has a slightly higher stock frequency. I dont think that stock frequency is a relevant parameter for a processor that can easily be overclocked.

For the sequential memory bandwidth benchmarks: This is indeed surprising and I found other sources that confirm the findings that DDR3 can yield higher sequential bandwidth (although they tried their best to cover this fact).
But to put this in perspective: you will easily find DDR4-3200MHz at a reasonable price. For the same money, you only get DDR3-2400MHz.

Hoverver, although sequential memory bandwidth is important, it is not everything. The larger L3-Cache of the I7-5820K is a relevant feature.
Concerning your thought experiment with the "information flow": unfortunately, it is not that easy. If all memory access was sequential, your conclusion would be valid. But RAM stands for random access memory and that is what memory access often is: random reads and writes of smaller chunks of data where sequential speeds are irrelevant and latency (and cache sizes to avoid this type of access) is also important.

In conclusion: I would still put my money on the newer platform.

Edit: Digging a bit deeper I think I found the reason why in your second link the memory bandwidth (and the L3-cache bandwidth and latency) were so much worse with the I7-5820k and DDR4. Although they did their best to produce similar conditions on both platforms, they missed one key parameter: uncore frequency. This is the relevant frequency for the memory controller and the L3-Cache. In Aida64 this is called NB frequency.
On the X79 platform this frequency seems to be linked to the CPU clock speed by default (still checking on this) but at least its default value equals the stock CPU frequency. The uncore frequency used in the benchmark was obviously 4200MHz for the I7-4930k. The default uncore frequency for the I7-5820k is only 3000MHz and it has to be overclocked separately. Since the screenshot taken shows 1200MHz (default idle value) it appears safe to say that the benchmark was run with an uncore frequency of only 3000MHz. This 40% difference perfectly matches the 41% difference in L3 cache latency. So with proper overclocking (not forgetting to increase the uncore frequency aswell) there is a high chance to get the benefit one would expect from DDR4 memory and a more recent CPU architecture.
http://us.hardware.info/reviews/5775...locking-uncore