|
[Sponsors] |
February 1, 2022, 13:32 |
|
#21 | |
Member
Guy
Join Date: Jun 2019
Posts: 39
Rep Power: 6 |
Quote:
You might find this video very interesting. https://www.youtube.com/watch?v=kEP8iVVQR-g The benchmark that showed the greatest memory speed sensitivity was WinRAR. I've attached screenshots comparing the speed difference due to rank, timing and memory speed. |
||
February 1, 2022, 13:37 |
|
#22 |
Member
Guy
Join Date: Jun 2019
Posts: 39
Rep Power: 6 |
Here is an interesting memory benchmark program.https://www.cs.virginia.edu/stream/
It appears to be the one used by AMD to benchmark their systems. See pages 7 and 8 here: https://prace-ri.eu/wp-content/uploa...-Guide_AMD.pdf |
|
February 1, 2022, 13:57 |
|
#23 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,400
Rep Power: 47 |
Stream is the de-facto standard for measuring memory bandwidth.
Numbers published by AMD themselves use a modified version though. As far as I am aware, they still have not disclosed which changes they made, let alone published their source code. So don't expect to get close to the theoretical limits with the default version of this benchmark. Results using Intel compilers with streaming stores enabled can get pretty close though. |
|
February 1, 2022, 13:59 |
|
#24 |
Member
Guy
Join Date: Jun 2019
Posts: 39
Rep Power: 6 |
I'll build Stream and test my 3900X and 7601 when I get time. I'll post results here.
|
|
February 1, 2022, 14:10 |
|
#25 | |
Senior Member
Join Date: May 2012
Posts: 548
Rep Power: 15 |
Quote:
Not sure why I should look at that video. It does not deal with memory sub-timings for single-rank vs dual-rank? Also, Winrar is not something I would use to benchmark CFD. If you have dual rank memory then you can try the settings I used in this post. I doubt it will work with your memory though. If you have GSkill Royal Z or similar then you might be able to push cl14 with a dual rank kit. Loaded latency is also interesting in some cases. https://www.intel.com/content/www/us...y-checker.html |
||
February 1, 2022, 14:27 |
|
#26 |
Member
Guy
Join Date: Jun 2019
Posts: 39
Rep Power: 6 |
As explained in the video link I provided, the memory controller and map was completely reworked between Zen 2 (Ryzen 3000) and Zen 3 (Ryzen 5000).
While the results you give in your thread are valid for your Ryzen 3000 part, they are not valid for Ryzen 5000 parts. WinRAR's performance is highly dependent on memory speed in comparison to processor speed. I agree that a CFD benchmark would be better, but absent that, WinRAR will have to do. I'm always happy to test things. Here is my RAM: Code:
Memory Device Array Handle: 0x0032 Error Information Handle: 0x0042 Total Width: 64 bits Data Width: 64 bits Size: 16 GB Form Factor: DIMM Set: None Locator: DIMM_B2 Bank Locator: BANK 3 Type: DDR4 Type Detail: Synchronous Unbuffered (Unregistered) Speed: 2133 MT/s Manufacturer: G-Skill Serial Number: 00000000 Asset Tag: Not Specified Part Number: F4-3600C17-16GTZ Rank: 2 Configured Memory Speed: 2133 MT/s Minimum Voltage: 1.2 V Maximum Voltage: 1.2 V Configured Voltage: 1.2 V Memory Technology: DRAM Memory Operating Mode Capability: Volatile memory Firmware Version: Unknown Module Manufacturer ID: Bank 5, Hex 0xCD Module Product ID: Unknown Memory Subsystem Controller Manufacturer ID: Unknown Memory Subsystem Controller Product ID: Unknown Non-Volatile Size: None Volatile Size: 16 GB Cache Size: None Logical Size: None This memory is 3600MT/s. I'm running it at 2133MT/s to test the benchmark speed difference. I don't have access to the DRAM Calculator in Linux. If you give me some settings, I'll test them and report back. BTW, the BIOS gets memory timings from the JEDEC table that the manufacturer provides on the RAM. Neither the motherboard nor the BIOS make these up. |
|
February 1, 2022, 14:53 |
|
#27 | |||
Senior Member
Join Date: May 2012
Posts: 548
Rep Power: 15 |
Quote:
Sorry I don't see your point. The memory sub-timings can be adjusted with Ryzen 5000 and while I do not have access to a Ryzen 5000 - you have. You can easily test the Ryzen DRAM calculator settings for Rank 1 and Rank 2 and try to tighten your timings as much as possible. I doubt you can use Rank 1 timings if you have Rank 2 memory, but this is not something I have tested. But if you read motherboard QVL lists it is apparent that the number of qualified kits with dual rank memory above DDR4 3600 are quite rare. When you try to improve your performance on the memory you will see that you will hit a wall in frequency as well as in timings. You will - in general - be able to have lower timings at lower frequencies, but this is not a 1:1 relationship for the memory controller. Long story short - you can pay extra for either bandwidth or memory latency. If you want both then you will have to pay a huge amount extra and prepare to run the memory at maximum voltage, while crossing your fingers that your CPU has a good memory controller and that the motherboard has several layers of good PCB lines etc. Quote:
I think this type of reasoning is completely wrong. You are of course free to base judgment on whatever you like. Edit: Quote:
You can try the settings I used, I posted screenshots of them, I am also on Linux but I run the DRAM Calculator on an old laptop. If you have a decent motherboard then you should be able to change all primary, secondary and tertiary timings of your memory. You should try to figure out what type of memory you have as well. I have the old golden standard, i.e. Samsung b-die. If your advertised memory speed is 3600@cl18 then you most likely have budget memory. If you manage a good memory - timing - overclock then you have "fooled the system" in that case. It is unlikely though now that the memories are so aggressively binned. Anyways good luck! Last edited by Simbelmynė; February 1, 2022 at 15:05. Reason: linuxguy added some extra information in his post |
||||
February 1, 2022, 15:05 |
|
#28 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,400
Rep Power: 47 |
What are we really doing here? Is this still about rank interleaving?
So your hypothesis is that because Zen3 doesn't use a feature called "interwoven memory" [sic], it will not benefit from rank interleaving. Thus getting the same memory bandwidth with one and 2 ranks per channel, assuming identical transfer rate? |
|
February 1, 2022, 16:40 |
|
#29 | |
Member
Guy
Join Date: Jun 2019
Posts: 39
Rep Power: 6 |
Quote:
This discussion has nothing to do with memory rank. |
||
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
General recommendations for CFD hardware [WIP] | flotus1 | Hardware | 18 | February 29, 2024 12:48 |
Used Memory Accumulates During Course of Simulation Until interFoam gets Killed | Ship Designer | OpenFOAM Running, Solving & CFD | 7 | October 6, 2023 01:26 |
4-core Workstation Builds | dominicafonso | Hardware | 9 | April 11, 2021 06:42 |
Epyc 7551 vs 6850K; Ansys Mechanical Bench | Duke711 | Hardware | 24 | March 26, 2020 10:16 |
AMD Epyc CFD benchmarks with Ansys Fluent | flotus1 | Hardware | 55 | November 12, 2018 05:33 |