Ram Configuration on New Threadripper Pro 8 Channel Platforms

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

ProphecyHope

New Member
Mar 29, 2021
1
0
1
I recently purchased a Threadripper pro processor (3955wx)with 16 core/32 threads. I have learned that it actually has 8 channel memory. I will be using the system for video editing – mostly 4k 10 bit footage with plenty of special effects. I will be using either a Nividia 3060 or 3080ti for my GPU. I fast 4th gen NVME drives to work from. My motherboard is the ASUS Prime TRX40 Pro. Here are my 3 questions:

1. What would be the optimal ram configuration to make themost out of these 8 channels?For example, for 64gb of ram, would it be 32gbx2, 16x4, or8x8?For 128gb, would it be 16x8 or 32x4?

I have noticed that the ram recommended for my motherboard(ASUS Prime TRX40 Pro) has better specs when I install less in my system (forexample, I can find 3600mhz with low latency at 32gbx2, but 32gbx4 or 33gbx8drops to 3200Mhz).

2. Would it be beneficial to sacrifice a bit of MHz (3600down to 3200) in order to populate more of the 8 channels. In other words, with video editing, would I really benefit from having faster ram, or is more ram better? The final part of my question is:

3. Since I am only editing 4k footage, I only really “need”between 32 and 64gb of ram. Would I BENEFIT (faster speed) in any way having 128gb of ram?Especially since processor has 8 channel memory?Any help would be appreciated. Thanks!
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
38
And that 16-core variant may technically have an 8-channel memory controller. But since it very likely only has 2 CCDs active, it will be limited to memory bandwidth closer to quad-channel.
With your use-case in mind, i think a regular TR CPU would have been enough/better.
 
  • Like
Reactions: lkm

bayleyw

Active Member
Jan 8, 2014
302
99
28
To answer the OP's questions (keeping in mind what @nnunn and @alex_stief said, the 3955WX is neither compatible with your board nor can it utilize all 8 channels)

1. To fully take advantage of 8-channel memory, you will need at least 1 DIMM per channel, so at least 8 DIMMs. If you are looking for 64GB, you will need 8x8GB.

2. TR Pro does not support memory faster than 3200 MHz. You should use 8x8GB 3200 MHz ECC DIMMs.

3. If you're not out of RAM, you won't really benefit from more RAM. If you are filling all 64GB, you will benefit from having some extra memory to avoid kicking things out of cache (the actual edits won't be faster, but your system will overall be more responsive). With DDR4 prices being what they are right now, I wouldn't skimp on the RAM - its only a couple hundred bucks extra on a probably $4K+ build to get 128GB.

Following up on what the others said, this doesn't feel like the right setup for video work. Modern video workflows are mostly GPU-accelerated, so what you need is the minimal platform that will host the storage and accelerators you want. You also didn't mention if this is for professional use or for a hobby; the mindset for building an editing rig you are making money with is slightly different since $1K+ for a 10% boost in performance might pay for itself in a few months, whereas you'd be hard-pressed as a hobbyist to pay that much for a marginal increase in throughput.
 

quixotheque

New Member
Feb 6, 2022
9
0
1
And that 16-core variant may technically have an 8-channel memory controller. But since it very likely only has 2 CCDs active, it will be limited to memory bandwidth closer to quad-channel.
With your use-case in mind, i think a regular TR CPU would have been enough/better.
Has this been tested anywhere, because I've heard similar on other occasions?! I'm specifically interested in ways that performance would be affected (negatively or positively) by an 8 channel configuration with use cases like 3d animation, compositing, editing.
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
38
The type of workload you mentioned typically is not memory bandwidth intensive. I.e. zero to very small benefit from more memory channels.
Which also means that TR Pro is not the first choice for this type of work. Depending on other requirements like core count, memory capacity and multi-GPU support, a regular Threadripper platform or even a desktop CPU might be better choices.
 

quixotheque

New Member
Feb 6, 2022
9
0
1
The type of workload you mentioned typically is not memory bandwidth intensive. I.e. zero to very small benefit from more memory channels.
Which also means that TR Pro is not the first choice for this type of work. Depending on other requirements like core count, memory capacity and multi-GPU support, a regular Threadripper platform or even a desktop CPU might be better choices.
Can you elaborate a little? I'm not usually doing simulation work since that can take as much ram as possible, but comp and 3d projects do usually use a lot of ram. Either way I was under the assumption that more channels would benefit with high ram usage.
 

alex_stief

Well-Known Member
May 31, 2016
884
312
63
38
There is not necessarily a correlation between memory usage and memory bandwidth utilization.
2 things need to be present for a memory bandwidth limit:
1) low computational intensity. I.e. for each byte of data fetched from memory, only few calculations need to be done
2) some level of optimization. If the code doesn't utilize the computing capabilities of the CPU somewhat effectively, pressure on the memory subsystem is still low, even if rule 1 is satisfied.

Or to put it differently: You can have an application that requires TB of memory. If that data is not accessed frequently and efficiently, memory bandwidth utilization is still low.
It turned out time and time again, that the -for lack of a better term- "creative" applications do not need tons of memory bandwidth. One of the reasons why regular Threadripper CPUs with only 4 memory channels are somewhat popular in this field.
Engineering applications based on FV or FEA on the other hand do benefit.

I can highly recommend the benchmarks that Puget System puts on their site. They regularly test with all kinds of hardware from mainstream desktop to high-end workstation like TR Pro. And they have a focus on the type of workload you are interested in.
Just one example, they have plenty of these tests for many applications in your field.

But let me emphasize one more that raw performance or price/performance isn't everything. E.g. if you need 1TB of system memory, the choice is much easier, because only few platforms support that.
 
Last edited:
  • Like
Reactions: nnunn

quixotheque

New Member
Feb 6, 2022
9
0
1
It makes a lot of sense but at the same time I want to ask you if you know of a method or a tool to diagnose memory usage. Especially on windows since it is notorious for having bad memory management. I've just ran a profiler in Nuke and it looks like it shouldn't take more than 6Gb but usage in windows is well above 15-20Gb so I'm not sure what to think atm.
But since that is the case (After Effects too, and both of these are manually limited at 20Gb otherwise I am running out of ram) I have to suspect that without limitations I could easily go to 32Gb-48Gb or even more per application so now the question is: is bandwidth relevant here and would 8 channels make a difference (especially if the app is multithreaded -nuke is)