[SOLVED] System build advice - 256GB RAM Threadripper

[buggaby]

I run CPU intensive simulations that are quite RAM limited. I need about 30 GB per core. That's why I chose the 16-core Threadripper. Fast per core, but still only need 8 cores to max out the 256GB of RAM. So 16 cores should work perfectly fine.

I will be running these machines headlessly, so only graphics needed would be terminal and initial OS installation.

Reason for not going server-class is that a) no need for ECC ram, which is slower, and b) it's more expensive per compute.

Any other thoughts? Thanks!

PCPartPicker Part List: https://pcpartpicker.com/list/YsZv7X

CPU: AMD Threadripper 2950X 3.5 GHz 16-Core Processor ($799.99 @ Amazon)
CPU Cooler: Noctua NH-U9 TR4-SP3 46.44 CFM CPU Cooler ($79.95 @ Amazon)
Motherboard: ASRock X399D8A-2T ATX sTR4 Motherboard ($464.95 @ Amazon)
Memory: G.Skill Trident Z Neo 256 GB (8 x 32 GB) DDR4-3600 CL18 Memory ($1499.99 @ Amazon)
Storage: Seagate Barracuda Compute 2 TB 3.5" 7200RPM Internal Hard Drive ($52.99 @ Adorama)
Video Card: MSI GeForce GT 710 2 GB Video Card ($76.98 @ Newegg)
Case: Rosewill TYRFING ATX Mid Tower Case ($36.99 @ Amazon)
Power Supply: Gigabyte P GM 750 W 80+ Gold Certified Fully Modular ATX Power Supply ($119.99 @ Amazon)
Total: $3131.83

 

[Herr B]

Ryzen CPU's do support ecc ram by stock eventhough they dont state so.

Further question: Is there any reason to choose the tr 2950x over the ryzen 9 5950x (which im currently running)? Is it because of max memory support?


It might be worthwhile to wait for the next generation tr cpus.

Is there any chance to optimize system memory use by program optimisation (i.e. Memory buffering or data streaming from disk rather than full List/Dataload from/to disk)?

A pcie 4 m.2 ssd would be around 10-20 times faster and might be a huge performance upstep for your data payloads. Be aware of increased wear from such data shoving in that case..

 

[buggaby]

Thanks for the reply!

Ryzen CPU's do support ecc ram by stock eventhough they dont state so.

Further question: Is there any reason to choose the tr 2950x over the ryzen 9 5950x (which im currently running)? Is it because of max memory support?

Yes, that's exactly the thinking. I actually need 3 of these machines, so if I stepped over to a Ryzen line with the 128GB limit, then I'd need 6 boxes. Given that I'm not sure there's a huge improvement in performance or cost, I would hesitate to go that route.

It might be worthwhile to wait for the next generation tr cpus.

Possibly a good idea if I had the time. Project starts soon so need these in the next week or 2.

Is there any chance to optimize system memory use by program optimisation (i.e. Memory buffering or data streaming from disk rather than full List/Dataload from/to disk)?

A pcie 4 m.2 ssd would be around 10-20 times faster and might be a huge performance upstep for your data payloads. Be aware of increased wear from such data shoving in that case..

Are you suggesting to effectively use an m.2 SSD as RAM extension? It's a good question. I haven't done the testing to know if that would work in these circumstances. But again, the opportunity cost to find out is more than I can afford given the time constraints.

EDIT:

Ryzen CPU's do support ecc ram by stock eventhough they dont state so.

As I understand it, Ryzen supported unbuffered ECC, which is much more costly given the limited market. And I don't need ECC at any rate, so not worth the extra burden.

 

[Herr B]

Are you suggesting to effectively use an m.2 SSD as RAM extension? It's a good question. I haven't done the testing to know if that would work in these circumstances. But again, the opportunity cost to find out is more than I can afford given the time constraints.

No that was not the main point. An m.2 should increase your speed in almost any circumstance big time when loading writing data. At this stage I guess, you are handling large datasets. Especially if your dataset is split into small chunks this would make a huge difference. 2 TB nvme drives are available, have one in my desk. but the price might make adifference.

Rest looks fine except I found three things which I might change:
Possibly, this rig will be used for artificial intelligence / machinelearning. If this is the case, I recommend a powerful gpu. this will speed up the learning process by a lot compared to cpu learning. If this is not the Application, forget about it.

The case might have some ventilation issues. If I see it correctly, one single fan is included in the back of the case. 3 further fans are available for the top side but in this case all fans are pulling air from the tiny gaps in the front. I would choose a case with the same amount of intake fans as exit fans and I would mount extra fans. Given that you have a noctua cooler, I would expect the Racks to be in office environment. I'd stick with Noctua in this case for silence in the office and the goodwill of your colleagues

Noctua is a great choice for CPU Coolers. However, this Cooler seems on the low end for me for this CPU and the selected workload.
Given that modern cpu's automatically throttle based on thermal cooling power, Id recommend a nh u14s tr4 sp3 (maybe even with an optional additional 2nd fan). This should increase performance of the cpu by a good margin. Additionally it would be a lot more silent if the rigs are kept in the office environment. Price of this cooler is roughly the same than the referenced one. Keep in mind to check if the height of the cooler fits into the case. (More information on the noctua tr4 2950x compatibility list ) Your case lists a max cpu cooler height of 16 cm (which would not fit the nh u14s either with it's massive 16.5 cm height) List of compatible cases here (pcpartpicker)

 

[buggaby]

Thanks for your contribution!

At this stage I guess, you are handling large datasets. Especially if your dataset is split into small chunks this would make a huge difference.

Very reasonable. However I'm not actually handling big data. I'm running agent-based models, which use some data, but generate their own. There's a lot of computation involved, but not a lot of data analysis until post-processing. And I'm actually outputting fairly aggregate data, so only 100-500 kb of data per hour per core or something. Not much.

Possibly, this rig will be used for artificial intelligence / machinelearning.

Again, not the case here. I have explored the possibility of speed up using GPUs. This is definitely possible with some simulation models, but agent-based models are really event driven with a lot of code branches which make for difficult porting to GPUs.

Your other suggests are great. However, based on feedback from another post, I might be switching to an Epyc build, possibly even a rack. So all the parts I listed will likely be changing entirely.