Question Advice on building a Workstation/ML/gaming rig

[quantleap]

Hi everyone, thanks for considering this thread. I'm new to the forum, and keen to be part of the community.

I'm spec-ing out a build for the following use cases:

Workstation - Big Data Software Engineering ( Top priority )
Ubuntu Linux / Kubernetes / Apache Spark / Apache Kafka / Scala

Machine Learning / Data Science (Research)
Tensorflow / PyTorch / Jupyter - Mostly Forecast / NLP / Deep Learning

Gaming ( Non Competitive)
Windows 10 Pro Dual Boot / 4k Gaming / (CyberPunk 2077 / Tomb Raider / Civilization VI / DOOM Eternal )


With that in mind, my current target:

CPU: Ryzen 3700X 8 C / 16 T 3.6/4.3 + Wraith Prism // No Overclocking intended
£350
Memory: Corsair Vengeance RGB PRO 64 GB (4 x 16 GB) DDR4 3200 MHz C16 XMP 2.0
£350
Motherboard: X470 AORUS ULTRA GAMING, AMD X470, DDR4, Dual M.2, 2-Way SLi/CrossFire, GbE, USB3.1 Gen2
£125

Storage:
1x 1TB NVMe Samung 970 Evo Plus
£220

1x 4TB Seagate Barracuda
£88

Case: Corsair Carbide Spec-06 w/ 2x 120mm Fans
£85
PSU: Seasonic Focus Plus+ 850 Watt, Full Modular, 80PLUS Gold, SLI/CrossFire, Single Rail, 70A, 120mm Fan
£90

-----------------------------------------------
£1300
-----------------------------------------------

GPU: Nvidia RTX 2080 (Asus Turbo 8GB) - Blower, single fan.
£730

-----------------------------------------------
£2030
-----------------------------------------------


Important notes/ doubts:

PCI4:
This setup ignores PCI 4, which there's very little detail on parts. It seems to me that in 2+ years, PCI 6.0 should be out, and PCI 4.0 and 5.0 would be as short lived as 1.0 and 2.0. By the time NVMes/GPUs are ready to take advantage of it, and price comes down, you'll need a whole set of new kit anyways.

Motherboard:
Bios upgrade doesn't seem to be available for the mentioned motherboard. I could upgrade it to X570 of the same model if the price isn't ridiculous. Big issue here is that it needs to be able to plug the latest Ryzen.

CPU:

More cores would be amazing, but for a lot of distributed applications, it sounds a reasonable price / performance ratio particularly when paired to RAM. It'll offer 4GB per vCore, so great for a large, self contained cluster.

GPU:

8GB is probably not a lot of GPU RAM, but it'll offer enough to get some interesting work done.
Leaving room for a potential second GPU, in case one is needed, or at least headroom for efficient use of the existing GPU.
Besides, games rarely support SLI correctly, and ML engines tend to struggle, often delivering about 80% of the combined power.
Given that we are one year in on the RTX 20xx generation, which is only the first product of a new line. Because of that it doesn't seem like a great idea to spend too much right now, so 2080Ti avoided. Probably in 2 years time, a much more powerful unit with a lot more Tensor Cores would trounce the 2080, which i could still keep, assuming i could pair different gens.
The single fan GPU was based on recommendations that dual fans just don't dissipate heat well when paired.

I've not built a PC in decades, so any advice on cooling, and anything to optimize the build would be greatly appreciated. Anything from PSU, case, you name it. I've joined the forum explicitly to get your feedback.

Here's the bill of materials (several key prices missing since items have not been released yet):
https://pcpartpicker.com/list/sFDpfH

Budget can be stretched, most important thing is to get a great system, but fit for purpose..

 

[King Dranzer]

Going for RTX2080Ti over RTX2070 Super is 35%+ improvement in performance for overall 20% increase in budget which is not bad at all.

For RAM I recommend going for 2 x 32GB over 4 x 16GB as if you want you can add 2 more 32GB sticks in future to total 128GB RAM instead of being limited to 64GB if required.

 

[Karadjgne]

So speaks the budget. Personally I never advocate adding ram. Ever. You want 128Gb? Buy 128Gb, sell off the old 64Gb.

1 stick is easy, it works or it doesnt. 2 sticks is iffy, factory guarantees they work. 4 stick? Good luck. Which stick doesn't play? One of the new ones? One of the old? So you return the kit for compatibility issues. Get another. And another. And another. Oh, bought online? Weeks later you might get all 4 sticks that play well, and sit well with that Ryzen.

Forget that. 1 kit, factory tested and guaranteed.

 

[quantleap]

Comparing against professional builds, this looks to be a great value:

https://www.pugetsystems.com/configure.php

ThreadRipper 2920x
64BG RAM
1 TB 970 EVo Plus
2 TB 970 Evo Plus
Nvidia 2080
Fractal Design R6
EVGA 850W

$4917.77

I'd be getting equivalent or better performance for half the price.

Why not 2080Ti ?

I'd expect to replace this card in 2-3 years time max. In fact, Ampere would be produced on 7nm, so since NVidia is holding their best cards for later, i'll do the same.

Buying the 2nd biggest GPU in the market (2080 and 2080 super are only marginally better) at this price feels correct.

From a gaming perspective, let's not forget that even DX12 and Vulkan still are been adopted by developers. Similarly with RTX, which only a handful of games support yet.

By the time that these hardware is put to serious use, a simple GPU upgrade will do the trick.

AnandTech scores:

https://www.anandtech.com/show/14586/geforce-rtx-2070-super-rtx-2060-super-review/4
2070 Super
SOTR: 42.6, 99th: 29.3
F1 2019 (Ultra, DX12): 59.9
ACO (Very High, DX11): 46.2,
Metro: Exodus (Ultra, DX12): 35, 99th: 22.8
Strange Brigade (Ultra, Vulkan): 75, 99th: 66.2
Total War: Three Kingdoms (Ultra, DX11) : 30
The Division 2 (Ultra, DX12): 40.2, 99th: 33.0
Forza Horizon 2 (Max, 4xMSAA, DX12): 58

GamersNexus:
GN: (AVG, 1% LOW, 0.1% LOW)

Sniper Elite 4: (High, DX12): 79.3
Strange Brigade (Ultra, Vulkan): 76.7, 68.3, 67.3
GTA V (Ultra, DX11): 57.3, 41.2, 30.6
SOTR (High, DX12): 49.9, 45.4, 44.4
Hitman 2 (Custom Ultra/High, DX12): 49.7, 20.6, 12.8

Regarding Tensorflow and other compute engines. I had several chats with data scientists, and most of them don't work with images, which are heavy memory bandwidth consumers. For most other ML, even a 2070 Super is more than enough. People used to win Kaggle competitions with substantially inferior hardware only 2-3 years ago, so it'll definitely be enough for me at this stage.

In particular, there was a rush to create models with ever more complex networks, in the hope that they'll become magically better. Often, very deep networks only show marginal improvements beyond 128 batch sizes, which is typically at the point where you need more than 8 GB RAM. Whilst it's true that bigger hardware will make things easier, it doesn't seem entirely justified right now. Even more the case with multi-GPU / SLI / NVLink use cases.

I'm still considering the Blower style card, even though it's unlikely it'll get used on a multi GPU setup. I'm yet to find a single review of this card. Any experiences to share regarding GPU implementation?




Regarding RAM, these are VERY fast modules, and Ryzen should be able to take advantage of it. But since most of the use case is many large processes as opposed to one single giant processes gobbling 64GB of RAM, i'm not too worried about memory performance. There's a nice video by Jay2cents where he points out that he'd rarely be able to tell the difference on memory performance. This is specially true when working with Java/Scala/JVM applications, where garbage colleciton will play a far larger role. Worse, many of the applications will run within containers, and not on bare metal, introducing substantial latency. And that's OK for my use case.

Finally, an interesting note. As detailed on Phoronix website, Ryzen does not even boot on Ubuntu 19.04 or other distros using latest Systemd.
https://www.phoronix.com/scan.php?page=article&item=ryzen-3700x-3900x-linux&num=2

Older Ubuntu 18.04 LTS is fine, which suggests that the latest and greatest software still needs to catch up.

 

[quantleap]

I've decided to change the GPU implementation, to an MSI Gaming X Trio, since i found great reviews (basically, 3 editors choice)

https://www.guru3d.com/articles-pages/msi-geforce-rtx-2070-super-gaming-x-trio-review,24.html

MSI GeForce RTX 2070 SUPER GAMING X TRIO Review

MSI's new GeForce RTX 2070 SUPER when overclocked battles the higher-end RTX 2080, all whisper quiet and running cool.

www.tweaktown.com

https://www.techpowerup.com/review/msi-geforce-rtx-2070-super-gaming-x-trio/34.html

https://us.msi.com/Graphics-card/GeForce-RTX-2070-Super-GAMING-X-TRIO/Specification

Temperatures normally hit a max of 66 degrees, and not particularly noisy.

It seems however, that getting hold of a R9 3900X is quite difficult these days, demand is quite high, which is unsurprising.

hopefully, i should be able to get the full system end of July/ beginning of August.