Return to Level1Techs.com

Workhorse build

Hi, I’ve been planning a pc upgrade for the past 3 years and my 5820k/asus x99m-ws setup gave it’s last breath after 6 years of service. I’m second guessing myself over and over if i should go for a full passthrough (full atx size) or a simple dual boot(pocket rocket itx/dtx) setup. I’m reaching out as i’ll probably be in the same dilemma until my ancient dell m4800 dies. I’ll be rocking it professionally for the next 3 to 5 years so my budget will be in between 3000 to 5000 euros/dollars (give or take a few, the cheaper the better :smiley: ).

What will i use it for:

  • Main OS is Linux. I’m an Infrastructure engineer that is currently going all-in on the SRE hype train. So development in Python/Go/Bash(yes, it’s a full blown language :smiley:) , a ton of virtualisation and containers, db benchmarks and stress tests,… I know some of this could be outsourced to a cloud provider but I do like my setup as close/responsive as possible.
  • I sometimes play the occasional game with friends (monthly)
  • 3d modelling (fusion 360 for CNC)
  • Growing interest in machine learning

I am as good as convinced that i’m going the amd 5000 series and get either the 5900x or 5950x. The big struggle are the motherboards/ram combos and gpu’s. I need my ram timings tight and the bandwidth as high as possible (I come from quad channel 3200mhz cl16 setup). I currently don’t care about ECC until i will :stuck_out_tongue: Thunderbolt is always welcome if it’s available. So in general, i need raw horsepower and reliability (16h a day on time and probably 6h a day crunchtime). I included the machine learning part as it’s a growing interest for a personal db tuning project. I know nvidia is currently top with CUDA, I also noticed ROCm getting a lot of love. My preference is going to AMD because of the OpenSource drivers and overall linux compatibility out of the box but if nvidia is just miles ahead, i would consider it. Oooh, nothing beats a silent system!

1 Like

Wouldn’t a HEDT platform make more sense then?
Not sure when AMD is going to launch their new TR platform.
But i assume probablly next year.
Of course HEDT like Threadripper wil be significantly more expensive than mainstream ryzen.
However it does give you more memory bandwith and also more pci-e lanes for expandability.

3 Likes

If the budget is 3-5k euros, a threadripper build might be possible, but it’ll definitely be closer to the costlier side.
I can already imagine this:
TR5970X + 256GB DDR4 3600 CL16, RX6800XT + Decent TRX40 Mobo (I’m thinking the Asrock one, Wendell said it was good with memory tuning).
This is probably around 3500 euros.

I do like the sound of HEDT on the memory bandwidth side but aren’t the cooling solutions on threadripper limited if you want near silent operation? The extra PCIe lanes though :drooling_face: There goes my pocket rocket fantasy :laughing:

Cooling solutions are indeed a bit limited for TR.

Of course a 5950X Ryzen system will also work fine.
And then there is more choice in regards to coolers.
However main stream platforms like Ryzen have only dual channel memory controllers.
Using 4 dimms on Ryzen only increases capacity but not the actual bandwidth.

I’ve bitten the bullet (or however you say this in english XD). Went with the asrock x570 creator, 5950x, 64gb 3600mhz cl16 and a few samsung 980 pro 1TB’s and a Fractal design Define 7 case. Currently i’, still running my 5700xt AE but am planning to either wait and buy the 3080ti or the 6800xt. Not sure which one i’ll go with :smiley:

I’ll probably try out RoCm and see if it currently satisfies my ML urges before i make a decision :slight_smile:

On the cooling part, i’m currently not so happy with the x570 chipset fan and might mod it. The noctua nh-u12a is currently silent enough but i have an old nexXxos 360 monster rad and ek dc pump + supremacy block lying around and will probably switch some time soon to make it even more silent. The waterblock on the GPU might free up some air for the chipset fan and give it some more breathingroom.

1 Like

Cool.

The pair of 980 pro evoS is exciting - 7.2GB/s - WOO!

If the mobo has 2x bifurcatable x8 pcie slots, & 8x pcie 4 lanes is sufficient for ur GPU, then the second x8 slot could host a pair of nvme & yield some powerful raid array options to simulate virtual memory.

The 980s are really nice! I have some integration tests that i run locally that actually spin up a pulsar stack, postgres cluster and a bunch of other services and when i do the performance regression tests, it’s just awfully fast. I automated and reran all previous tagged releases to get my baseline in order and it just went through it in less than 24h. I was planning on running those benchmarks on memory disks but i’ll just leave it as is. It’s fast enough :smiley: The performance itself will probably earn my system back in about a year :octopus:

In the meanwhile i man aged to get a zotac 3090 for msrp by botting :stuck_out_tongue: i’ve let faith decide and started botting for a 6800xt, 3080 or a 3090 and apparently the 3090 won :smiley: The extra cash that i saved by not going with a threadripper is now gone though XD. Running it on manjaro (oh man, pacman -Sy nvidia is so fcking nice!) and it just works. I can remember the gtx 460 days on linux and those were not happy days.

My next adventure will probably be vfio but i first have to get some real work done and get some cash in before i start tinkering with the setup again :stuck_out_tongue:

Have you used Thunderbolt, how well does it work? What can you tell us about the system after a few months?
I’m thinking about building something similar (5950x + x570 Creator), but I’m still on the fence.