Build your deep learning box: wiki thread


(Francisco Ingham) #102

Just wanted to share a picture of my build.

Zero experience in building and now I have this beauty. If any of you is doubting whether to build or not I would argue strongly for you to build your own box if you are serious about DL.

Reasons:

  1. Skin in the game
    Investing a substantial amount of money in your new build proves to yourself that you are serious about this DL stuff.
  2. Motivation
    Finishing the build, with the hassle it entails to choose the parts, source them and get everything to work smoothly (both hardware and software) is HIGHLY satisfactory.
  3. Flexibility
    The fact that you worked this out for yourself from the ground up means you know each step you took and can easily change if necessary (e.g. change a computer part for a new one or modify your dev setup).
  4. Ease of work
    If you can afford good hardware, everything is easier than relying on external compute. Running a notebook is smooth and fast, no lags. This facilitates experimentation enormously.


(Jason Patnick) #103

i just started setting up my box, so i basically set up the fastai environment and thats pretty much it. right now i have 114/223 GBs used on my ssd and 0 used on my hdd. which seems like theres a ton of stuff on the ssd that doesnt need to be. the ssd is setup as the C: drive and the hdd is the D:. is this how it should be? or should the drives be switched around and just save the ssd for data in projects? or something else?


(Joseph Woodson) #104

Here’s my personal box I got for gaming (with DL in mind when I purchased it): https://pcpartpicker.com/list/hHB2JV

I do wish I had gone for a 1GB SSD (and perhaps a 960 Pro for the speed), and nowadays I’d recommend an 8700K instead of Ryzen (and perhaps slightly cheaper RAM because Intel processors tend to not be as hungry for low memory latency), but overall I’m happy with my purchase. Specifically note the WD Gold – it’s just as fast (and expensive) as WD’s better-known Black gaming drives, but is made for more resilience in a heavy-use environment such as a datacenter.

Here’s a picture (I assure you all, I have done a bit of cable management since I took it):


#105

Would love some feedback on my build.

https://au.pcpartpicker.com/list/Fq6VbX (edited for PSU)

Aiming for 2-3 GPUs within a reasonable budget.


(Sahil Gupta) #106

I think Ant PC has dedicated deep learning workstations which is quite superior than these see this link, They customize as well:

https://www.ant-pc.com/home/ai_and_deep_learning


#107

(Edit: realised Sahil you were responding to a previous post but someone might still find this useful)

Thanks for the link, I’m deciding to build it my own to make it easier to upgrade later.

The ant pc link has a good reference for a dual 1080 Ti build, it is actually quite similar to mine with the exception of ninth Gen Intel CPU (i9 7900X with 10 cores). In Australia , that’s at least x3 the price of a i7-7800X. I agree it’s superior but doesn’t fit the budget since I only think I need 6 cores (I’d prefer to spend it on the third GPU!)


#108

Has anyone looked seriously at the new mac min with eGPU? Any thoughts on eGPU setup, am reading but it seems like the eGPU talk for DL has sort of died down recently


(Pawel) #109

Could you have a look and give come comments for this build?

CPU
Intel - Core i7-8700K 3.7 GHz 6-Core Processor $358.89

CPU Cooler
be quiet! - Dark Rock 3 67.8 CFM Fluid Dynamic Bearing CPU Cooler $61.79

Motherboard
EVGA - Z370 Classified K ATX LGA1151 Motherboard $179.00

Memory
G.Skill - Ripjaws V Series 32 GB (2 x 16 GB) DDR4-3200 Memory $226.99

Storage
Samsung - 970 Evo 500 GB M.2-2280 Solid State Drive $100.00

Video Card
NVIDIA - GeForce RTX 2080 8 GB Founders Edition Video Card $879.95

Case
Thermaltake - Core P3 ATX Mid Tower Case $120.00

Power Supply
EVGA - SuperNOVA G3 750 W 80+ Gold Certified Fully-Modular ATX Power Supply $95.88


(adrian) #110

Looks good, but psu may be a bit underdone?, something like a750w/800w would give you more of a buffer.


#111

I’d go with the RTX 2070 over the 2080 and put the savings towards a larger SSD and power supply. I have a 2080 but would opt for the 2070 if I had a do-over.

Or actually, EVGA has a bundle of a 2080 and a 1000W power supply cheaper than your FE 2080. There were some 2070 bundles even cheaper, but those are sold out.

https://www.evga.com/products/featured-bundles.aspx


(Pawel) #112

Thank you for the advice. Why I might need bigger power supply? PCpartpicker estimates 360W for this setup.


#113

EVGA’s power tool also has lower than expected wattage for your build, but NVIDIA recommends a 650w for the 2080. It may be the difference between the peak draw and having some headroom for stability.


(adrian) #114

Following on from @Ralph bigger psu doesnt necessarily main more power use, just means psu can supply that power. Also psu efficiency curves from what ive seen (and each psu is gonna be different) have better efficiency at ~<90% or 80% load (cant exactly recall).
Also may want to consider if you want to get a second gpu at some stage may want psu to hadle 2 :smiley:


(Pawel) #115

Thanks. So I have changed storage for bigger & faster and as suggested power supply for 750W. And what do you thing about CPU is it worth going with i7-9700K?


(adrian) #116

Benchmarks ive seen incl below say 5-10 pc better sincle thread and 10-20% multithread performance (even with fewer threads) for the 9700k. httpsrowser.geekbench.com/processor-benchmarks. But probably not going to notice unless youre doing a lot of single thread work/heaps of data munging. Also both have 16 pcie lanes.
Im guessing the 9700 is priced at top dollar as is new and may be able to get a good deal on the 8700k?. :thinking:


(David) #117

I am planning to build a new pc for gaming and Deep learning. I put this together on pc part builder. Does anyone have any comments or feedback at all?

https://uk.pcpartpicker.com/list/4XzqQZ

Cheers!


#118

Beginner who had some perishable credit on electronics vendor with limited computer parts supply. Got my hands on a evga rtx 2070 xc 8GB and a big case. Started the fastai course but the workflow with cloud resources got too annoying among other things. Ready to dive back in.

So i have my hands on the rtx 2070 already. Don´t want to waste too much of its capability for most beginner/intermediate use cases but do not have that much to spend.

the impression i get from the latest guide update by Tim Dettmers is that CPU and RAM does not matter much. If so even a i3 7100 or g4560 would not hamper the gpu or only slightly (and those cpu´s are really cheap).

Can´t have multiple gpus anyway with the 2070. one of Cpu´s above, two ram slot micro mobo with dd3r 1x8gb(optionally another added on) would be really cheap. Bad idea? if so why?


#119

CPU and RAM don’t matter much for training, but you will be doing a good bit of preprocessing, and you will appreciate having a bit more beef. I’d at least add the second stick of RAM, and review pricing to bump the CPU up a notch if possible.


(Sahil Gupta) #120

Agreed with your point GPU power is essential training models, My 4 x 2080 Ti powered Ant PC is quite a performer with these beats… I have seen people investing heavily on Processor no clue why?


#121

Hi Guys, I am struggling with with my build.
Currently I am trying to decide mobo, gpu and cpu (I think the rest will be easy)

Here is what I have right now:

https://pcpartpicker.com/list/cyxXFt

I want a CPU with 12 cores as I want to do a lot of preprocessing on images.

As for GPU, I am thinking with 2x 2070s and I will wait and see If I want to upgrade.
It seems I should use FP 16 and go with 2070s according to some blogs I read.
EVGA - GeForce RTX 2070 8 GB Black Video Card.

It seems I have to be careful with the type of cooler on these GPUs if I use multiple GPU.
Is this one okay?

I am not sure about the MOBO. Is seems most mobos for this CPU cost close to $300 which is a bit surprising.

Should I just go with https://pcpartpicker.com/product/kjmxFT/asrock-x399-taichi-atx-tr4-motherboard-x399-taichi?

Thank you for any feedback.