eGPU for Deep Learning

Nima Dehkordi
3 min readDec 5, 2020

--

Due to the Covid-19 pandemic I wasn’t able to work from the office so I considered an eGPU setup at home. At first I was very reluctant about using an eGPU but after some research off internet I decided to go for it.

System Specs

First thing you need for an eGPU setup is of course a laptop which supports Thunderbolt 3. The laptop I am using is:

  • Dell XPS 13 7390
  • 16 GB Memory
  • 512 GB Storage
  • i7–10510U Comet Lake Processor
  • NVIDIA GeForce GTX TITAN X

eGPU Enclosure

For the PCI-e to Thunderbold 3 connection there are two options. First is the messier DIY solution which is also cheaper.

For that you can use the The ADT-Link R43SG-TB3 or similar + an ATX power supply and you’re good to go this option costs around 150€ and you end up with something like this:

DIY eGPU Solution — Image adapted from ADT-Link

Another option is a dedicated eGPU enclosure which is more expensive but keeps your desktop tidier. There are many options available but an informative buyer’s guide can be found in the eGPU.io website.

In the end I decided to go with Razer Core X which is among the cheapest and most reliable enclosures and not surprisingly on top of the list in the buyer’s guide.

Installation

it was pretty much just plug and play without much of hassle.

Driver

CUDA Toolkit

Cuda installation guide you can run the following code in your terminal

or based on your operating system run the script form NVIDIA website.

Tensorflow / PyTorch

Performance

The benchmark is conducted using the https://www.userbenchmark.com/ website. The following shows the performance with eGPU connected:

UserBenchmarks: Game 57%, Desk 81%, Work 49%
CPU: Intel Core i7–10510U — 69.8%
GPU: Nvidia GTX Titan X — 75.8%
GPU: Intel UHD Graphics — 4.9%
SSD: Pc601 NVMe SK hynix 512GB — 203.7%
USB: PHD 3.0 Silicon-Power 1TB — 33.1%
RAM: Hynix H9CCNNNCLGALAR-NVD 2x8GB — 71.4%
MBD: Dell XPS 13 7390

Without eGPU the performance drops to :

UserBenchmarks: Game 11%, Desk 74%, Work 11%
CPU: Intel Core i7–10510U — 72%
GPU: Intel UHD Graphics — 4.8%
SSD: Pc601 NVMe SK hynix 512GB — 208.9%
USB: PHD 3.0 Silicon-Power 1TB — 34.5%
RAM: Hynix H9CCNNNCLGALAR-NVD 2x8GB — 68.9%
MBD: Dell XPS 13 7390

--

--

Nima Dehkordi

I am a software developer and machine / deep learning enthusiast. I am interested in a broad range of topics including: AI, Economy, Psychology, and History.