Elixir Deployment Options - with GPU

There’s a large thread on general deployment options here (as well as many more in the #deployment tag). This thread is for deployment options where you need a GPU :023:

I wonder how people are going to approach this in production. While dedicated servers can be rented at fairly reasonable prices, those with GPUs jump up quite a bit per month, so my current thoughts are to just build a linux workstation and run it from home. I’d then set it to check the live server once per minute for any new jobs, and if there are any to fetch them and push the results back to the server. The project I have in mind is going to be aimed at consumers (and will cost them nothing) so I’d be willing to give this a try, simply telling them that we’ll email them once their photos have been completed and what the current wait time is. I wonder if anyone else is aiming to approach things similarly?

(If this gets any replies we can split it into a new thread.)


Kurt from fly.io posted this today:

If it is like fly machines where I can pay only for the storage and boot in seconds a machine for my workload it will be awesome.

1 Like

Ah that’s awesome! It’ll be great being able to recommend Elixir-centric hosts instead of AWS or Google cloud/compute :003:

Let’s move this to a dedicated thread as I think this is going to be a big topic in the coming years :023:

What are the options?

  1. Probably the most obvious and quickest way to get started: use AWS, Google Compute Engine or some other cloud service. (Hopefully soon Fly, Render and other Elixir-centric hosts will offer GPU services too.)

  2. Rent a dedicated server with a GPU.

  3. Buy/build your own server and co-locate it.

  4. Buy/build your own server and run it from your office/home.

Assumptions: Your app is already hosted somewhere and so this is primarily for your GPU related requirements.

What are the costs?

1 - AWS

Elastic Graphics Accelerator Size Graphics Memory Price
eg1.2xlarge 8 GiB $0.400
eg1.large 2 GiB $0.100
eg1.medium 1 GiB $0.050
eg1.xlarge 4 GiB $0.200

I’ve not used anything like AWS but I’ve read that it’s something like $3 per hour - not sure how accurate or well that translates to real-world usage.

2 - Dedicated server with GPU

The costs seem to vary, but I’ve selected this one because it contains a decent amount of RAM and a GPU that you can easily buy yourself (for comparison with the next option). From serverroom.net (I’ve actually never heard of them before but they are one of the companies advertising when searching for ‘Dedicated server with GPU’:

There are cheaper alternatives available, but they usually start at at least $150 p/m.

3 - Build your own server and co-locate it

Server costs will depend on your spec (but you could go with what you can afford and then upgrade the GPU later). If you built a server with the equivalent spec as the dedicated server example above, I guess you’d recoup your costs within half a year or so (Geforce RTX 3079 is about £500 to buy). Co-location costs will vary too depending on where you’re located and what you need, but in the UK a 1U rack will cost about £25 a month (2U £40, 3U £113, 4U £135 - prices from servercolocation.uk).

4 - Build your own server and run it from your office/home

This is the most interesting to me because I imagine most people will be hosting their user-facing apps elsewhere (either some cloud service or a dedicated server) and you’ll only be using this server for computational tasks.

I reckon this could be a good option for when those tasks are not mission critical or not required in the quickest possible time (so for instance where you’re happy to tell your users you’ll email them when their stuff is ready - perfect for what I personally want to do!) The only costs here then are for the server, your usual internet connection and electricity, and if you need better contingency or speeds, you could get a leased-line (these start from around £100 a month in the UK depending on what you need and where you’re located).

Pros and Cons?

Can you think of any?

Here are some of my initial thoughts:

  • A cloud service will be the easiest and probably the cheapest for low-use. But will probably be the most expensive once you start getting into moderate to high use.

  • Dedicated servers (rented) will probably be ideal for those who are moderate to high use and are happy to manage the server/s but don’t want the hassle of co-locating or building their own datacenter.

  • Co-locating will probably be good for those with moderate to high-use, or those who want to save money in the long term (higher initial investment but you save a lot over time). For the largest users, they could build their own datacenter.

  • Office/home with or without a leased line will probably be good for those who want to experiment with the technology or are on a tighter budget. (For those who think this is a crazy idea, years ago a large number of companies hosted their own sites in-house with leased-lines and back-up generators.) Another advantage of this is that you can use the server to train your models from the comfort of your own office, perhaps being its primary use before being deployed.


For people that are willing to go this route an important recommendation is to get in contact with the datacenter before to know about restrictions in hardware, you can mount a server with a couple of RTX cards only to discover that the data center will not accept it because active cooling is problematic for them.


Good point Thomas. Check power requirements too, 1U will usually come with 0.3 AMPs as standard, whereas 4U will give you 2.0 AMPs as standard.

I have a feeling we’re going to see a lot of people/companies opt for option 4, particularly for services like Avatar AI - where it’ll be fine to email people when their stuff is ready (think current wait time at Avatar AI is about 30 minutes, and they’re using a cloud service).