Professionals in Data Science have a variety of options when it comes to computing. As far as the algorithms can be paralleled (it’s exactly the case with most machine learning models — with a lot of matrix multiplication), model training can be done using CPUs, GPUs and also TPUs. What does it mean from the business perspective?
You can either
- rent AWS, Google or other cloud;
- compute your models locally, using your own hardware;
- join Megamind platform and use distributed computing resources:
- Sell your own computing resources;
- Buy distributed cloud resources.
Well, technically you can even use some free computing power offered by such services as e.g. Google Colab, but you can’t have more than one GPU at a time if you use it for free. So it doesn’t count and we don’t consider it as an option, because it will take you days and weeks to process a somewhat serious amount of data using only 1 GPU.
So a number of data scientists all around Reddit, Medium and towardsdatascience calculated (that’s what they are good at, right?) that buying and operating your own PC can be 10 times cheaper than AWS or GCP rent. Jeff Chen, an AI professional, believes that a suitable machine can be built for about $3,000 without including tax. Once your personal rig is built, the only recurring cost to pay for is power. It costs $3 (£2.28) an hour to rent a GPU-accelerated system on AWS, whereas it’s only 20 cents (15p) to run on your own computer.
Assuming a computer with 1 GPU geared for deep learning depreciates to $0 in 3 years, the chart below shows that if you use it for up to 1 year it’ll be 10 times cheaper than Amazon web services and EC2 (including electricity costs):
However, there are a few drawbacks with a local deep learning workstation, such as slower download speed, the need for static IP for access outside your home network, and you might consider updating your GPUs sometime in the future but the money you will save is so great it’s worth it.
One of the biggest assumptions in this case is that GPUs in your PC might work 24/7 for months. Whereas Machine learning takes a lot of hardware time, it is rare that GPUs are always 100% busy. Ultimately, letting your hardware ‘rest’ for quite some time may shift your breakeven point farther in the future.
However with Megamind you can consider the following:
- you don’t buy a PC and just rent distributed cloud resources — according to our forecasts, Megamind’s prices for a typical ML practitioner can be up to 3-5 times lower than AWS, GCP solutions;
- OR you can buy a PC for Machine learning, and when it is not in use you can sell its power in Megamind platform. When your PC is not enough for a particularly complex task you can also buy additional computing resources from Megamind.
In the near future we will perform a more in-depth and thorough economic analysis within the scope of the proof of concept. We plan to release Megamind’s PoC in October.