Hacker News Reader

Comment by mbrookes | Hacker News Reader

The course is excellent, and thank you for making it and offering it for free - but a word of caution for those considering following it: Along the way you will incur not-insignificant costs for Amazon EC2 GPU instances, and, even if your instance is shut down, SSD EBS storage costs.

Edit: To be clear, I'm not suggesting it's not worth it, just highlighting that theres's more than a time commitment to budget for.

View on HN

Yep, the course is free, but you'll need to pay for the computing power one way or another.

If you have a workstation with a fairly recent Nvidia GPU (I used a GTX 980 TI) and bunch of spare disk space, you don't need AWS at all. You'll still pay for the electricity, of course, but it's not what I would call a significant extra cost. That is, if you already have the hardware.

What's the minimum nVidia GPU spec (VRAM) and SSD/HDD disc space (TB) required, from your perspective?

Get a GTX 1070 at a minimum. GTX 1080 Ti if you have a bit more cash to spend. Disk space is entirely dependent on your data set size. We're not talking huge numbers for most things you would be doing

I have GTX 970 and how much spare space do we need?

btw, I have a doubt, the course has specific video to setup aws, for local machines, what to do? should I install the required packages python, keras or whatever they use?

The install-gpu.sh script found in fastai's github can be used to set up the required environment on an Ubuntu box.

A GTX 970 should do. The exercises are designed for GPUs with more memory, so you'll need to use smaller batch sizes here and there.

To give a more exact estimate - if you don't use spot instances ($0.20/hour) it'll cost $0.90/hour, and if you do the suggested 70 hours of work for the course, that's $63. And then there's around $6/month cost for the EBS volume.

Not exactly.

The time you spend on the course, and the time spent by your instance running workloads are two completely separate things.

Your cost per-hour is accurate (plus a few bucks for an IP address), but the number of hours is off by an order of magnitude.

Also, I'm paying double that per instance per month for SSD using the provided scripts to build the instances. It's the smallest part of the cost, but I mention it because it can take newcomers to EC2 by surprise if they have an instance shut-down but consuming disk space.

I'm going through this course, but using Google cloud instead of AWS. I can confirm that at least the first Jupyter notebook works well for me.

I had to adapt the aws-install.sh script, but it was easy enough. I ended up using a snapshot instead of a persistent volume, as the monthly cost when you're not running is much cheaper. So I have a script to create a new instance and then restore that snapshot. It's faster than installing the dependencies each time).

Google Cloud has a $300 free trial for new users: https://cloud.google.com/free/

(Disclosure: I work for Google, but have no involvement in Google Cloud other than being a happy user)

Any chance you can share the adapted scripts? I assume they are easy to make, but the less barriers to entry the better.

Sure:

1) Create the instance the first time: https://gist.github.com/rahimnathwani/40ed3b6f496d377e3d17de...

2) SSH into the instance and run this script (replace anaconda mirror if your instance isn't in Asia, and definitely set the password hash): https://gist.github.com/rahimnathwani/b63ebc5900b832743d9e22...

3) After you've tested the server, snapshot the disk (using the web console) and destroy the instance and the persistent disk.

4) Run this script to recreate the instance and disk from the snapshot (replace with your snapshot name): https://gist.github.com/rahimnathwani/9019df33d4ad7ec4607413...

Buy a laptop with a GPU. If you are serious about learning AI, it's worth spending few hundred dollars to get a proper workstation.

I heard that nowadays it's possible to buy a laptop and a external GPU?

I have a mac, is it possible to really buy an external GPU??

Yes, but GPUs should only be consider for acceleration when there is insufficient local CPU power to accomplish something valuable. Otherwise, it's like buying a Ferrari to get groceries.

https://www.journaldulapin.com/2014/12/04/a-nvidia-maxwell-c...

https://tompaw.net/cuda-on-thunderbolt-egpu/

http://xlr8yourmac.com/tips/MBP_ThunderBoltVideoCard.htm

AKiTio Node | Thunderbolt3 External PCIe Box for GPUs https://www.amazon.com/dp/B06WD8KS52

Gigabyte GeForce GTX 1080 XTREME Gaming Premium Pack Video Card (GV-N1080XTREME-8GD Premium Pack) https://www.amazon.com/dp/B01HHC9Q3U

There's been some work on using AWS spot instances for the course, which can save quite a bit:

https://medium.com/slavv/learning-machine-learning-on-the-ch...

http://wiki.fast.ai/index.php/AWS_Spot_instances

Like how much? If you are going to get into deep learning for real seems like it might be worth building a multi-GPU workstation?

Don't buy equipment before you have demonstrated a real need for it.

This applies across basically all of life, and it's so frustrating to see people ignoring it, because what ends up happening is they use a string of 'gonnas' to justify buying stuff they don't need. Gonna get fit - buy $1500 worth of gym gear. Gonna learn electronics - buy oscilloscope, power supplies, tons of components. Gonna get your motorcycle license - buy brand new bike and stick it in the garage.

If you have a desktop computer, you're good to start. When you've done enough that your available CPU/GPU is limiting you on your own projects (not on something you pulled off github) then you can look at upgrading.

/rant

>Gonna learn electronics - buy oscilloscope

A fairly accomplished electronic engineer told me that they'd never once solved a problem using an oscilloscope, but that it helped to keep them occupied while they were mulling over what might have gone wrong. (That's presumably why the better ones have so many knobs and dials to play with, like one of those children's toys.)

I've certainly solved problems with a storage scope before, but not for a long time, and they were mostly software problems rather than hardware problems (ie. using it as a poor man's logic analyzer to infer what's going on with the code via a couple of spare IO pins). I really kinda want one though.

There is actually an accepted term for what you describe in many circles, called GAS -- Gear Acquisition Syndrome.

One forum user put together a <$700 complete PC with an adequate GPU. You definitely don't need >1 GPU.

Or you can use spot instances, as a sibling comment mentions - about $0.20/hour generally.

Ah, cool. I am familiar with AWS and spot instances, but after that comment was assuming it required multiple instances for training or something.

Build a fast deep learning machine for under $1K https://news.ycombinator.com/item?id=13605222 a month ago with many tweaks recommended in the comments.

gcp: Contrary to the claims there, the CPU does matter

dsacco: yes, this will work, it will quickly become suboptimal [...] as you scale your hobby into something resembling more professional work

brudgers: I'd start with a used Dell Precision T7xxx series off of Ebay for a <$300 including RAM and a Xeon or two.

croon: If the limit is a firm $1000, I would get something like this: https://pcpartpicker.com/list/XHV9Fd

I'd consider $60 bucks to pick up deep learning skills an "insignificant cost".

Well, that depends on how you define your cost function. :-)

"Along the way you will incur not-insignificant costs for Amazon EC2 GPU instances"

Is there some particular constraint involved that I cannot run the computations at home?

Heat death of the universe.

Buy a pc/laptop with an NVIDIA GPU and setup Linux on it. It is a better option than using EC2 instances.

How much VRAM is required?

Reply on news.ycombinator.com