I know we have a dedicated thread for each individual platform, but I thought it might be a good idea for us to have one thread where people share their reviews/experiences with all the platforms to make it easier for new students to choose one.
I personally have had a very bad experience with Paperspace Gradient and now want to switch to a new service, but am not sure what would be best. Please share your experience with whatever service you’re on in terms of ease of use, user interface, support and so on…
Paperspace has lost my code multiple times on one of their C7 notebook instances (my fault for trusting them and not keeping a local/github copy). This cost me ~8 hours of work one day, and ~2 hours of work another. I am sure it was on their end as it was reproducible and I did so multiple times. They stopped replying in the fastai support thread, seem to be almost completely inactive on twitter (and didn’t respond to my tweet about this). Their official support responded telling me they were passing it to an engineer, but I heard nothing for over a week.
I think they have a great ease of setup and I love the interface. It’s usable and easy to setup/switch machines, as well as install new software, but losing data and unresponsive support in the face of that type of problem means I’m switching to whatever people here recommend most.
Edit: I want to add that shortly after posting this thread (but apparently unrelated), one of their engineers reached out to me and has been more responsive. They also mentioned they would start more active support in the forums here. I’m still pretty disappointed by this bug in the first place, and how they initially handled it, but hopefully things turn around because they do have a nice platform.
Have you tried GCP? It is really unbeatable in quality and ease of setup.
For cheap projects I really like vast.ai
Also vast.ai let’s you test all the consumer GPUs like the RXT 2070, 80, 80ti for super cheap.
I’ve had a machine on Paperspace. Their bandwitth mush have gone down recently, as it’s been pretty slow.
Currently trying vast.ai. It’s cheaper, faster, and more flexible. One thing is if you stop your instance you might have to wait a while for it to be available again.
Paperspace convinced me to come back by issuing me credit and promising improved support (one engineer who stepped up and was very kind/helpful). Unfortunately there’s been zero response in the fastai thread, and I just experienced a bug for a third time. This one was on a different machine, and wiped out absolutely everything I hadn’t stored in /storage or backed up offline. I’m definitely in part to blame for not backing up more of my experiments, but this was the last straw for me and I’m signing up for GCP per Thomas’s recommendation.
I’m not in love with the idea of everyone using Google for every single service, but at least I can trust that my data won’t be erased at random with no explanation. I’ll report back on my experience with GCP.
Hey guys, just wanted to give a review for GCP. So far it’s been amazing! There are lots of little features that Paperspace didnt have that make a world of difference. Their jupyter terminal functions like a normal unix one. You can tab complete, press up-arrow to get your last command, and so on. Paperspace didn’t have this functionality for whatever reason (I thought it was all Jupyter terminals that were broken). I’ve also found their filesystem to be a lot faster so when I want to copy files to make test sets I don’t get stopped waiting for 30-120 seconds for each (an easy time for me to get distracted haha).
Definitely give GCP a shot if you’re not sure. The setup isn’t bad and the guide here is great. I’d love to hear about some smaller independent platforms though. If something like Salamander, FloydHub, or Crestle has the same level of functionality I’d love to give them my business.