Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!


TensorDock: Affordable, Easy, Hourly Cloud GPUs From $0.32/hour | Free $15 Credit! - Page 4
New on LowEndTalk? Please Register and read our Community Rules.

TensorDock: Affordable, Easy, Hourly Cloud GPUs From $0.32/hour | Free $15 Credit!

1246711

Comments

  • lentrolentro Member, Host Rep

    @jmgcaguicla said: Heroku oops page after clicking Register and again when I try to sign in

    Can you PM me your email? I see one failed registration somehow, not sure if that's you.

  • Ohh nice. Thx for more GPU options!

    Thanked by 1lentro
  • lentrolentro Member, Host Rep

    @ElliotJ said: Is your goal also to make the world a better place?

    LOL yes yess

    "Making the world a better place through scalable, hourly, affordable GPU-accelerated HPC computing"

    In all seriousness, yes. It'd be cool to make money while enabling ML people to compute easier. I've talked to so many startups that have no idea about the infra side of things. Many actually buy super expensive servers because of how expensive cloud platforms are. I hope to run a small DigitalOcean for GPUs.

  • lentrolentro Member, Host Rep

    @jugganuts said: Thx for more GPU options

    Thanks! Yes, you can choose exactly what GPU you need and nothing more :)

    A100 SXMs are running low right now so I've been thinking of adding Oracle as an upstream supplier. Other than that, I think every GPU has some stock.

  • lentrolentro Member, Host Rep
    edited December 2021

    @drunkendog said: I've used Dash-Cloud, and it's definitely a premium service with good prices and great performance

    Yayyy great to see you! Thanks!

    If you ever need credit, PM me, I'd love for you to continue running on my servers :)

    Thanked by 1drunkendog
  • lentrolentro Member, Host Rep

    @KENTKING said: paypal payment

    I've had bad experiences with PayPal and chargebacks. Also, their fees are slightly higher. With Stripe we're requesting 3DS on all payments and blocking anything with a remotely high risk score. Not sure if I can do a similar thing with PayPal...

    But thanks for the suggestion! If we add another payment method, we'd probably add crypto first. What do you think?

  • lentrolentro Member, Host Rep

    @Erisa said: wrong string lol

    Good catch, fixed and pushed! You should see new text in the dashboard soon :)

    Thanked by 1Erisa
  • lentrolentro Member, Host Rep

    @lentro said:

    @derekyang said: @lentro surprising i can't find any link to reset the password , my browser did not remember ad-hoc generated password

    Ok password resetting is literally working right now in our dev branch. I'll review it tomorrow and get it pushed.

    Don't worry, you'll still get your credits :)

    Remind me if I forget

    ~Rich

    @derekyang, Does this work for you?
    https://console.tensordock.com/reset_password

    Thanks!

  • Looks awesome! Good luck, mate ;)

    Thanked by 1lentro
  • lentrolentro Member, Host Rep

    @Ympker said: Looks awesome! Good luck, mate

    Thanks @Ympker! Let me know when you need credits and I'll top up your account! :)

    Thanked by 1Ympker
  • @lentro said:

    @KENTKING said: paypal payment

    I've had bad experiences with PayPal and chargebacks. Also, their fees are slightly higher. With Stripe we're requesting 3DS on all payments and blocking anything with a remotely high risk score. Not sure if I can do a similar thing with PayPal...

    But thanks for the suggestion! If we add another payment method, we'd probably add crypto first. What do you think?

    OK!I see what you mean.I'll go along with you.

  • Cloud GPUs at https://tensordock.com/, ID recdhpkmeg

  • lentrolentro Member, Host Rep

    @dom1024 said: recdhpkmeg

    Done, check your account! Remember to post feedback to get another $10 once you try out the service :)

  • Cloud GPUs at https://tensordock.com/, ID recsoqvcof

  • Cloud GPUs at https://tensordock.com/, ID rec2b5pici

  • lentrolentro Member, Host Rep

    @jerry_me said: recsoqvcof

    Done, check your account! Remember to post feedback to get another $10 once you try out the service

  • lentrolentro Member, Host Rep

    @Iroshan464 said: rec2b5pici

    Done, check your account! Remember to post feedback to get another $10 once you try out the service :)

    Thanked by 1Iroshan464
  • Yes, it works well for me now, thanks btw for the extra credit!

    @lentro said: @derekyang, Does this work for you?

    https://console.tensordock.com/reset_password

    Thanked by 1lentro
  • lentrolentro Member, Host Rep

    @derekyang said: extra credit

    Shhhh, keep it quiet! :wink:

    Thanked by 1yoursunny
  • When I create the instance, it prompts:

    Do not leave this page. VMs take 45 seconds to 5 minutes to deploy. Once deployed, each takes another two minutes to boot. Please contact us if you are waiting more than 10 minutes.

    In fact, it only takes a few seconds to create it successfully.

    I got the IP is:216.153.50.X

    Nvidia-docker is pre-installed in the Ubuntu 20.04 system, but user users does't have operating permissions, so I added permissions.

    Unfortunately, tensorflow is not pre-installed. After all, the compatibility of various versions is really troublesome.

    I decided to run a benmark(cifar):

    git clone https://github.com/tensorflow/benchmarks
    python3 benchmarks/perfzero/lib/benchmark.py --git_repos="https://github.com/tensorflow/models.git;benchmark" --python_path=models --gcloud_key_file_url="" --benchmark_methods=official.benchmark.keras_cifar_benchmark.Resnet56KerasBenchmarkSynth.benchmark_1_gpu_no_dist_strat
    

    But...

    2021-12-29 07:48:01.386804: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1850] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.

    Run after reinstalling CUDA support(lib):

    2021-12-29 08:01:38,318 INFO: TimeHistory: 25.68 seconds, 49.84 examples/second between steps 0 and 10
    2021-12-29 08:01:38,318 INFO: TimeHistory: 25.68 seconds, 49.84 examples/second between steps 0 and 10
    2021-12-29 08:02:00,398 INFO: TimeHistory: 22.08 seconds, 57.97 examples/second between steps 10 and 20
    2021-12-29 08:02:00,398 INFO: TimeHistory: 22.08 seconds, 57.97 examples/second between steps 10 and 20
    2021-12-29 08:02:22,293 INFO: TimeHistory: 21.89 seconds, 58.46 examples/second between steps 20 and 30
    2021-12-29 08:02:22,293 INFO: TimeHistory: 21.89 seconds, 58.46 examples/second between steps 20 and 30
    2021-12-29 08:02:44,184 INFO: TimeHistory: 21.89 seconds, 58.47 examples/second between steps 30 and 40
    2021-12-29 08:02:44,184 INFO: TimeHistory: 21.89 seconds, 58.47 examples/second between steps 30 and 40
    2021-12-29 08:03:06,145 INFO: TimeHistory: 21.96 seconds, 58.29 examples/second between steps 40 and 50
    2021-12-29 08:03:06,145 INFO: TimeHistory: 21.96 seconds, 58.29 examples/second between steps 40 and 50
    2021-12-29 08:03:27,979 INFO: TimeHistory: 21.83 seconds, 58.63 examples/second between steps 50 and 60
    2021-12-29 08:03:27,979 INFO: TimeHistory: 21.83 seconds, 58.63 examples/second between steps 50 and 60
    2021-12-29 08:03:49,474 INFO: TimeHistory: 21.49 seconds, 59.55 examples/second between steps 60 and 70
    2021-12-29 08:03:49,474 INFO: TimeHistory: 21.49 seconds, 59.55 examples/second between steps 60 and 70
    2021-12-29 08:04:11,165 INFO: TimeHistory: 21.69 seconds, 59.01 examples/second between steps 70 and 80
    2021-12-29 08:04:11,165 INFO: TimeHistory: 21.69 seconds, 59.01 examples/second between steps 70 and 80
    2021-12-29 08:04:32,905 INFO: TimeHistory: 21.74 seconds, 58.88 examples/second between steps 80 and 90
    2021-12-29 08:04:32,905 INFO: TimeHistory: 21.74 seconds, 58.88 examples/second between steps 80 and 90
    2021-12-29 08:04:54,538 INFO: TimeHistory: 21.63 seconds, 59.17 examples/second between steps 90 and 100
    2021-12-29 08:04:54,538 INFO: TimeHistory: 21.63 seconds, 59.17 examples/second between steps 90 and 100
    2021-12-29 08:05:16,538 INFO: TimeHistory: 22.00 seconds, 58.18 examples/second between steps 100 and 110
    2021-12-29 08:05:16,538 INFO: TimeHistory: 22.00 seconds, 58.18 examples/second between steps 100 and 110
    110/110 - 244s - loss: 0.8967 - sparse_categorical_accuracy: 0.9909 - 244s/epoch - 2s/step
    

    The performance seems to be 2.5 times that of my 3080, but there is no training data for the time being, and I look forward to a good experience in the future.

    Finally, I used the API to delete the instance.Of course, I deliberately made a wrong attempt, but when the wrong API KEY was entered, the content returned was not JSON, so it didn't feel very friendly.

    curl --location --request GET 'https://console.tensordock.com/api/delete/single?api_key=xxx&api_token=xxx&server=recWwTgbFrjpgipSb'

    user ID for promotions: recx6gpwrt

    Thanked by 2yoursunny lentro
  • jmgcaguiclajmgcaguicla Member
    edited December 2021

    It would be nice if you could get an AMI-like system going so people can spawn multiple VMs at once from a template or scale vertically by switching to a beefier GPU instance without re-deploying their stuff.

    I believe most people's use case would be to scale-up/down on-demand, and not leave it on 24/7 (hence the hourly pricing model).

    Even something simple like cloud-init or startup scripts (ala Linode's StackScripts) can still be pretty useful here.

    Thanked by 2yoursunny lentro
  • lentrolentro Member, Host Rep

    @mcgree thank you so much for your review!

    @mcgree said: reinstalling CUDA support

    This seems to be a pain point, as @yoursunny mentioned earlier. So we will work on making all software versions compatible with preinstalled TensorFlow, Keras, as well as Jupyter Notebook options.

    @mcgree said: 2.5 times that of my 3080

    2.5x faster?! Wow! the GPU that you deployed is a Quadro RTX 4000 which should be worse, so that's really great! Though, the RTX 4k apparently has more tensor cores than the 3080, so I guess it could be possible.

    I see that your test VM rate was $0.32/hr, if the performance was comparable to a 3080's performance... then we cost 70% less than this competitor :joy:
    https://www.genesiscloud.com/pricing#rtx-3080

    @mcgree said: not JSON

    We'll look into this. Agreed that the errors always have to be in JSON. Good catch!

    Thanks so much for your detailed review! Check your account for credits, I hope it'll be enough to last you a while :)

  • lentrolentro Member, Host Rep

    @jmgcaguicla said: AMI-like system going so people can spawn multiple VMs

    Gotcha, agreed. Defo would be cool. That's how the backend for deployments is done so it shouldn't be that hard to add support for you to be able to deploy your own OS images, we'll probably add it within the next few months :)

    @jmgcaguicla said: beefier GPU

    This is interesting. We will consider it!

    @jmgcaguicla said: cloud-init or startup scripts

    Will do! Same thing with SSH keys. Check your account for a small feedback bonus :)

    Really thanks for all the support! Getting a lot of useful feedback, thanks everyone!

  • jmgcaguiclajmgcaguicla Member
    edited December 2021

    Small UX feedback on the panel, bordering on being a nitpick: on a 1080p display you need to scroll all the way down to be able to see the "Deploy" or otherwise "Out of stock. Try another location?" button.

    I spent a good minute adjusting the specs only to find out it's out of stock when I finally reach the bottom of the page. Not really a big deal, just think an item being OOS could be expressed much clearer from the get go while the user is configuring the instance.

    Thanked by 2lentro Ympker
  • Cloud GPUs at https://tensordock.com/, ID rec7tj02q5

    great console btw

  • lentrolentro Member, Host Rep

    @sirluis said: rec7tj02q5

    Done, check your account! Remember to post feedback to get another $10 in account credits once you try out the service

  • HalfEatenPieHalfEatenPie Member
    edited December 2021

    Just getting around to seeing this. Interesting to see that there's some more affordable options within this space. We work with a ton of ML models and I've built our own cluster with about 40 V100s.

    Definitely great to see additional competition in this space so that $$$ isn't as big of a barrier to entry for training models.

    My questions are less technical than @yoursunny's, but rather understanding the workflow.

    As you know, many ML models requires access to large datasets, some frequently needing higher I/O bandwidth.

    1. How's your storage structured? Is it local storage or NAS storage? If it's NAS storage, what kind of port speed/limitations?
    2. How much is the bandwidth cost? Ingress and Egress, as I'm sure most of the time it makes sense to simply download a copy of the dataset locally first.
    3. Related to my previous question, what's the expected workflow for the system? Do you provision your VM with a GPU and end up having to pay for it as you import your dataset or is it possible to provision an instance first to get a local copy of your dataset?

    Can you give an example walkthrough of how a normal "team" should use your system? I'd like to compare it to the usability of Microsoft Azure ML Studio. I was recently involved in a deal where our client awarded Microsoft Azure ML Studio with a significant amount of $$$, but if it's possible (without too much problem) to do our workflow using your system, I think it'd make sense to give it a shot.

    Looking forward to trying this out later.

    Thanked by 2yoursunny lentro
  • lentrolentro Member, Host Rep

    Hi @HalfEatenPie!

    Thanks for the questions, and feel free to ask more!

    1. Storage on each VM is structured as an NVMe-based Ceph cluster backed by 40gbps internal networking, but we can also run NFS shares. For now, we've been doing those kinds of setups for larger clients.

    2. Both egress and ingress is free. Just don't abuse your port by using it 100% 24x7 on a single $0.50/hr instance. We have 10gbit uplinks, which should be more than enough for most people.

    3. For now, yes, you have to download your dataset on your GPU instance. In the coming months, we'll add CPU instances. These will cost like $0.03/hour with the option to be upgraded into a GPU instance, so you'll only pay for the GPU compute when you need it.

    Eventually (6+ months), we'll probably add fully-managed networked storage, so you will be able to upload files through our console and then select which folders/files to then mount to your VM during the deployment process.

    I'd be happy to give you some credits to try us out. Just PM me your email address or user ID if you sign up :)

    At the moment, the biggest differentiator is cost. Azure ML Studio, AWS Sagemaker, and Paperspace Gradiant's Workflows are managed ML services that integrate into your CI/CD process. TensorDock, for the time being, is more like cheap GPU servers.

    The reason why we're doing raw compute power for now is because when we were doing our fundraise, we decided that we would rather get 6 figures from friends and family than traditional VCs, where we could get 7 figures. We decided that instead of burning cash, growing fast, and becoming a unicorn, we were more interested in running a financially sound business that wouldn't crash without constant cash injections and that provides long term value for customers over the next 20 years or more. That is not to say that investors are bad: we will work with investors who share the same long-term thinking as us.

    Anyways, the profits from this iteration of the company will enable us to self-fund the development of additional features that will turn us into an even easier-to-use service (at the same low prices that you see here, of course).

    Here's a YABS if you're curious about disk performance or network: https://pastebin.com/raw/Vxecr2AV

  • @lentro

    I have already received the credit, but I will probably use it officially in a few months, after all, the data set on hand is relatively small.

    Thanked by 1lentro
  • lentrolentro Member, Host Rep

    @mcgree said: but I will probably use it officially in a few months

    No worries, thanks -- good luck! Lmk if you ever need any assistance :)

Sign In or Register to comment.