edit

FloydHub Documentation

Please see here for Billing FAQs

Legacy Plans FAQs

Note that these plans are no longer available. The following FAQs are for users who signed up before November 20, 2017. For the latest FAQs, please see Plans FAQs

Preemptible Instances deprecation notice

Starting March 7, 2018 we are no longer supporting premptible instances on FloydHub. All instances will now be dedicated instances. This change was made in response to AWS (our cloud provider) making changes to the way Spot instances behaved for the purposes of running Machine Learning training jobs. The spot instances were preempted too frequently and led to a poor user experience and low SLA. So we made the decision to upgrade all the instances our fleet to dedicated instances.

If you are using any preemptible instance Powerups they will be deprecated soon and you will not be able to purchase new powerups of the same kind. Existing preemptible powerups will continue to work but autorefreshes will be turned off on April 5, 2018. So we recommend that you switch to the new Powerups before the cut-off date. Contact support@floydhub.com if you have any questions about this.

Which Plan is right for me?

Free Plan

If you're just exploring, the Free plan is for you! You are automatically enrolled in the Free plan when you sign up on FloydHub. It includes 20 hours of free CPU when you sign up. You cannot, however, use a GPU or run multiple jobs in parallel.

Data Scientist Plan

Our Data Scientist plans offer varying levels of job concurrency, GPU computing hours and storage. You can also purchase Powerups to add more compute hours to supplement your plan.

For more details on our different plans, visit our pricing page. Please see the feature comparison table for a full list of each plan's features.

What is in the Trial plan?

All users that sign up on FloydHub are automatically enrolled in the Free plan. Refer to our pricing page for details on the Free plan.

Once you've signed up, take FloydHub for a spin with our Quick Start Guide

Do the plans come with preemptible or dedicated instances?

The GPU and CPU compute hours included in your plan (Free or Data Scientist) are preemptible instances. This means that there is a small chance that your job will be terminated without notice. In practice, this happens infrequently and this is perfect for most users. If you need dedicated instances for your jobs, you can buy the GPU+ or CPU+ Powerups.

Do my remaining compute credits roll over each month if I don't use them all?

No, your monthly Plan compute credits are not rolled over. However, your Powerup credits will remain valid for one year from purchase date.

What happened to the old Pay-as-you-go Individual Plan?

We are transitioning from the Individual Plan, which offered a pay-as-you-go payment method, to our current pricing plan. The Individual Plan is no longer available for new users.

I am in the Pay-as-you-go Individual Plan. What will happen to me?

If you signed up for the Individual Plan before August 20th 2017, you will be grandfathered till October 1st 2017. After this, you will be automatically enrolled in the Free plan. Please note that any remaining promotional credits will also expire on this date.

Please upgrade to one of the Data Scientist plans to continue using FloydHub without interruption. We will also be reaching out to you with more information about this transition.

Why did I not get 100 free GPU hours when I signed up?

We offered 100 hours of free GPU for all users during our promotional period. This has ended.

Are there any academic discounts for students?

We don't have discounts. However, a lot of students create content for us. If you are willing to contribute high quality content to FLoydHub, we will give you free GPU credits in exchange!

Content we are looking for:

  • Technical blogs on deep learning and AI
  • FloydHub tutorials, text or video
  • Port popular deep learning projects to FloydHub
  • Create interesting datasets
  • <Insert your own idea here>

If this is interesting to you, please let us know about it here.

Compute

What is job concurrency?

Job concurrency is the number of jobs you can run in parallel. Each plan has a limit on the number of concurrent jobs you can run. For example, in the Free plan, you can only run 1 job at a time. In the Data Scientist Pro plan, you can run up to 8 jobs in parallel.

Having a higher concurrency is useful when you want to parallelize your training, for example while hyperparameter sweeping.

What will happen to my running job when I run out of computing credits?

You job will be shutdown immediately when you run out of computing credits.

If you run long-running jobs and expect them to exceed the computing hours offered by your plan, you can purchase Powerups.

You can also enable auto-refresh on your Powerups to ensure your long-running jobs are never killed because you ran out of computing hours. We'll automatically refresh your selected Powerup so that your job can continue running.

Preemptible vs. Dedicated Instances

Preemptible Instances

Preemptible instances have medium job uptime SLA of 98%. This means that there is a small chance that your job can be terminated (preempted) at any point during its runtime by FloydHub if it requires access to those resources for other, higher priority tasks.

Preemptible instances (CPU / GPU) offer top notch compute at affordable prices, in exchange for fault tolerance.

Note that SLA refers to what we can guarantee. In practice, this happens infrequently. Historically, less than 0.1% of jobs run on FloydHub have encountered interruption. However, you need to be aware that there is the possibility.

Why Do You Use Preemptible Instances?

To be able to offer you compute at a much lower cost.

We have a fixed pool of resources that we have to allocate amongst all our users. Some of our users require dedicated instances and are willing to pay the premium for uninterrupted access. But, the majority of our users can tolerate a 98% job uptime SLA for the significant price savings that preemptible instances offer.

Will I Get a Refund if My Job is Preempted?

No.

Our preemptible instances have a 98% job uptime SLA. By using them, you are accepting a small chance of your job being terminated without notice, in exchange for paying a much lower price than dedicated instances.

How Will I Know When My Job Is Preempted?

Your job's state will turn from Running to Shutdown. We will send you a notification informing you about this. Unfortunately, we are currently unable to warn your ahead of time of an impending preemption.

What is the SLA of Preemptible Instances?

Preemptible instances have 98% job up time SLA.

Dedicated Instances

Dedicated instances have high job uptime SLA of 99.95%. Use dedicated instances for your jobs if they are critical or not fault tolerant. You can purchase '+' Powerups (CPU+ / GPU+) to utilize dedicated instances.

Why Do You Use Dedicated Instances?

If your job is not fault tolerant and cannot withstand a small (<2%) chance of your job being shutdown without notice, you should use our "+" Dedicated instances.

Price sensitivity also plays a factor - dedicated instances are more expensive than premptible instances.

Given that deep learning models typically train over long periods of time, it is good practice to build your application to be fault tolerant by regularly checkpointing your training.

What is the SLA of Dedicated instances?

Dedicated instances have 99.95% job up time SLA.

What is the difference between GPU vs. GPU+ and CPU vs. CPU+?

GPU and CPU are preemptible instances. GPU+ and CPU+ are dedicated instances.

Powerups

What are Powerups?

Your subscription plan comes with a monthly quota of CPU and GPU computing hours. If you need more computing hours, you can buy Powerups to supplement your plan.

What Powerups should I buy?

This depends on your computing needs. We offer multiple tiers of Powerups:

  • Preemptible vs. Dedicated: CPU / GPU are affordable preemptible instances, CPU+ / GPU+ are high-reliability dedicated instances.
  • 10 vs. 50 vs. 100 hours: Purchase a pack that suits your computing needs. Note that the larger packs offer compute at a much cheaper rate/hour than smaller packs.
  • Auto-refresh: You can enable auto-refresh on any pack.

If you are just starting out and need more computing hours than your plan offers, you can start with the GPU10 Powerup.

If you run long-running jobs, you should purchase the GPU100 with auto-refresh enabled, to ensure that you never run out of computing credits.

If you run critical jobs that are not fault-tolerant, you should purchase the GPU+ Powerup.

How can I buy Powerups?

You can purchase them from your Powerups Dashboard

Why would I enable auto-refresh?

Auto-refresh ensures your long-running jobs are never killed because you ran out of computing hours. We'll automatically refresh your selected Powerup so that your Job can continue running.

Can I buy a Powerup if I am in the Free Plan?

No. You have to be enrolled in one of the Data Scientist plans to be eligible for purchasing Powerups.

Do Powerups expire?

Yes. Powerups are valid for 1 year from the date of purchase.

How will my Powerups be used?

Your compute hours will be consumed in the following order:

  • Hours from your subscription plan
  • Hours from free credits
  • Hours from Powerups

Storage

How much storage do I get?

Each plan comes with its own storage limit. Please see the feature comparison table for details.

What counts against my storage?

Storage is consumed by the datasets that you upload, your code and the data that your jobs output.

Note that you are only responsible for the data that you own. For example, if you use a public dataset in your job, you won't be charged for it.

Can I buy more storage than my plan offers?

If our Data Scientist Pro plan's storage doesn't meet your needs, please contact us at support@floydhub.com.

We will soon have a storage Powerup that can you buy to add more storage to your base plan.

Help make this document better

This guide, as well as the rest of our docs, are open-source and available on GitHub. We welcome your contributions.