Hello Matthias_Kleine, If I understood your question well, You’re looking for a platform with high computational capacity to carry out ML and DL calculations? There are three options Cloud computing, On-premises, and Hybrid. But each of them will come at a cost depending on the size of your projects.
For On-premises you’d need:
specialized skill and expertise
computational special-purpose hardware which is kinda costly for development, infrastructure, and workforce.
Machine learning frameworks often face issues during scaling up as they need more computers.
For Cloud Computing
It is by far more scalable, Access to the newest hardware at the click of a button.
Hybrid solutions
I would prefer it for your case If you already have on-premises hardware but all of the above there must be a trade-off between cost and efficiency just try to weigh in which work best for you.
But if the project is not that much huge you can use Google Colab as pointed out by @Kic
As both Kic and Isaak have mentioned, Colab is a good platform to try first for Cloud Computing. It uses Jupyter Notebooks as the interface, so it’s easy to get started. The way it handles file storage is slightly different than the notebook platform that the DLAI courses use, but it’s pretty easy to adjust things. They also provide GPU support and you can use it for free, as long as you’re willing to wait in a queue to execute when things are busy with the paying customers. In free mode, it’s also a bit tricky to get training to run for more than single digit hours: you have to trick the notebook into thinking you are interactively using it otherwise it just suspends. But if you implement checkpointing of your training, then you can just restart and pick up where you left off. You’ll need that capability anyway to save your trained models.
I’ve only used Colab in free mode and have not probed the limits of how much GPU power you can really get in the limit. There are more general platforms like AWS that provide essentially unlimited hardware power, but I have not tried DL training there. I’m sure they support it, but I don’t know if there is a “free trial” mode.
New hardware will help, but the problem with buying your own hardware is that it is always limited and adding more capability takes time and money. These days the demand for GPUs is very high because of all the crypto miners, so it’s not necessarily easy or quick to build a custom “training rig”. Of course it also matters a lot the scale of the problems you are trying to solve. Maybe you start small and a powerful laptop is enough, but what if your needs grow over time? With Cloud Computing, you just “dial in” as much as you need. Of course it’s not free once you are “at scale”, but it’s cheaper than owning the hardware yourself. And the flexibility is worth something as well.
Of course you can also do a “hybrid” strategy as Isaak described: use your own hardware for the problems that are at that scale and then when you need more h/w power you can access your Cloud Service. With Cloud Services, you just pay for what you actually use in most cases. Check the billing models of the various services for more info.
I recommend Google Colab with google drive as persistent storage. If you are working on a Kaggle problem, you can use their platform too. Note that free access has certain time limitations.
As @rmwkwok mentioned, you can try Kaggle with some limitations. Regarding the local environment, I do not recommend a laptop because to train a model you must have good hardware and, in this area, when we compare a laptop with PC, the PC is much cheap. I started with a laptop but had to move to a PC because the laptop was heating a lot. In PC you can even have a water cooling, if necessary. Anyway, if money is not the problem, NVIDIA has a good laptop for AI.
As people have mentioned Google Colab is an excellent option. You could also sign up for Google Cloud and receive 300 USD to spend within 3 months. I don’t think a credit card is necessary either.
There is also Saturn Cloud, this platform has the interesting option of attaching an extra server to use to host a dash cluster - super easy to implement.
The other main Cloud companies, Microsoft Azure and IBM, also do free trials, but Google’s is the most $ and the longest with the most flexibility in setting up VM’s. To use GPU’s you need to send in a quota request but TPU’s are fair game (and so unbelievably fast).
Then there is Paperspace - sort of like Colab but not as good.
