Big Data and no Processing Power? Leverage Google’s Cloud TPU to Accelerate ML Tasks
- Google has added TPU support to it’s Cloud ML offerings to speed up the model building process
- With this addition, Google’s Cloud Platform customers can train high-end machine learning models with differentiated performance per dollar
- Cloud TPU recently came top in the ImageNet training cost category of Stanford’s DAWNBench competition
One of the biggest challenges in machine learning is finding enough computational resources to deal with the data at hand. When you take part in ML competitions, you must have faced the same problem when the data set size went into gigabytes. Standard machines do not have the capability to handle such large amounts of data. This is where GPUs and TPUs have come into play recently, but they can be expensive as well.
So to eliminate this issue for data scientists and smaller organizations, Google has announced that it will be offering TPU support for it’s Cloud ML platform. This will accelerate training machine learning and deep learning models, without having to sacrifice on precision and accuracy. The Cloud TPU quota is available for all current Google Cloud Platform customers to use right now.
Google first launched its Cloud ML catalogue back in early 2017 as a managed TensorFlow service and has seen its demand and popularity grow exponentially. Since that release, the team behind this technology has released many other features, like support for NVIDIA’s V100 GPUs, improvements to the hyperparameters feature, among other things.
These Google Cloud TPUs are still in beta mode but are available for training virtually any machine learning model. According to Google’s blog post, these TPUs enable “you to train a variety of high-performance, open-source reference models with differentiated performance per dollar. Or, you can choose to accelerate your own models written with high-level TensorFlow APIs”.
To add to the jewels already in Cloud TPUs crown, it recently came top in ImageNet’s Training Cost category of Stanford’s DAWNBench competition. In second place? Also Google’s Cloud TPU. At the current pricing, this is one of the cheapest options available in the market compared to all competitors at it’s level (like AWS).
Our take on this
This release by Google should help small to medium sized organizations take better advantage of machine learning capabilities. One of biggest positives of Google’s Cloud ML platform is that it handles all the hardware part – like the infrastructure, computation resources, scheduling, etc. on your behalf.
As a data scientist, this is a dream scenario where you can narrow down your focus on dealing with data and developing your models. If you are still wondering if Google’s platform is worth it for ML tasks, check out these 2 real-life examples of how it has produced better results than any other tool currently around. Incredible, isn’t it?
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!