Itching to hurl cash at machine learning? Google unveils TensorFlow Enterprise


Google has introduced TensorFlow Enterprise to the machine learning community, offering support, better performance, and managed services to those willing to pay.

The still in beta service seems to aim at companies that have already started using TensorFlow for their machine learning endeavours but have run into problems that usually go along with the use of open source in the enterprise. The main selling point however seems to be Google’s “unique position” to offer support and insight, since its team created and open-sourced the popular library.

One of the issues enterprises face is the higher than usual release pace of open source projects, which can’t always be followed if the project is integrated into a larger context. TensorFlow Enterprise therefore promises security patches and select bug fixes for certain TensorFlow versions “for up to three years”. 

Which versions Google will choose for that hasn’t been announced yet, but the support will feature on the company’s Cloud with fixes also available in mainline TensorFlow. The latter might lead to Enterprise users slightly influencing the direction of development, since the TensorFlow Enterprise website lists “prioritised requests” (for bug fixes and patches nonetheless) as one of the product’s benefits.

Since machine learning can be quite computationally expensive, the second leg of the Enterprise offering is quite simply scale. This comes in the form of Google Cloud services such as Deep Learning VMs, Deep Learning containers (those are also still in beta), and a couple of optimisations. The latter for example include a BigQuery reader to ingest data directly from Google’s data warehouse and some tweaks to the way TensorFlow Dataset reads data from Cloud Storage, which should add up to a speed up in data reading times by up to three times.

And if you’re still unsure if the new offering is meant to lure enterprise devs over to Google Cloud, a glance at the “white-glove service” that can be part of the offering might help. This “engineer-to-engineer assistance” can’t be bought directly but companies have to apply to be considered. To qualify, customers need to “have spent $500,000 annually or are willing to commit to a spend of $500,000 annually on Deep Learning VMs, Deep Learning Containers, or AI Platform Training and Prediction”. Which perhaps gives some idea of the resources necessary to run serious machine learning.

As it is usual with these Google offerings, “TensorFlow Enterprise is available at no additional cost”. The resources customers need to use to get the full experience however aren’t, so this should be kept in mind when signing up for the service.