Post has attachment
We have a new GDG Cloud meetup on the September 29th. The topic of the day: Learn all about the Cloud Machine Learning offerings and Tensorflow.
Add a comment...

Post has shared content
Tensor Processing Units (TPUs)
I'm very excited that we can finally discuss this in public. Today at Google I/O +Sundar Pichai revealed the TPU (Tensor Processing Unit), a custom ASIC that Google has designed and built specifically for machine learning applications. We've had TPUs deployed in Google datacenters for more than a year, and they are an order of magnitude faster and more power efficient per operation than other computational solutions for the kinds of models we are deploying to improve our products. This computational speed allows us to use larger, more powerful machine learned models, expressed and seemlessly deployed using TensorFlow (tensorflow.org) into our products, and to deliver the excellent results from those models in less time.

TPUs are used on every Google Search to power RankBrain (https://en.wikipedia.org/wiki/RankBrain), they were a key secret ingredient in the recent AlphaGo match against Lee Sedol, they are used for speech and image recognition, and they are powering a growing list of other smart products and features.

+Norm Jouppi and the rest of the team that developed this ASIC did a fabulous job, and it's great to see it discussed in public!

Blog post:
https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chip.html

Link to the part of the keynote where Sundar discusses TPUs:
https://www.youtube.com/watch?v=862r3XS2YB0&feature=youtu.be&t=7300

WSJ article:
http://www.wsj.com/articles/google-isnt-playing-games-with-new-chip-1463597820

Edit: Added a link and some text.
Add a comment...

Post has attachment
DataFlow Java SDK 1.5.0 is release. Here are some highlights:

This release done some refactoring in preparation to the move to Apache Beam (incubating).

* Enabled an indexed side input format for batch pipelines executed on the Google Cloud Dataflow service. Indexed side inputs significantly increase performance for View.asList, View.asMap, View.asMultimap, and any non-globally-windowed PCollectionViews.
* Upgraded to Protocol Buffers version 3.0.0-beta-1. If you use custom Protocol Buffers, you should recompile them with the corresponding version of the protoc compiler. You can continue using both version 2 and 3 of the Protocol Buffers syntax, and no user pipeline code needs to change.
* Added ProtoCoder, which is a Coder for Protocol Buffers messages that supports both version 2 and 3 of the Protocol Buffers syntax. This coder can detect when messages can be encoded deterministically.
* Added withoutResultFlattening to BigQueryIO.Read to disable flattening query results when reading from BigQuery.
* Added BigtableIO, enabling support for reading from and writing to Google Cloud Bigtable.
* Improved CompressedSource to detect compression format according to the file extension. Added support for reading .gz files that are transparently decompressed by the underlying transport logic.

Improvements to the service:

Scalability and performance improvements available when using Cloud Dataflow SDK for Java version 1.5.0:
* The service now scales to tens of thousands of initial splits when reading from a BoundedSource. This includes TextIO.Read, AvroIO.Read, and BigtableIO.Read, among others.
* The service will now use Avro instead of JSON as a BigQuery export format for BigQueryIO.Read. This change greatly increases the effiency and performance when reading from BigQuery.
Add a comment...

Post has attachment

Post has attachment
Next week I will be filling in for a Google Engineer at the Kubernetes & CoreOS roadshow. It's free, signup now! #gde
Add a comment...

Post has attachment
Finally I can make a distinction between the services of all my open tabs on the Google Cloud Platform console. They changed the icon from the general GCP icon to the icon of the individual service.

This is a real life save for someone like me, who has like a bazillion tabs open to each service. Great!
Photo
Add a comment...

Post has attachment

Post has attachment

The DataFlow Python SDK is available on GitHub. It's *local* pipeline execution only for now, but you finally can try the DataFlow/Beam semantics in Python (till now it was only available in Java) on your own machine.
Add a comment...

Post has attachment
DataFlow SDK becoming an Apache project? This is kind of a big deal I would say.
Add a comment...
Wait while more posts are being loaded