News ▶ General

Deep learning capabilities

MIT recently unveiled a chip they claim can give smartphones, tablets and even IoT devices the kind of horsepower which allows them to perform deep learning roles.

The 168-core monster aims to shift the processing point from the cloud to the local device. 

Currently, most devices are used as not much more than data collectors from their sensors. The could then performs the analysis after the data has been uploaded. Problems with this include network connectivity, latency, congestion, security and power continuity. These ultra-powerful chips can offload may of the cloud roles to the local device.

As well as this awesome processing power, MIT announced breakthroughs in its energy efficiency - a huge concern for devices with no permanent wired power supply.

Neural networks

The prevailing architecture over the past few years which has driven AI to the stage we see today is founded on the neural network. Part hardware, part software, this system was thrashed out in academia until making its appearance in the latter part of the last century. From the 1970's, the work remained largely theoretical because there was nothing like the raw processing power available to make them actually run. The new MIT chip has 168 cores - or processing units - which in itself means it is already geared up to support them.

One of the novel features of the chip is its customized circuitry which spreads the work load. This facilitates crunching through more data in the shortest time possible before fetching more from main memory.

desert tree stumps

Vivienne Sze, who's group at  MIT's Department of Electrical Engineering and Computer Science powered the chip, said:

Deep learning is useful for many applications, such as object recognition, speech and face detection. Right now, the networks are pretty complex and are mostly run on high-power GPUs. You can imagine that if you can bring that functionality to your cell phone or embedded devices, you could still operate even if you don’t have a Wi-Fi connection. You might also want to process locally for privacy reasons. Processing it on your phone also avoids any transmission latency, so that you can react much faster for certain applications.

Here's Mike Tyka's TED talk explaining the power of the neural network:

More nodes means more power

The architecture is based on a machine learning convolutional neural network. In this design, the cores can communicate with each other directly without being forced to access the main memory. This is loosely based on the way the human brain is thought to work, and is essential when multiple nodes are processing the same data.

The last piece in the puzzle is a new scheduling system which efficiently allocates tasks across the cores. Without this low-level co-ordination, the automatic re-configuration required for such tasks as image recognition wouldn't be anywhere near as efficient.

Google, IBM, Microsoft, Qualcomm, Apple ...  all the big names in the computer industry are investing heavily in AI and betting big on its central role in every aspect of our lives in the future.

Part funded by DARPA, the MIT chip is already 10 times as efficient as a state-of-the-art mobile GPU.