Movidius has just launched the Fathom Neural Compute Stick, a computer dedicated to using Deep Learning training on ultra-mobile, low-power, platforms such as drones other embedded applications where every milliwatt of power, and every gram of weight, counts.
The Fathom Neural Compute Stick uses the company’s Myriad 2 chip (model MA2450), a superbly power-efficient vision processor (VPU) which we covered back in 2014, and the USB stick format provides a casing for the tiny motherboard, memory (512MB) and everything else required to form a complete computer. In practical embedded A.I applications, USB 3.0 would provide enough bandwidth for known workloads.
At a very high level, there are typically two phases in deep learning A.I: Training and Inference. Training requires huge computing resources and storage. For example, it’s possible to train computers to recognize a certain number of “object types” (cars, people, cats…) by exposing them to enormous amounts of images that contain these objects. The result of the training only takes a fraction of the space required for the training material itself.
Once the training has been done and saved. The training is then re-used in an application. That phase is called Inference. Inference still requires the machine to process vision, but instead of trying to learn more on the spot, it will use the knowledge acquired during training, thus requiring fewer resources and storage. In turn, needing fewer resources is necessary to avoid having to communicate with a cloud.
That said, Inference still requires computing capabilities that evade the overwhelming majority of drones and other embedded systems today. That’s where Fathom comes into play because this stick computer will give the extra “brain” capacity required for the task at hand, and will do so in a power envelope (sub 1W), which won’t be an overall game-changer for the platform.
During our discussion with Movidius’ CEO, Remi El-Ouazzane, he was confident that his platform can be 5X to 10X more power-efficient than the competition.
For example, Fathom allows the construction (or extension) of drones that can then become more autonomous. They could become smart enough to spot a flat area large enough to land safely. They could avoid collisions, and be more “aware” of their general surrounding and take proper action.
In fact, any computerized platform equipped with a camera and a USB port could benefit from Fathom. Surveillance cameras or rescue robots could now spot people, recognize who they are. Movidius can’t comment on 3rd party products which will use its platform, but the company is known to work very closely with Google.
At the moment, Movidius is sampling products to OEMs to accelerate the development or finalization of applications. After that, we know that the company intends to position Fathom aggressively against competing platforms, and it seems that Movidius will easily have the latitude to do so.