The VPU report
Posted by Robert Dow on January 29th 2017 |
This report covers video processor units from four leading suppliers in the field. The application for such processors ranges from sophisticated IoT devices to security systems, license plate registration, autonomous vehicles and much more. The functionality of these devices depends on computational photography algorithms such as super resolution, anti-shake and continuous focus as well as on vision algorithms to implement face detection, gesture based user interfaces, depth perception and image classification.
With such a range of application, performance requirements, price and environmental constraints, its challenging to make comparisons of one firm’s products to another, especially as some are SoC level products while others are IP (which may be announced but not yet in silicon production). This should be borne in mind when comparing raw performance data.
A signature feature of this market is that it is in rapid flux; most of the suppliers started in the early 2000s and have had to be very nimble to stay in contention. The report summarizes the evolution of their products over time, giving an insight into how the market has developed.
For edge devices, two broad trends have emerged over the last couple of years as suppliers reacted to the changing market dynamics. First with a pivot back to programmability as it became clear that the range of applications now extended as much to vision as to computational photography and then to architectures optimized for neural network inference as this technology gained acceptance for classification and other tasks. The implications of these have not yet fully played out but it is already clear that mapping two-dimensional data through a variety of multi-dimensional formats onto the rather limited number of available compute structures is key to efficiency, making the memory access subsystem as important as the processor itself.
For cloud applications, where training as well as a non-overlapping set of functions are required, a new class of device is starting to emerge. Prefigured by Nvidia's "datacenter GPU", this is the logical endpoint of the GP/GPU movement and is set to become a heavily contested space, with multiple entrants coming over the next few months.
We have not attempted a shipment or market share analysis, but may do that in future editions as data becomes available.