At this year’s Embedded Vision Summit, Cadence Design Systems announced the Tensilica Vision P6 digital signal processor (DSP) targeted at convolutional neural network (CNN) applications as well as advanced imaging and classic imaging. We spoke with Chris Rowen, CTO of the IP group at Cadence and an IEEE Fellow. Rowen worked at MIPS, Silicon Graphics, and Synopsys before founding his own company, Tensilica, in 1997. That company was acquired by Cadence 3 years ago.
“Soon all those IoT sensors will create so much data we will need to massively filter it just to get it down to the amount of today’s Big Data,” said Rowen.
The Tensilica Vision P6 DSP
The Tensilica brand is best known for its customizable microprocessor core, the Xtensa processor architecture and the Xtensa Processor Generator. The Xtensa architecture is one of the most popular licensable processor architectures, and is shipped in products ranging from sensors to supercomputers. Architects use the generator to select and describe all the features needed in cost-, energy-, and performance-sensitive applications; the generator automatically builds the RTL, the compilers, the models, and even the RTOS for that optimized processor and memory system.
The Xtensa platform is widely used to produce DSPs tuned to run complex algorithms in imaging, video, and computer vision. The DSP has new instructions, increased math throughput, and other enhancements. Cadence has used this same environment to create many of the leading DSP standards in audio, wireless, and vision processing. The Vision DSP family has been widely adopted in mobile phones, games, automotive, security, and IoT platforms. Xtensa has a library of vision DSP functions and numerous imaging applications from an established ecosystem of Cadence partners. It also shares the Tensilica partner ecosystem for other applications software, hardware, and services. The image processing libraries and robust development environment promise a power-efficient implementation of neural nets in a shortened development cycle.
“Anything that can written in C can run on a P6,” said Rowen.
One key trend in vision systems is the rapid adoption of deep neural networks, a fundamental computation method that relies on automatic training, rather than manual programming, to create sophisticated image recognition systems. Deep neural networks are remarkably flexible, but also computationally intensive. Cadence has evolved its Vision DSP family to provide exactly the kind of horsepower, especially on large multiply-rich convolution operations, needed for neural networks. In internal imaging and computer vision benchmark tests, the latest Vision DSP, the Vison P6, has been proven to increase performance by up to 4 times the speed of last year’s Tensilica Vision P5 DSP, on a range of key tasks, including neural network-based vision.
How Does It Work?
The increased use of computers resulted in an increased need for representing analog signals as digits. Hence DSP technology was born. To digitally analyze a signal, it must be converted from analog to digital values. This is done by sampling the signal at various intervals and partitioning it into equivalence classes. Then the signal is quantized by replacing the analog signal with representative but not exact signal approximated by values from a finite set of numbers. The Tensilica Vision P6 DSP follows this process, as well, by calculating neural net values with 16-bit instead of 32-bit precision. The resulting neural net is less accurate as a result, but usually this does not make a difference to most applications.
“Someday neural nets may create computer assists for elders. Hearing aids will have tiny cameras, be able to recognize people, and whisper their names in your ear,” said Rowen.
According to Tractica’s research, in most cases, the revenue from hardware will be many times the amount spent on software. The Tensilica Vision P6 DSP has a very large potential payback if it is designed into devices where both power consumption and recognition performance are critical, such as cars or mobile phones. In our recently published Deep Learning for Enterprise Applications report, we forecast that spending on hardware as a result of deep learning projects will grow from $436 million in 2015 to $41.5 billion by 2024, while software spending will only grow to $10 billion during that same period.