Embedded neural networks: Cadence’s latest DSP target
The new IP extends the Tensilica product portfolio further into the fast-growing vision/deep learning applications areas. It quadruples multiply-accumulate (MAC) performance compared to the previous generation Vision P5 DSP, targeting convolutional neural network (CNN) applications which are dominated by available MAC performance.
Compared to commercially available GPUs, the Tensilica Vision P6 DSP can achieve twice the frame rate at much lower power consumption on a typical neural network implementation, claims the company. For a wide range of other key vision functions, such as convolution, FIR filters and matrix multiplication, the Tensilica Vision P6 DSP increases performance by up to 4X by utilizing its improved 8-bit and 16-bit arithmetic. In addition, the new IP implements on-the-fly data compression to sharply reduce memory footprint and bandwidth requirements for demanding “fully connected” neural network layers.
Compatible with the Vision P5 DSP, this newest vision DSP offers an optional 32-way SIMD vector floating-point unit that includes the IEEE half precision standard (FP16). Floating-point performance capability is double that of the Vision P5 DSP, enabling easy use of existing floating-point neural network implementations.
”Cadence is investing heavily on advanced vision and deep learning,” said Chris Rowen, CTO for the IP Group at Cadence. “We are devoting intense efforts to discovering improved structures and training for neural networks, providing rich software environments for fast application development and offering breakthrough vision DSP architectures for embedded vision and learning deployment. The Tensilica Vision P6 design is the direct result of this investment and significantly raises the bar in both vision efficiency and scalability.”
The Tensilica Vision P6 DSP is based on the Cadence Tensilica Xtensa architecture, and combines flexible hardware choices with a library of vision/imaging DSP functions and numerous vision/imaging applications from our established ecosystem partners. It also shares the comprehensive Tensilica partner ecosystem for other applications software, emulation and probes, silicon and services and much more.
Visit Cadence at www.cadence.com
Related articles:
Licensible IP core accelerates neural networks
Xilinx invests in neural network startup
Neural network stick promises new smart applications, products
Qualcomm offers neural network SDK for Snapdragon processor
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
