ztachip is an opensource framework for Domain Specific Architecture.
Domain Specific Architecture defines the hardware/software architectures that can accelerate a particular class of applications very efficiently. The class of applications that ztachip accelerates today are AI and vision processing tasks.
ztachip has the full hardware implementation available in VHDL source code.
ztachip hardware can be deployed to FPGA hardware or custom ASIC.
ztachip is fully software programmable by using a special tensor programming paradigm.
Unlike many other AI hardware architectures, ztachip is flexible enough to run not just neural-network functions, but also a wide range of image processing such as image resizing, edge detection, image blurring, optical flow, harris corner feature extraction,...
Latest results based on SSD-MobiNet AI inference performance show ztachip to be 5.5x more computational efficient than Nvidia's Jetson Nano and 37x more computational efficient than Google's TPU edge.
This is a great YouTube video about why the future of computing is Domain Specific Architecture. Stanford Seminar - New Golden Age for Computer Architecture presented by John Hennessy, 2017 Turing Award Recipient / Chairman, Alphabet
To follow latest update/news of this project, please follow us on Twitter
For more information and download, visit github.com/ztachip/ztachip
DemonstrationAccelerate edge detection+objectDetection+HarrisCorner+MotionDetection at same time...
Accelerate MobiNet Image Classification.
Accelerate edge detection
The demo is running on DE10-NANO from terasic.com
Comments