New Edge AI processor accelerates key workloads while offering flexibility for changing AI needs
SiFive, Inc., the leading provider of commercial RISC-V processor IP and custom silicon solutions, and Osaka R&D venture ArchiTek Corporation today announced ArchiTek’s first in-house developed AI processor, AiOnIc®, featuring the ArchiTek Intelligence® Pixel Engine (aIPE), and SiFive E3-Series RISC-V processor Core IP. The ArchiTek AiOnIc® chip is developed to demonstrate an edge AI concept of processing sensor data in real-time to create meta-data for reduced latency of AI inference processing while offering increased security and privacy.
“The ArchiTek AiOnIc® processor featuring ArchiTek aIPE offers support for different algorithms by dynamically rebalancing hardware use, combining the advantages of CPUs, GPUs, and dedicated LSIs,” said Shuichi Takada, CEO & CTO of ArchiTek. “SiFive E3-Series Core IP enabled us to hit our performance, power, and area targets, while SiFive Core Designer and SiFive partner DTS-Insight helped reduce our development time.”
“Edge AI is a massive opportunity for new architectures that can deliver low-power, area-optimized designs,” said Chris Lattner, President of Engineering & Product, SiFive. “The power of SiFive RISC-V Core IP combined with AI accelerators and other compute elements enable the flexibility and programmability needed for shifting AI applications.”
The ArchiTek AiOnIc® architecture was designed for use in Edge AI applications, offering a low-power, high-efficiency chip that enables fanless designs, suitable for harsh environments. The aIPE architecture reduces processing time for Simultaneous Localization and Mapping (SLAM), reducing processing time to 1/20th of the time seen for general-purpose CPUs, and improves the operating speed of perceptual computing using OpenPose for human pose estimation by 3.8X vs. GPUs. The dynamic configuration capabilities of the AiOnIc chip enable various algorithms to be supported and can be used for autonomous vehicles, cybersecurity systems, robotics, and security cameras.
The ArchiTek AiOnIc® processor includes built-in engines for signal processing, general-purpose sort, multi-functional DMA, inverse matrix operation, fast Fourier transforms (FFT), general-purpose matrix multiply-add, and a high-performance general-purpose GPU. The processor is built using TSMC 12nm N12FFC process technology, with a die size of 4.5mm2. The chip features the SiFive E3-Series processor core, a 32-bit RISC-V processor with 5-stage pipeline configured to include hardware floating-point unit. The AiOnIc® processor also features an 8MB SRAM, DDR4, Ethernet, UART, I2C, and other standard interfaces. SiFive Core IP has been adopted worldwide and is available in products by leading technology companies including BeagleBoard, Bouffalo Lab, Coherent Logix, eTopus, Innovium, FADU, Huami, MicroChip, Synaptics, Qualcomm, Samsung, and Tynker with BBC Learning.
The Architek AiOnIc® proof of concept silicon demonstrates the versatility and programming capabilities of the chip to address the requirements of modern AI solutions and opens the door to deliver the promise of AR, VR, smart cities and countless other 21st century use cases by proliferating AI on the Edge.
