EdgeCortix has made a strategic shift from selling AI intellectual property (IP) to selling its own edge-AI inference chips for line-powered systems. The new die, dubbed Sakura, started as a test chip, but the company says customer interest convinced it to offer the chip as a product. The chips come mounted on one of two cards: a dual-M.2 (M-key) card and a low-profile PCIe card. Based in Japan, EdgeCortix has supplemented its $13,5 million in total funding with revenue from FPGAs and ASICs that employ its soft IP. It targets perception – vision, lidar, and related technologies – are aimed at transportation, augmented/virtual reality, industry, smart cities, and drones.
Sakura, revealed first at the recent Linley Spring Processor Conference, implements the company’s dynamic neural accelerator (DNA) engine, adding on-chip SRAM, two LPDDR4X ports, and I/O. The chip has no host CPU, so it operates under the control of an external host. Sakura has a maximum performance of 40 TOPS; on ResNet-50, it achieves 0,4 ms latency at 4,7 W, yielding 533 inferences per second per watt (IPS/W).
Sakura combines six engines for standard 3D convolutions, six engines for 2D depth-wise convolutions, and three vector engines for activations and other miscellaneous operations. Standard convolutions combine all three colour channels, whereas depth-wise convolutions separate the channels, combining results later to reduce computing operations. To boost utilisation, the Mera tools can convert between standard and depth-wise convolutions to keep all engines occupied. Together, the engines are capable of 25 000 MACs per clock cycle.
Within a vector engine, dedicated blocks implement activation functions. Each block is configurable for a range of variants in a family of activation functions. Although this approach yields better performance, it’s less flexible than a fully programmable engine. A 20 MB on-chip SRAM can hold weights and activations or act as a scratchpad. The runtime software reconfigures the network-on-a-chip (NoC) on the fly, reassigning engines as part of EdgeCortix’s focus on achieving high hardware utilisation.
RFiber Solutions has added this group to its line card.
| Tel: | +27 12 667 5212 |
| Email: | [email protected] |
| www: | www.rfibersolutions.com |
| Articles: | More information and articles about RFiber Solutions |
© Technews Publishing (Pty) Ltd | All Rights Reserved
printer friendly version