Kneron KL530 Targets Automotive
November 30, 2021 - Author: Aakash Jani
With its newest AI processor, Kneron is shifting gears to the automotive market. The KL530 delivers 0.5 trillion INT8 operations per second (TOPS) and supports INT4 format for greater throughput. Kneron is expanding quickly, boasting more than 200 employees and $100 million in funding. It has achieved moderate success in AI-enabled IoT devices and is now expanding into new markets. With the KL530, it plans to initially pursue aftermarket ADAS vendors and later sell directly to carmakers.
The KL530 debuts the company’s new deep-learning-accelerator (DLA) architecture. That design operates similarly to previous Kneron products but adds a twist: it increases throughput for quantized layers by handling INT4 data in addition to INT8. The company claims the KL530 reduces inference times by up to 66% with mixed INT4 and INT8 networks compared with standalone INT8 networks. On networks such as ResNet-50, EfficientNet Lite, and MobileNet v2, the new SoC maintains nearly the same accuracy for mixed versus INT8-only models, varying by less than one percent.
INT4 support isn’t the only feature new to the KL family; the KL530 also integrates an Andes D25F CPU for AI coprocessing. That RISC-V core comes with the P extension, supporting packed SIMD instructions. Andes uses the extension, which it helped develop, for efficient audio, media, and image processing. The D25F helps the DLA by performing other activation functions and postprocessing. It works in lockstep with the KL530’s Cortex-M4, which handles system control.
Subscribers can view the full article in the Microprocessor Report.