+886 2 8797 8337
12F-2, No.408, Ruiguang Rd., Neihu Dist., Taipei City 11492, Taiwan
Center of Innovative Incubator R819, No. 101, Section 2, Kuang-Fu Road, Hsinchu, Taiwan
In this era of AI-driven Enlightenment, the tasks that AI can perform on smartphones are diverse. For example, photo enhancement, translation, music identification, etc. An AI chip for a smartphone needs to be able to handle such diverse application scenarios. Furthermore, new and more powerful AI models are always being invented, and this leaves outdated AI chips behind. ONNC enables AI chips to be easily adopted in frequently evolved AI models by automatically compiling them into the chip’s machine code. This makes it possible for AI chip providers to deliver more more reliable and sustainable products and embrace broader markets.
The development time of an AI chip is at least 18 months, long enough for AI applications to evolve several new generations. For example, YOLO, the most popular object detection model, evolves from v4 to v7 in less than two years. Yolo v7 uses newer architecture and surpasses all previous versions and even other object detection models in terms of both latency and accuracy. AI chips that cannot support such state-of-the-art models are hardly survival in the market. Only a forward-compatible AI chip can prevent itself from dead on arrival.
ONNC is a retargetable toolchain that boosts your development of AI system software. ONNC Compiler tackles new models with "progressively lowering" technology and transform newly-innovated deep learning layers into the target instruction set. ONNC Compiler also features with off-the-shelf CPU fallback mechanism for popular CPUs like ARM and RISC-V.
Another trend in the development of deep learning models is their size. The number of parameters in the latest models is increasing as their accuracy improves. ONNC optimizes inference performance via various algorithms, such as software pipelining, scheduling, memory allocation. Such optimization algorithms greatly help you prepare your AI chip for tomorrow's challenge.
With above features, your AI chip can cover more deep learning models, not only for today, but for models from tomorrow. ONNC compiler prolongs the lifespan of your AI chips and make them ready for more diverse AI applications on smartphones.
ONNC is a modularized and retargetable toolchain for development of AI system software. With the helps from ONNC, you can:
Increase Product Competitiveness
ONNC helps AI chip vendors increase model coverage which means increase the competitiveness of their IC product. ONNC also enhances performance of AI chips and increase the value they can provide.
Shorten Time to Market
As deep learning models evolve, AI chip vendors need to continuously update their support. With ONNC, deep learning models can be automatically compiled into your AI chips’ machine code and heterogeneous architectures are well supported. Such features reduces your R&D risk and help you shorten time to market.