Luba Tang is an internationally recognized expert on compiler optimization and virtual machine design. His research interests include both electronic system level (ESL) design, compilers, and virtual machines.
He was the chief programmer of Starfish DSP simulator; the original writer of Marvell iterative compiler; the software architect of the MCLinker project – the fastest and smallest system linker in the world; and the architect of the ONNC project – the first heterogeneous architecture compiler for AI.
He is the co-founder of the biggest and most professional engineer’s meet-up – Hsinchu Code Serf’s Meeting – in Taiwan. He also started the professional compiler company – Skymizer Taiwan Inc. – since 2013.
Anchors While Both Hardware and Software Are Changing in AI Semiconductor.
The global AI market has grown at 35.6% annual CAGR during forecast period of 2021-2026. It creates strong demands for the semiconductor industry.
AI semiconductor is forecast to be $7.1 billion by 2025, growing by a five-year CAGR of 25.29%. More than 3x faster than the global semiconductor market.
People said it’s a new golden age for computer architecture.
However, modern neural networks require huge computation power and it leads most AI chips are sort of heterogeneous multicore systems. The systems are composed of complex software systems on various hardware components. Conventional hardware/software codesign and electronic system level design come back to the center of development process.
In this talk, we will introduce the challenges coming from both hardware and software sides. We will discuss the reactions when a team is facing to the dynamic schedule and system requirement. In the end of the talk, we will introduce our solutions – ONNC – a system software addressing these great challenges.
Skymizer provides AI-on-chip system software subscription services, including compilers, calibrators, runtime systems, and various AI models (Model Zoo) to the complete source code of the basic application (application). Skymizer also provides customize software and relevant source code for the IC design company.
Company Products and Services
Open Neural Network Collection (ONNC) is a collection of system software for AI system-on-chip, including ONNX model zoo, TorchVision, Pytorch, compiler, calibrator, runtime, virtual platform, and FPGA. ONNC supports various AI frameworks, including Pytorch and ONNX, and supports rich accelerators, including CMSIS-NN, NVDLA, customized SIMD, and customized DLA.
We provide these services:
(1) engagement service – customize AI System Software for specific DLA
(2) subscription service – lastest release of ONNC source code
(3) maintenance – maintain a special version of ONNC
ONNC on NVDLA AI Starter Kit- Faster time to market tool for designing new AI chip https://skymizer.com/nvdla/
Tiny ONNC- Run PyTorch on MCU seamlessly