Skymizer, a leading AI system software provider, today announced that it had achieved top-tier performance in the MLPerf Tiny v1.1 benchmarks. Skymizer's submissions used both hardware/software co-optimization and pure compiler optimization, and both approaches achieved impressive results.
In the hardware/software co-optimization case, Skymizer worked with hardware vendor Nuvoton to develop a custom microcontroller that was optimized for AI inference. This microcontroller, the M467HJHAN, operates at a frequency of 200MHz and has only 512KB of SRAM. Despite these limitations, Skymizer was able to achieve top-tier performance on the MLPerf Tiny v1.1 benchmarks in the entire ARM Cortex-M4 segment.
In the pure compiler optimization case, Skymizer used its ONNC compiler to optimize the performance of AI models on typical microcontrollers. ONNC achieved a 1.24x speed-up on average compared with the open-source compiler. ONNC's performance gains were particularly impressive on the IC (Image Classification) benchmark, where it achieved a 1.31x speed-up. This was due to ONNC's ability to split large tensors into smaller sizes that fit within the memory constraints of the microcontroller. ONNC also achieved a significant speed-up on the VWW (Visual Wake Word) benchmark, where it achieved a 1.52x speed-up. This was due to ONNC's ability to handle sparse tensors. ONNC's unique optimization algorithm can remove sparsity in the tensor and keep the precision of the neural network at the same time.
Skymizer's results demonstrate the company's leadership in AI system software. Skymizer's software stack is re-targetable to a wide range of hardware platforms, and it can achieve high performance on even the most resource-constrained devices. ONNC has been applied in many neural network accelerators, including recommendation systems for data centers, general accelerators for data centers, smart phones, surveillance systems, and microcontrollers. Skymizer is committed to providing its customers with the best possible tools and services, and ONNC is a crucial part of that commitment.
For more information, please visit https://skymizer.com/ and request a demonstration.
For the best ARM Cortex-M4 MCU solution, please visit https://www.nuvoton.com/.