Aoha Light
AI semiconductor-related software that accelerates, lightweights, and reduces the power consumption of AI systems, optimizes for various semiconductors, supports a range of AI frameworks and models, including LLMs, and enables validation during both hardware and software development stages
Aoha's Technology
-
Aoha's post training quantization (PTQ) has high accuracy that can exceeds quantized-aware training (QAT)
-
Automatic optimization of combination with lower INT like INT 4, 2 and 1 in addition to INT 8
-
The table below is a reference example (as of 2023). The calibration data used are random for reference, and in actual operations, further accuracy improvements can be achieved in this aspect as well
Accuracy Simulator
-
Accurate simulation at HW-development stage to support SW-defined development
HW Design Simulator
-
HW-design simulation using workload models after optimization with EDA tools
Model | FP32 | INT8 | INT8+INT4 |
|---|---|---|---|
Classification: MobileNetV3 | 74.04 | 73.64 | 73.37 |
Segmentation: DeepLabV3+ | 79.6 | 79.6 | 79.6 |
Detection: Faster R-CNN | 35 | 35 | 34.4 |