Real-time digital video coding became a mandatory feature in current consumer electronic devices due to the popularization of video applications. However, efficiently encoding videos is an extremely processing/energy-demanding task, especially at high resolutions and frame rates. Thus, the limited energy resources and the dynamically varying system status (such as workload, battery level, user settings, etc.) require energy-efficient solutions capable to support run-time energy-quality scalability. In this work, we present an energy-quality scalable SAD Unit hardware architecture for the HEVC intra-frame prediction targeting real-time processing of UHD 8K (7680 × 4320) videos at 60 frames per second. Approximate computing is used to provide energy-quality scalability by employing configurable imprecise operators. The proposed Energy-Quality scalable architecture supports four operation points: precise computing, and 3-bit, 5-bit or 7-bit imprecision. When implemented in a 45-nm technology using Nangate standard cells library and running at 269 MHz, the proposed architecture consumes from 8.42 to 7.38 mJ to process each UHD 8K frame, according to the selected imprecision level. As a drawback, the coding efficiency (measured in BD rate) is reduced from 0.28 to 1.72%. Compared to the related works, this is the only intra-frame prediction SAD unit able to provide energy-quality scalability.

中文翻译：

---

使用优化和可配置的不精确加法器的UHD 8K能量质量可扩展的HEVC帧内预测SAD单元硬件

由于视频应用的普及，实时数字视频编码已成为当前消费电子设备中的强制性功能。但是，有效地对视频进行编码是一项极其处理/耗能的任务，尤其是在高分辨率和帧频下。因此，有限的能源和动态变化的系统状态（例如工作负载，电池电量，用户设置等）需要能支持运行时能源质量可伸缩性的节能解决方案。在这项工作中，我们为HEVC帧内预测提供了一种能量质量可扩展的SAD单元硬件体系结构，以每秒60帧的速度对UHD 8K（7680×4320）视频进行实时处理。近似计算用于通过采用可配置的不精确运算符来提供能量质量可伸缩性。提议的能源质量可伸缩体系结构支持四个操作点：精确计算以及3位，5位或7位不精确度。当使用Nangate标准单元库在45纳米技术中实现并运行在269 MHz时，根据所选的不精确度，所提议的体系结构将消耗8.42至7.38 mJ来处理每个UHD 8K帧。缺点是，编码效率（以BD速率测量）从0.28％降低到1.72％。与相关工作相比，这是唯一能够提供能量质量可伸缩性的帧内预测SAD单元。根据所选的不精确度，以38 mJ处理每个UHD 8K帧。缺点是，编码效率（以BD速率测量）从0.28％降低到1.72％。与相关工作相比，这是唯一能够提供能量质量可伸缩性的帧内预测SAD单元。根据所选的不精确度，以38 mJ处理每个UHD 8K帧。缺点是，编码效率（以BD速率测量）从0.28％降低到1.72％。与相关工作相比，这是唯一能够提供能量质量可伸缩性的帧内预测SAD单元。