Convolution engine

Abstract

This paper focuses on the trade-off between flexibility and efficiency in specialized computing. We observe that specialized units achieve most of their efficiency gains by tuning data storage and compute structures and their connectivity to the data-flow and data-locality patterns in the kernels. Hence, by identifying key data-flow patterns used in a domain, we can create efficient engines that can be programmed and reused across a wide range of applications. We present an example, the Convolution Engine (CE), specialized for the convolution-like data-flow that is common in computational photography, image processing, and video processing applications. CE achieves energy efficiency by capturing data reuse patterns, eliminating data transfer overheads, and enabling a large number of operations per memory access. We quantify the tradeoffs in efficiency and flexibility and demonstrate that CE is within a factor of 2-3x of the energy and area efficiency of custom units optimized for a single kernel. CE improves energy and area efficiency by 8-15x over a SIMD engine for most applications.

HEVC hardware vs software decoding: An objective energy consumption analysis and comparison

“…P Other corresponds to the remaining elements present in the SW video decoding platform. Finally, we compare the SW video decoding energy to that of the HW video decoding by calculating the ratio between them, using Equation (11).…”

“…We then assess the impact ofP Other , which corresponds to the power consumed related to the remaining elements present in the platform, such as memory, on the global platform power consumption using Equation (8). Finally, we compare the SW video decoding energy consumption to that of the HW video decoding by calculating the ratio between them,r sw/hw , using Equation (11). For that, Kimono video test sequence, which possesses the characteristics given in Table II (resolution is 1080p), is decoded as an example.…”

“…Table VI depicts the ratio between the HEVC SW and HW video decoding energy, calculated using Equation (11), as a function of the video resolution. Compared to the state-ofthe-art study where it is widely accepted that the dedicated processors outperform the GPPs by around 1000× in terms of energy efficiency, for low video resolutions (720p and 1080p), from the operating system level point of view, the gap is not as high as that of the state-of-the-art work (a ratio of less than 3×).…”

“…In state-of-the-art work, it is widely accepted that the dedicated processors outperform the general purpose processors (GPPs) by around 1000× in terms of energy efficiency [10]. This is achieved thanks to the use of specialized processing units [11] which eliminate the power consumption related to instruction decoding and control logic characterizing GPP [12]. As a consequence, most modern smartphones are equipped with an HDIP that consumes less energy with respect to the real-time video decoding constraint [13] [14].…”

Modelling, exploration and optimization of hardware accelerators for deep learning applications

“…Accelerators such as CNP [45], Convolution Engine [46], Neuflow [47] and TPU [5] have introduced customized logic by leveraging high parallelism and flexibility to efficiently map convolution operations to hardware. Google also introduced TPU Edge [48] processor for inference applications at the edge.…”