PulseDL: A reconfigurable deep learning array processor dedicated to pulse characterization for high energy physics detectors
PulseDL: A reconfigurable deep learning array processor dedicated to pulse characterization for high energy physics detectors
复制标题
PulseDL：一种可重新配置的深度学习阵列处理器，专用于高能物理探测器的脉冲表征
DOI：
10.1016/j.nima.2020.164420
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Chen Junling
中科院分区：
文献类型：
--
作者：
Ai Pengcheng;Wang Dong;Huang Guangming;Shen Fan;Fang Ni;Xu Deli;Wang Hui;Chen Junling
神经网络模型在高能物理的在线和现场数据分析中显示出良好的速度和精度。在这份报告中，我们讨论了一个多功能的神经计算芯片称为PulseDL脉冲特性的测量。为了平衡功耗和性能，我们在PulseDL中采用了外部RISC CPU和处理引擎的结构。寄存器传输级数字逻辑的设计重点是线程级的并行性。在硬件方案的基础上，我们共同设计了网络架构，以最大限度地利用片上资源。将网络的卷积、反卷积和全连接矩阵乘法等功能嵌入硬件中，并在运行时进行重构。该芯片采用GSMCR 013 130 nm工艺流片，面积4.9 mm× 4.9 mm，工作频率至少为25 MHz，内核电压为1.2 V。通过布局后的模拟测量，芯片的峰值功率效率估计为每秒每瓦12千兆运算。
Neural network models show promising speed and accuracy for online and on-site data analysis in high energy physics. In this report, we discuss a multi-functional neural computing chip called PulseDL for measurement of pulse characteristics. We adopted a structure with outside RISC CPU and processing engines in PulseDL for balanced power and performance. Digital logic at register transfer level was specially designed with emphasis on thread level parallelism. Based on the hardware scheme, we co-designed the network architecture to best utilize the on-chip resources. Convolution, deconvolution and fully-connected matrix multiplication of the network were fitted into the hardware with reconfiguration during runtime. The chip has been taped out under the GSMCR013 130 nm process, with 4.9 mm× 4.9 mm area, at least 25 MHz working frequency and 1.2 V core voltage. Measured by post-layout simulations, the peak power efficiency of the chip was estimated to be about 12 giga operations per second per watt.