Memristor based computing-in-memory chips have shown the potentials to accelerate deep neural networks with high energy efficiency. Due to the inherent filament-based conductive mechanism of the memristor, the reading and writing noises are hard to eliminate. Besides, the precision of the large-scale memristor array is still limited. However, when the noise of the memristor is large, the existing training methods to reduce the accuracy loss of memristor based computing-in-memory chips will face challenges. Hence, we proposed the array-level boosting method with spatial extended allocation to reduce the accuracy loss induced by the limited precision and large noises. To optimize the spatial allocation number of each layer in the neural network, the greedy spatial extended allocation algorithm is also proposed. The image processing and classification tasks are demonstrated based on fabricated 32 × 128 memristor arrays to valid the performance of the proposed method. The chip-in-loop results show that the recovered accuracy of ResNet-34 on CIFAR-10 with array-level boosting method is 92.3%, which is closed to software-based accuracy of 93.2%.

基于忆阻器的内存中计算芯片已显示出以高能效来加速深度神经网络的潜力。由于忆阻器固有的基于细丝的导电机制，因此难以消除读写噪声。此外，大规模忆阻器阵列的精度仍然受到限制。然而，当忆阻器的噪声很大时，减少基于忆阻器的基于内存的计算芯片的精度损失的现有训练方法将面临挑战。因此，我们提出了一种具有空间扩展分配的阵列级增强方法，以减少由于精度有限和噪声较大而导致的精度损失。为了优化神经网络中每一层的空间分配数，还提出了贪婪空间扩展分配算法。基于制造的32×128忆阻器阵列演示了图像处理和分类任务，以验证所提出方法的性能。芯片在环测试结果表明，采用阵列级增强方法，CIFAR-10上ResNet-34的恢复精度为92.3％，接近基于软件的93.2％的精度。