The CUDA core of NVIDIA GPUs had been one of the most efficient computation units for parallel computing. However, recent rapid developments in deep neural networks demand an even higher level of computational performance. To meet this requirement, NVIDIA has introduced the Tensor core in recent generations. However, their impressive enhancements in computational performance have newly brought high pressure on the memory hierarchy. In this paper, first we identify the required memory bandwidth in the memory hierarchy as the computational performance increases in actual GPU hardware. Through a comparison of the CUDA core and the Tensor core in V100, we find that the tremendous performance increase of the Tensor core requires much higher memory bandwidth than that in the CUDA core. Moreover, we thoroughly investigate memory bandwidth requirement over Tensor core generations of V100, RTX TITAN, and A100. Lastly, we analyze a hypothetical next-generation Tensor core introduced by NVIDIA through a GPU simulation, through which we propose an inter-warp multicasting microarchitecture that reduces redundant shared memory (SMEM) traffic during the GEMM process. Our evaluation shows that inter-warp multicasting reduces the SMEM bandwidth pressure by 33% and improves the performance by 19% on average in all layers of ResNet-152 and BERT-Large.

Mobile and edge devices become common platforms for inferring convolutional neural networks (CNNs) due to superior privacy and service quality. To reduce the computational costs of convolution (CONV) , recent CNN models adopt depth-wise CONV (DW-CONV) and Squeeze-and-Excitation (SE) . However, existing area-efficient CNN accelerators are sub-optimal for these latest CNN models because they were mainly optimized for compute-intensive standard CONV layers with abundant data reuse that can be pipelined with activation and normalization operations. In contrast, DW-CONV and SE are memory-intensive with limited data reuse. The latter also strongly depends on the nearby CONV layers, making an effective pipelining a daunting task. Therefore, DW-CONV and SE only occupy 10% of entire operations but become memory bandwidth bound, spending more than 60% of the processing time in systolic-array-based accelerators. We propose a CNN acceleration architecture called MVP, which efficiently processes both compute- and memory-intensive operations with a small area overhead on top of the baseline systolic-array-based architecture. We suggest a specialized vector unit tailored for processing DW-CONV, including multipliers, adder trees, and multi-banked buffers to meet the high memory bandwidth requirement. We augment the unified buffer with tiny processing elements to smoothly pipeline SE with the subsequent CONV, enabling concurrent processing of DW-CONV with standard CONV, thereby achieving the maximum utilization of arithmetic units. Our evaluation shows that MVP improves performance by 2.6 \( \times \) and reduces energy by 47% on average for EfficientNet-B0/B4/B7, MnasNet, and MobileNet-V1/V2 with only a 9% area overhead compared to the baseline.

Recurrent Neural Networks (RNNs) spend most of their execution time performing matrix-vector multiplication (MV-mul). Because the matrices in RNNs have poor reusability and the ever-increasing size of the matrices becomes too large to fit in the on-chip storage of mobile/IoT devices, the performance and energy efficiency of MV-mul is determined by those of main-memory DRAM. Therefore, computing MV-mul within DRAM draws much attention. However, previous studies lacked consideration for the matrix sparsity, the power constraints of DRAM devices, and concurrency in accessing DRAM from processors while performing MV-mul. We propose a main-memory architecture called MViD, which performs MV-mul by placing MAC units inside DRAM banks. For higher computational efficiency, we use a sparse matrix format and exploit quantization. Because of the limited power budget for DRAM devices, we implement the MAC units only on a portion of the DRAM banks. We architect MViD to slow down or pause MV-mul for concurrently processing memory requests from processors while satisfying the limited power budget. Our results show that MViD provides 7.2× higher throughput compared to the baseline system with four DRAM ranks (performing MV-mul in a chip-multiprocessor) while running inference of Deep Speech 2 with a memory-intensive workload.