Wider SIMD units suffer from low scalability of gather/scatter instructions that appear in sparse matrix calculations. We address this problem with an out-of-step pipeline which tolerates bank conflicts of a multibank L1D by allowing element operations of SIMD instructions to proceed out of step with each other. We evaluated it with a sparse matrix-vector product kernel for matrices from HPCG and SuiteSparse Matrix Collection. The results show that, for the SIMD width of 1024 bit, it achieves 1.91 times improvement over a model of a conventional pipeline.

The increase in complexity and depth of deep neural networks (DNNs) has created a strong need to improve computing performance. Quantization methods for training DNNs can effectively improve computation throughput and energy efficiency of hardware platforms. We have developed an 8-bit quantization training method representing the weight, activation, and gradient tensors in an 8-bit fixed point data format. The shared exponent for each tensor is managed dynamically on the basis of the distribution of the tensor elements calculated in the previous training phase, not in the current training phase, which improves computation throughput. This method provides up to 3.7 -times computation throughput compared with FP32 computation without accuracy degradation.

Metagenomic shotgun sequencing enables us to explore diverse DNA sequences from viruses, bacteria, and eukaryotic microbes in complex samples. As the continuous advancement of sequencing technology generates a massive amount of sequencing data, its overall computational complexity has become a major challenge for traditional database sequence comparison methods. Studies have shown that deep learningoriented methods have been widely adopted to solve many classification problems, including those in the bioinformatics field, and have demonstrated this method's accuracy and efficiency for analyzing large-scale datasets. The aim of this study attempts to investigate how deep learning (LSTM model) can be used to learn sequential genome patterns through virus detection from metagenomic data. This study provides three major contributions. First, we provide the background and steps for the task of DNA sequencing classification from data collection, preprocessing, and normalization. Second, we analyze the effect of sequence length on LSTM classification accuracy and split the raw sequencing data to proper subsequences to improve the outcome of virus detection. Third, to enhance both the classification accuracy and processing speed, we introduce the concept of discrimination function that enables prediction results for multiple subsequences results and accelerated these processes through GPU parallel computing. Two case studies of HCV and influenza detection were conducted to elaborate upon the accuracy and computational efficiency of our proposed approach. Our test result showed that the proposed LSTM model obtained similar pathogen detection accuracy to the conventional BLAST method with a speed that was about 36 times faster.

Energy efficiency is the most important factor in the design of wireless modem LSIs for mobile handset systems. We have developed an energy-efficient SIMD DSP for LTE-A modem LSIs. Our DSP has mainly two hardware features in order to reduce energy consumption. The first one is multiple VLIW configurations to minimize accesses to instruction memories. The second one is an advanced memory access unit to realize complex memory accesses required for wireless baseband processing. With these features, performance of our DSP is about 1.7 times faster than a base DSP on average for standard LTE-A Libraries. Our DSP achieves about 20% improvement in energy efficiency compared to a base DSP for LTE-A modem LSIs.