{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,6,26]],"date-time":"2026-06-26T09:48:29Z","timestamp":1782467309358,"version":"3.54.5"},"reference-count":23,"publisher":"MDPI AG","issue":"7","license":[{"start":{"date-parts":[[2026,6,26]],"date-time":"2026-06-26T00:00:00Z","timestamp":1782432000000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"name":"Guangdong S&amp;T Program","award":["2024B0101010003"],"award-info":[{"award-number":["2024B0101010003"]}]}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Computers"],"abstract":"<jats:p>It is widely acknowledged that General Matrix Multiplication (GEMM) serves as a foundational kernel across numerous application domains. Complex numbers exhibit distinctive mathematical properties that enable their widespread adoption across engineering computing scenarios, including signal processing and signal transformation. This study investigates high-efficiency CGEMM, namely, complex-valued GEMM, for NPU hardware, broadening the application scope of NPUs beyond mainstream low-precision AI computation workloads. The major contributions of this study are as follows: (i) numerical precision and hardware utilization of the 3M and 4M decomposition schemes on Ascend NPUs are analyzed, and the 4M method is selected as the preferred CGEMM implementation under our tested hardware constraints to fit the bandwidth limitations of modern accelerators for both precision-sensitive and performance-critical matrix computation scenarios; (ii) a complete high-performance CGEMM design based on the 4M scheme tailored for Ascend NPUs is proposed, with an AIC\/AIV dual-stream pipeline scheduling strategy equipped to coordinate padding operations, matrix\u2013matrix multiplications, and element-wise instructions across multi-level memory hierarchies and compute units; (iii) a fine-grained task scheduling and assignment mechanism is implemented to maximize Cube core occupancy across diverse matrix dimensions, improving hardware utilization for various computation workloads. Our experimental measurements show that the proposed CGEMM achieves a competitive hardware utilization rate of 83.6% across all tested matrix configurations, enabling efficient exploitation of available computing resources. Meanwhile, we observe a measured average speedup of 1.14\u00d7 relative to the AscendSipBoost implementation tested on an identical Ascend NPU, alongside a measured 3.17\u00d7 speedup compared with cuBLAS running on the Nvidia GPU platform adopted in our experiments across all evaluated matrix sizes. These results reflect the promising capability of Ascend NPUs for high-precision complex-valued computing workloads within the tested experimental setup.<\/jats:p>","DOI":"10.3390\/computers15070407","type":"journal-article","created":{"date-parts":[[2026,6,26]],"date-time":"2026-06-26T08:43:57Z","timestamp":1782463437000},"page":"407","update-policy":"https:\/\/doi.org\/10.3390\/mdpi_crossmark_policy","source":"Crossref","is-referenced-by-count":0,"title":["Design and Optimization of GEMM for Complex Numbers on Ascend NPU"],"prefix":"10.3390","volume":"15","author":[{"ORCID":"https:\/\/orcid.org\/0000-0001-9872-3095","authenticated-orcid":false,"given":"Erkun","family":"Zhang","sequence":"first","affiliation":[{"name":"School of Future Technology, South China University of Technology, Guangzhou 510641, China"},{"name":"Pengcheng Laboratory, Shenzhen 518071, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Yu","family":"Zhang","sequence":"additional","affiliation":[{"name":"School of Computer Science & Engineering, South China University of Technology, Guangzhou 510006, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-2273-1504","authenticated-orcid":false,"given":"Pengxiang","family":"Xu","sequence":"additional","affiliation":[{"name":"Pengcheng Laboratory, Shenzhen 518071, China"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-6372-7088","authenticated-orcid":false,"given":"Lu","family":"Lu","sequence":"additional","affiliation":[{"name":"Pengcheng Laboratory, Shenzhen 518071, China"},{"name":"School of Computer Science & Engineering, South China University of Technology, Guangzhou 510006, China"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1968","published-online":{"date-parts":[[2026,6,26]]},"reference":[{"key":"ref_1","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1145\/3503469","article-title":"Energy Efficient Boosting of GEMM Accelerators for DNN via Reuse","volume":"27","author":"Cicek","year":"2022","journal-title":"ACM Trans. Des. Autom. Electron. Syst."},{"key":"ref_2","first-page":"1","article-title":"A LAPACK Implementation of the Dynamic Mode Decomposition","volume":"50","year":"2024","journal-title":"ACM Trans. Math. Softw."},{"key":"ref_3","unstructured":"Nair, H., Vellaisamy, P., Chen, A., Finn, J., Li, A., Trivedi, M., and Shen, J.P. (2023, January 21\u201325). tuGEMM: Area-Power-Efficient Temporal Unary GEMM Architecture for Low-Precision Edge AI. Proceedings of the 2023 IEEE International Symposium on Circuits and Systems (ISCAS), Monterey, CA, USA."},{"key":"ref_4","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1145\/3378176","article-title":"Enabling Highly Efficient Batched Matrix Multiplications on SW26010 Many-core Processor","volume":"17","author":"Jiang","year":"2020","journal-title":"ACM Trans. Archit. Code Optim."},{"key":"ref_5","unstructured":"Miji\u0107, N., and Davidovi\u0107, D. (2022, January 23\u201327). Batched matrix operations on distributed GPUs with application in theoretical physics. Proceedings of the 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO), Opatija, Croatia."},{"key":"ref_6","first-page":"30318","article-title":"GPT3.int8(): 8-bit Matrix Multiplication for Transformers at Scale","volume":"Volume 35","author":"Koyejo","year":"2022","journal-title":"Proceedings of the Advances in Neural Information Processing Systems"},{"key":"ref_7","unstructured":"Xia, H.J., Zheng, Z., Wu, X.X., Chen, S.Y., Yao, Z.W., Youn, S., Bakhtiari, A., Wyatt, M., Zhuang, D.L., and Zhou, Z.Z. (2024, January 10\u201312). Quant-LLM: Accelerating the Serving of Large Language Models via FP6-Centric Algorithm-System Co-Design on Modern GPUs. Proceedings of the 2024 USENIX Annual Technical Conference (USENIX ATC 24), Santa Clara, CA, USA."},{"key":"ref_8","unstructured":"AMD (2026, June 23). rocBLAS 5.2.0 Documentation. Available online: https:\/\/rocm.docs.amd.com\/projects\/rocBLAS\/en\/latest."},{"key":"ref_9","first-page":"37524","article-title":"Understanding Int4 Quantization for Language Models: Latency Speedup, Composability, and Failure Cases","volume":"Volume 202","author":"Krause","year":"2023","journal-title":"Proceedings of the 40th International Conference on Machine Learning"},{"key":"ref_10","doi-asserted-by":"crossref","unstructured":"Yoneyama, R., Kawamura, M., Terashima, R., Yamamoto, R., and Toda, T. (2025). Comparative Analysis of Fast and High-Fidelity Neural Vocoders for Low-Latency Streaming Synthesis in Resource-Constrained Environments. arXiv.","DOI":"10.21437\/Interspeech.2025-1819"},{"key":"ref_11","first-page":"1","article-title":"Implementing High-performance Complex Matrix Multiplication via the 3m and 4m Methods","volume":"44","author":"Smith","year":"2017","journal-title":"ACM Trans. Math. Softw."},{"key":"ref_12","doi-asserted-by":"crossref","unstructured":"Hassan, M.W., Dabah, A., Ltaief, H., and Fahmy, S.A. (2023, January 15\u201319). Signal Detection for Large MIMO Systems Using Sphere Decoding on FPGAs. Proceedings of the 2023 IEEE International Parallel and Distributed Processing Symposium (IPDPS), St. Petersburg, FL, USA.","DOI":"10.1109\/IPDPS54959.2023.00020"},{"key":"ref_13","doi-asserted-by":"crossref","first-page":"800","DOI":"10.1109\/TCOMM.2024.3440830","article-title":"One-Bit Transceiver Optimization for mmWave Integrated Sensing and Communication Systems","volume":"73","author":"Lin","year":"2025","journal-title":"IEEE Trans. Commun."},{"key":"ref_14","first-page":"217","article-title":"Chapter Seven - Architecture of neural processing unit for deep neural networks","volume":"Volume 122","author":"Kim","year":"2021","journal-title":"Hardware Accelerator Systems for Artificial Intelligence and Machine Learning; Advances in Computers"},{"key":"ref_15","doi-asserted-by":"crossref","unstructured":"Liao, H., Tu, J., Xia, J., Liu, H., Zhou, X., Yuan, H., Zhang, L., Wang, Y., Li, W., and Chen, K. (March, January 27). Ascend: A Scalable and Unified Architecture for Ubiquitous Deep Neural Network Computing: Industry Track Paper. Proceedings of the 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), Seoul, Republic of Korea.","DOI":"10.1109\/HPCA51647.2021.00071"},{"key":"ref_16","doi-asserted-by":"crossref","unstructured":"Zhang, E., Xu, P., and Lu, L. (2026). Efficient Low-Precision GEMM on Ascend NPU: HGEMM\u2019s Synergy of Pipeline Scheduling, Tiling, and Memory Optimization. Computers, 15.","DOI":"10.3390\/computers15010039"},{"key":"ref_17","unstructured":"Huawei (2026, June 01). Atlas 300T Pro Training Card User Guide (Model 9000) 06. Available online: https:\/\/support.huawei.com\/enterprise\/en\/doc\/EDOC1100285917\/426cffd9\/about-this-document?idPath=23710424|251366513|22892968|252309113|256118195."},{"key":"ref_18","doi-asserted-by":"crossref","first-page":"389","DOI":"10.1016\/S0167-8191(06)80021-9","article-title":"The communication challenge for MPP: Intel Paragon and Meiko CS-2","volume":"20","author":"Hockney","year":"1994","journal-title":"Parallel Comput."},{"key":"ref_19","unstructured":"Kunkel, J.M., Balaji, P., and Dongarra, J. (2016). Performance, Design, and Autotuning of Batched GEMM for GPUs. International Conference on High Performance Computing, Springer International Publishing."},{"key":"ref_20","doi-asserted-by":"crossref","first-page":"495","DOI":"10.1016\/j.procs.2017.05.138","article-title":"The Design and Performance of Batched BLAS on Modern High-Performance Computing Systems","volume":"108","author":"Dongarra","year":"2017","journal-title":"Procedia Comput. Sci."},{"key":"ref_21","doi-asserted-by":"crossref","first-page":"1","DOI":"10.1145\/3431921","article-title":"A Set of Batched Basic Linear Algebra Subprograms and LAPACK Routines","volume":"47","author":"Abdelfattah","year":"2021","journal-title":"ACM Trans. Math. Softw."},{"key":"ref_22","unstructured":"Nvidia (2023, October 24). NVIDIA A800 80GB Active Graphics Card. Available online: https:\/\/www.nvidia.com\/en-us\/products\/workstations\/a800."},{"key":"ref_23","unstructured":"Huawei (2026, June 23). Domain-Specific Libraries. Available online: https:\/\/www.hiascend.com\/eng\/cann\/aal?tab=1."}],"container-title":["Computers"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/www.mdpi.com\/2073-431X\/15\/7\/407\/pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,6,26]],"date-time":"2026-06-26T09:22:19Z","timestamp":1782465739000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.mdpi.com\/2073-431X\/15\/7\/407"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,6,26]]},"references-count":23,"journal-issue":{"issue":"7","published-online":{"date-parts":[[2026,7]]}},"alternative-id":["computers15070407"],"URL":"https:\/\/doi.org\/10.3390\/computers15070407","relation":{},"ISSN":["2073-431X"],"issn-type":[{"value":"2073-431X","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,6,26]]}}}