20230824_东方证券_计算机行业算力网络系列深度报告全国一体化算力网络建设加速_62页.pdf

|2020 AI 1.AI+2.3.4.2025 33+(603019)(688041)(000977)(600602)-U(688256)(002261)(600839)(603019)-W(688158)(300846)A(000032)(300895)(301085)2.昇.WUAZvMqNoNmNtNtQsPoRoP9P9R7NmOnNmOnOjMqQyQiNpNqN9PoOuNMYmOsOwMoPqP 3.4.3.1 4.3.2 4.3.4.昇.4.13 AI,.21 Z100 NV A100 AMD MI100.昇.25 2023 8.26 2023 8.5.43.44 4.0.45.46 10.47.48.49.50.51.52.55 IO500.56 III AI.57 昇 AI.58 II.60.61.6.76 X785-G30 HPC/.77 X785-G40 GPU.78 5A.79.80 DCU.81 CPU DCU.82 2023 3 31.83.84.昇.7 1.1 AI 1.1.1 2020 12 2021 7 2021-2023 2022 2 2023 2 2023 4 17 2025 1 2023.04 2025 2023.02,5G 2023.02 2022.02 2022.01 3 2021.12 1.1.2 2022 2 2022 2 17 8 8 10 1 2022 2 8 10 1.1.3 AI ChatGPT J.Sevilla Compute Trends Across Three Eras of Machine Learning,2022 International Joint Conference on Neural Networks(IJCNN)2012 20 5-6 2015-2016 10 2 3(OOM)2022 ChatGPT AI Bert GPT4 GB 2 J.Sevilla,L.Heim,A.Ho,T.Besiroglu,M.Hobbhahn and P.Villalobos,Compute Trends Across Three Eras of Machine Learning,2022 International Joint Conference on Neural Networks(IJCNN),Padua,Italy,2022,pp.1-8,doi:10.1109/IJCNN55064.2022.9891914 9 AI 2022 AI AI IDC 2022-2023 2023 50%AI 3 IDC AI AIGC AI 20%80%AI AI AI AI 1.2 1.2.1 4.3 10 0%1%2%3%4%5%6%7%8%0%10%20%30%40%50%60%70%80%90%2021 2022 10 4 5 1.2.2 6 23 7 1.2.3 0100200300400500600/11 1993 7 NII,360 1.3 1.3.1 1961 John McCarthy 1990 Ian Foster 12 8 1.3.2 AI AR VR 9 13 1.3.3 10 2.1 2023 5 2023 5 2023-2025 2023 5 2023 2024 2023 6 2022 9 14 2023 4 2023 3（2023-2025）2023 1 2.2 1.2.3 4.11 15 3.1 CPU AI AI Int8 12 80%AI AI 13 AI,3.2 2022 45%28%16 1000GFlops IMD 14(2022)2022,2021-2026 52.3%IDC 2021 155.2EFLOPS 2026 1,271.4EFLOPS 2021-2026 52.3%15 EFLOPS IDC AI 2022 AI 385 AI 2027 AI 2164 16 AI 45%28%27%02004006008001000120014002019 2020 2021 2022E 2023E 2024E 2025E 2026E(EFLOPS)17 2022 60%2027 88%2018 12%2022 60%2023 70%2027 88%AI 17 2018-2027E 3.3 2022 6()60%60%AI 16(FP16)400PFLOPS 32(FP32)200PFLOPS 16(INT16)400POPS 2 60%60%1953273855668271210167521640%10%20%30%40%50%60%70%80%050010001500200025002020 2021 2022 2023E 2024E 2025E 2026E 2027E AI 12%27%41%51%60%70%76%81%85%88%88%73%59%49%40%30%24%19%15%12%0%20%40%60%80%100%2018 2019 2020 2021 2022 2023E 2024E 2025E 2026E 2027E 18 400PFLOPS(16(FP16);32(FP32)200PFLOPS 16(INT16)400POPS 3 5 2023-2025 2025 18,000 PFLOPS FP32 2023 2024 2025 5000PFLOPS FP16 60%2025 80 400 10EFLOPS 3 2023 2024 2025 10.8 25 50 80 225 260 330 400/3000 11 20 23 26 PUE/1.41 1.4 1.3 1.2/56.5%57%61%65%Eflops 0.81 2 5 10 Eb 11.34 24 45 60 20 18 16 14/30%50%100%19 3.4 30+40+30 40+30 AI AI 昇 20 5A 4/5A/昇 昇 100P FLOPS 500P 1000P 100P FLOPS 300P FLOPS FP16 300P FLOPS 1000P 100P FLOPS 400P FLOPS 200P FLOPS 25 1000P 1.14Eops 4Eops 40P FLOPS 100P 昇 800P OpS/100P FLOPS 300P FLOPS 100P FLOPS 100P FLOPS 100P FLOPS 300P-40P 1.4P 20 99P FLOPS 100P FLOPS 200P 200P 500P 100P FLOPS FP16 20 850 1400P FLOPS/800P OpS 300P FLOPS 10 1 500POPS 12EFLOPS 3EFLOPS 5A 1024P 100P FP16+5P FP64 300P FP16+15P FP64 105P 400P/100P 3.4.1 5A 5A 5A 18 5A-21 19*DCU GPGPU DCU CPU DCU DCU 2022 AI AI DCU DCU 20 DCU GPGPU DCU GPGPU CUDA CUDA NVIDIA C/C+NVIDIA GPU CUDA DCU GPGPU DCU AMD ROCm CUDA GPU TensorFlow Pytorch PaddlePaddle CUDA ROCm DCU A100 AMD MI100 DCU HCZB-2021-ZB0364 DCU Z100 8192 FP64 10.8TFlops 32GB HBM2 AI 22 A100 FP64 9.7 TFlops 40/80GB HBM2 AMD MI100 FP64 11.5 TFlops 32GB HBM2 21 Z100 NV A100 AMD MI100 AMD 3.4.2 昇 昇腾 20 昇 AI Atlas 昇 Atlas AI AI Atlas 900 昇 910 AI HCCS PCIe 4.0 100G RoCE 80%256P 1024P FLOPS FP16 22 Atlas 900 AI 2023 Altas 900 昇 98%2019 4,000 16,000 AI 16,000 1750 100B 30 昇 AI AI 昇 23 2020 1000P AI 25 5000P 昇 AI 23 2023WAIC,昇 昇 AI 昇 IHV 昇 13 昇 24 昇 昇 AI 100%24 昇 25 2023 8 26 2023 8 wind wind 昇 AI AT800 Model 9000 AI 昇 Atlas 900 AI 300 PFLOPS 15 27 AI 昇 昇 昇 AI CA9900 AI 25 AI 昇 昇腾 AI 昇 昇 28 AI AT800 Model 9000 29 3.4.3 2022 10 10 2.56 昇 计算峰值速度达到 100PFLOPSFP16，支持 MindSpore1.7 TensorFlow1.15 PyTorch 26 88000 44%22000 11%6 1,000 100%88,000 44%22,000 11%20,000 10%40,000 20%30,000 15%1,000 100%200,000 100%3.4.4 AI GPU,AI 2 GPU 100 100 100 100 3 AI 100 27 40P 70 200P 650 AI 100 AI ChatGLM LLaMA Aquila 32 AI AIGC E OAM PUE 1.1 AI AI AI 2.0 200P AI TF32 33 2.5 i20 34 31 28 WAIC 4.1 35.J.,2023,49(5):9-14.29 2023 4 1,000 PFLOPS FP16-7 2023 4 1,000 PFLOPS FP16 2023 2024 1 III 2023 2-1 3 20 4.2 30 4.2.1 2 2023 2 36 2023 2 4.2.2 2023 2 4.2.3 2.0 31 2022 5 2.0 2023 5 2.0 18 30PFlops 80Pb 64 4 37 2023 5 487 32 4.2.4 2023 2 38 32 Digital Shanghai 14 2023 4 14 39 4.2.5 2023 4 40 33 4.2.6+2023 7 1000P 9000P 41 4.3 IDC 34 4.0 1.0 42 4.3.1 4.0 2021 7 2022 10 4.0 Portal 35 43 44 4.0 10+3000 AI GeoEast 45 46 10 IT3.0 47 36 2019 5G CENI 48 2025 1 10 16 2030 100 2035 3000 49 37 4.3.2 China Computing NET,C NET 2019 AI 2.3E NPU MLU GPU 2022 6 2022 5 29 China Computing NET,C NET 2022 6 15-20 50 51 52 38 IP 30%IP Edge-Disjoint KSP 30%2022 12 UCS UCS CNCF Karmada 2025 2019 100P 500 450 II 4000 昇 AI 4000 A100 GPU 1000P II AI IO500 39 III 23 2023 2025 III 6G 3.2 16000P II 55 II IO500 56 III AI 昇 AI II 昇 AI 4096 AI 2048 CPU 4 II 昇 910 CANN AI 昇 MindSpore MindX ModelArts 1000 GPT-3 2600 AI III 10%57 昇 AI 昇 40 II 99%1/4 3/4 AI AI AI 58 II 4.3.3 1.0 AI/59 41 IT 4.3.4-60 2023 6 42-61 4.3.5 2015 16 80,000 1,000PFlops 800PB 62 43 2017 8 13 63 4.3.6 ICT 64 44 5.1 67 45 5.2 2023 2024 2025 35 300P FLOPS FP16 30 5.3 46 A800 2023 5000R 12000R 30%-50%GPU NVIDIA V100S A100 A800 68 69 UCloud UCloud+24 94 5 CDS NVIDIA A100 A40 A5000 2022 30 40 12000 15 80%AI AI AIGC Pre-train Fine-tune 70 94 5 47 IDC 20 2.2 11%71 72 PKS 2020 CECSTACK CeaCube CeaStor CeaSEC 73 48 IDC CBD IDC IDC 74 2022 49 2025 33+75 2023 7 28 1000 1.AI+50 2.3.4.2025 33+7.1 2006 2014 X785-G30 X785-G40 76 X785-G30 HPC/77 X785-G40 GPU 5 7 50-5A 5A 51 78 5A 79 7.2 CPU GPGPU CPU GPGPU CPU DCU,GPGPU CPU DCU AMD CPU DCU 80 DCU CPU DCU CPU DCU 2018 DCU 2020 1 2022 6 52 81 CPU DCU 2022 FP64 CPU DCU AI AI TensorFlow PyTorch Caffe2 AI 1000 7.3 AI IT IDC 2021 AI AI 7.4 1986 1987 2015 2023 3 31 5%28.03%6.5%82 2023 3 31 53 ToG ToG 110KV 2021 2021-2022 2022 9 83 84 7.5 AI AI 2016 AI 1A 2017 970 Mate10 AI MLU100 MLU270 MLU290 54 85 7.6 昇 1996 2008 31 10+1500 2017 86 昇+AI AI 昇 AI AI AI AI AI 昇 昇 昇 AI+55 8 33 87 7.7 2020 6+昇 昇 AI 300P AI 15 昇 100 200 88 89 AI AT800 Model 9000 AT800 Model 9000+昇 AI AT800 Model 9000 ChatGPT AI AT800 2.56 PFLOPS FP16 56 8*100G RoCE v2 1070%1.3 2.56 PFLOPS/5.6 kW 7.8 PKS 2021 PKS IaaS PaaS SaaS PKS P CPU K S CPU 60%PKS CPU 74%87%90 PKS 2021 802.6 2023 1203.9+CPU AI 7.9 AI IDC IDC 60 3,000 10 57 2022 GEC 91 7.10 IDC+IDC IaaS PaaS SaaS CBD IDC 92 58 7.11 AI 2012 IaaS PaaS 31 11 CPU 93 AI UCloud 2018 GPU 4.4-8.8kW GPU GPU A100 V100S MI100 AI 94 95 59 7.12:60 1%61 12 A 300 500 15%5%15%-5%+5%-5%5%-5%+5%-5%