20231204_华龙证券_AI算力行业深度研究报告：智算供给格局分化国产化进程有望加速_24页.pdf
y 2023 12 04 AI 2023 10 4.0 GPT-4 2023 1 7 64 AI AI 30 MindSpore AI 17 Chiplet AI 1 2 IP 3 4 华龙证券研究所 S 0 230 523 0800 04 sun b w h l zq gs.co m-20%0%20%40%60%80%100%2022-12-022023-02-172023-04-282023-07-072023-09-152023-12-01沪深300 AI 算力指数 智 算 供给 格 局分 化，国 产 化进 程 有望 加 速 AI 算力 行业 深度 研究 报告 1.1 1.1.1 1.1.1.1 1.1.2.3 1.2 AI.5 1.2.1 AI.6 1.2.2 AI.7 1.3 AI.10 1.3.1.10 1.3.2 AI.11 1.3.3 AI.12 2 AI.12 2.1 AI.12 2.2.13 2.2.1 AI.13 3 AI.14 3.1.14 3.2:Chiplet.15 3.2.1 Chiplet.15 3.2.2.16 4.16 4.1-U 688256.SH.16 4.2 603220.SH.17 5.19 XVBYzQsPpMrQpRsRqNoRrQ7N8Q6MoMnNnPmPeRnMtRfQpPtO8OpPwPuOnMqMuOpOsQ 1 2018.1 2 2017-2023.2 3.3 4.4 5.4 6.4 7.5 8 MindSpore-.8 9 Megatron.9 10 2023 10 18.10 11 370.16 12 2020-2025.17 13 2020-2025.17 14 2020-2025 ROE().17 15 2020-2025.19 16 2020-2025.19 17 2020-2025 ROE().19 18 2020-2025 PE.19 1.6 2 2023.7 3 2023.7 4 2023.9 5 AI.11 6 H800.14 7.18 1 1 1.1 AI 1.1.1 2018 NLP 2023 7 2 48 2023 1-7 31 2021 2023 1-7 1 2018 2017-2023 7 2017 Transformer 251438483101020304050602018 2019 2020 2021 2022 2023 1-7 2 2018 6 OpenAI Transformer GPT-1 1.2 10 BERT 3 2019 OpenAI 15 GPT-2 2019 9 83 Megatron-LM 110 T5 170 Turing-NLG 2020 OpenAI GPT-3 1750 10 5300 Megatron-Turing 2021 1 Switch Transformer 1.6 2022 OpenAI GPT-3.5 ChatGPT GPT-3.5 2023 OpenAI GPT-4 1.8 2 2017-2023 Attention Is All You Need GPU 2021 8 Deepak Narayanan GPU 175B A100 1024 Token 300B,34 1T A100 3072 Token 450B 84 0.12 0.34 1.5 8.3 11 17 175 530 600 1,600 1,800 05001,0001,5002,000GPT-1BERTGPT-2Megatron-LMT5T-LNGGPT-3Megatron-TuringGshardSwitch-TransformerGPT-4 3 2020 6 OpenAI GPT-3 2021 1 Switch Transformer 1.1.2 2019 2023 7 130 2023 1-7 64 3 2023 3 1.0 4 5 7 3.0 2023 10 4.0 GPT-4 123132640102030405060702019 2020 2021 2022 2023 1-7 4 2023 7 40%60%14 13 10 7 6 6 4 5 6 L0 L1 L2 5+N+X 5 N X 14 21 1 171613644611442 211031405101520253035404550 60%40%5+7 1.2 AI 2023 8 FLOPs=6 Training Tokens FLOPs=8 Training Tokens FLOPs=2 Prompt Tokens GPT-3 3.15E+23 FLOPs GPT-4 2.15E+25 FLOPs GPT-4 MoE 2770 6 1(TOKENS)FLOPs)Baichuan2 2023 10 7B 13B 2600B 1.092E+23 2.028E+23 Llama 2 2023 7 Meta 7B 13B 34B 70B 2000B 8.4E+22-8.4E+23 2023 6 104B 1600B 9.98E+23 GPT-4 2023 3 OpenAI 1800B 1300B 2.15E+25 GLM-130B 2022 10 AI 130B 400B 3.12E+23 GPT-3 2020 6 OpenAI 175B 300B 3.15E+23 Llama 2:Open Foundation and Fine-Tuned Chat Models 1.2.1 AI 2023 10 4.0 GPT-4 2023 GPT-4 7 GPT-3 2023 2023 200 134 80 72 500 2023 3.03E+25 FLOPs 7 2 2023 FLOPs FLOPs 9E+22 72 6.48E+24 GPT-3 3.15E+23 7 2.2E+24 GPT-4 2.16E+25 1 2.16E+25 FLOPs)3.03E+25 80%20%GPT-4 2 Tokens 80 4.10E+17 FLOPs GPT-3 1 Tokens 80 1.30E+17 FLOPs 500 5000 Tokens 80 1.85E+16 FLOPs 2023 2.65E+18 FLOPs 3 2023 FLOPs FLOPs GPT-4 4.10E+17 1 4.10E+17 GPT-3 1.30E+17 7 9.07E+17 1.85E+16 72 1.33E+18 FLOPs 2.65E+18 2023 3.03E+25 FLOPs 1.2.2 AI 8/2023 9 AI 2022 8 AI MindSpore AI AI 2023 40%2023 54 8 MindSpore-MindSpore 2023-500 5,000 Tokens 8.1E+24 FLOPs 3000 Tokens 80 2023 6E+17 FLOPs 9 2023 8.1E+24 FLOPs 4 2023 FLOPs FLOPs 1.5E+23 1.11E+16 54 54 8.1E+24 6E+17 FLOPs 8.1E+24 2023 3.84E+25 FLOPs AI 2023 5 GPT-3 GPU GPU 44%52%GPU GPU GPU 44.8%GPU GPT-3 A100 A100 FP16 140TFLOPS 9 Megatron NVIDIA GPU 60-90 GPU 50%2023 31,648-47,472 A100 32641282565121024153619022520307244%44%46%43%44%45%47%50%52%52%0%10%20%30%40%50%60%05001000150020002500300035001.7B 3.6B 7.5B 18.4B 39.1B 76.1B 145.6B 310.1B 529.6B 1008B B:GPU 10 2025 18 OpenAI 2012 3.4 2020 6 GPT-3 2023 3 GPT-4 7 2025 5 133 1.3 AI 1.3.1 2023 10 18(TTP)(PD)Chiplet 10 2023 10 18 BIS 11 1.3.2 AI 10 8 2023 5.5 E+19 FLOPS 2025 300 EFLOPS 35%1.05 E+20 FLOPS AI A100/A800 IP 910 A100/A800 2023 5 AI 30 5 AI 910-A100/A800 310 FP16 8TOPS 12nm-BR-100 FP16 1000TFLOPS 7nm 590-A100/A800 370 INT8 256TOPS 7nm-2 FP16 128TFLOPS 7nm-3-4nm()-800 INT8 825TOPS 12nm-12 1.3.3 AI 2.5D 3D IP 2023 10 7/6nm 40%60%70%5/4nm 75-80%CoWoS 2.5D 3D 120%10 2024 50 7nm 3-6%2 AI 2.1 AI 1-2 13 2.2 2.2.1 AI AI IDC P A100/A800/H800 H800 8 H800 228 8 GPU 16P 2023 H 29.9/5%5 12/P/H800=P/P()=/()+()+()+()=14()=()()H800 100%H800 5 H800 17 2023 11 A100 100%6 H800/H800/228 228 P 16/p/12 192/22.80 2.98 1.20 1.1 38.38 16.70 1./10%2.6KW 9 545 3.1000 4.0.5%3 AI 3.1 15 2022 2 2023 6 3.2:Chiplet 3.2.1 Chiplet 18 24 AMD IEDM 250 45nm 1 14/16nm 2 7nm 4 Chiplet Chiplet Chiplet 2021 Chiplet 370 16 Chiplet 11 370 3.2.2 Chiplet 2022 3 3 AMD Arm Google Meta Chiplet Chiplet Universal Chiplet Interconnect Express UCIeUCIe PCIe PCIe CXL Raw Mode IP UCIe Chiplet Chiplet 2023 1 13 CCITA IP 4 4.1-U 688256.SH 17 AI 290 270 370 Cambricon-1M Cambricon-1H IP 220 IP AI 590 A100 910/920 A100 12 2020-2025 Wind 13 2020-2025 14 2020-2025 ROE()Wind Wind 4.2 603220.SH 2023 8 458.93721.05 729.03913.631,736.472,676.970204060801000500100015002000250030002020A 2021A 2022A 2023E 2024E 2025E()(%)-434.51-824.95-1,256.56-623.77-459.87-112.14-100-50050100-1400-1200-1000-800-600-400-2000()(%)-6.76-14-25.88-15.27-13.43-6.26-30-25-20-15-10-502020A 2021A 2022A 2023E 2024E 2025EROE()(%)18 8.5 AI 2023 11 15 67,160 2023 11 12 18/P/1920P 128 15P 34,560 H800 8 GPU 16P H800 H800 228/29,184 11 7 H800 P 1920/P/)18/P/)34560 29184/2918/146/122.18/154/3340.10 5 11.22 Wind 19 15 2020-2025 16 2020-2025 Wind Wind 17 2020-2025 ROE()18 2020-2025 PE Wind Wind 1 2 IP 3 4 5()IP 2,066.532,641.462,643.133,326.173,966.834,653.500510152025300.001,000.002,000.003,000.004,000.005,000.002020A 2021A 2022A 2023E 2024E 2025E()(%)56.93181.65108.79193.17242.33295.83-100-500501001502002500501001502002503003502020A 2021A 2022A 2023E 2024E 2025E()(%)3.5410.626.110.0311.4612.4024681012142020A 2021A 2022A 2023E 2024E 2025EROE()(%)85.8529.3631.6676.8561.2550.180204060801002020A 2021A 2022A 2023E 2024E 2025EPE 20 6-12 A 300 300 10%300 5%10%300-5%5%300-10%-5%300-10%300 300 300 189 F1 100033 638 21 730030 0931-4635761 720 11 200000
