资源描述
AI|S0800519070001 13072123839 2023 10 10 2 1 3 12-1.93-8.76 32.89 300-1.48-4.15-0.97-6%5%16%27%38%49%60%2022-10 2023-02 2023-06 300 GPU AI AI A100 H100 AI AI 1 1 1750 1024 A100 GPT-4 A100 2 ChatGPT Copilot GPU A100 GPU Meta GPU 1 3 AI 1.AI 2.+3.AI 5000P 4.2500P 5.AI 6.2 ZXCXuNrMsRmNrPtQnQrOmN7NaO6MsQmMnPtQkPmNmMlOrQoP7NrQtQxNqMyRuOpMpPCONTENTS CONTENTS 02030401 AI 05 4AIGC AI2023.4 ChatGPTGPT4InstructGPT/GPT3.5/ChatGPTGPT3Transfor-merPyTorch Azure OpenAI API Attention Transformer Decoder 1750 RLHF 2023 2022 2021 2020 2019 2018 2017 2015 2011 Transformer 1 OpenAI 3 BERT OpenAI 1.17 GPT-1OpenAI 15 GPT-2 OpenAI 1750 GPT-3 5300 MT-NLG 1.6 Switch Transformer OpenAI InstructGPT ChatGPT Stability AI Stable Diffusion OpenAI GPT-4 PaLM2 Emergent Abilities of Large Language Models 5 FLOPs 12 1.2 5GB48 15 40GB96 2.7B/6.7B/13B/175B 45TB96 1.3B/6B/175B GPT-1(2018.6)GPT-2(2019.2)GPT-3(2020.5)ChatGPT/GPT3.5(2022)GPT-4(2023)AI CPU GPU62003 2010 2017 2019/AI/PC(Windows)/IoT/.OS/UNIX/SolarisLinux/OpenStackK8S/Spark TensorFlow/Caffe/Torch PowerPC/SparkIntel/AMDIntel/AMD NVIDIA-GPU ADSL/2GPCIE 1.0100M 3GPCIE3.0(8GT/S)10G4GPCIE4.0(16GT/S)25G 5GPCIE5.0(32GT/S)100G/NLP:ChatGPTAIGC CPU CPU+vCPU/CDN-AIAI/ML/AI GPU AI7/SFT RM GPT-3.5 Small(1.25)2.6 PFLOPS*GPT-3 XL 27.5 PFLOPS*GPT-3 175B 3640 PFLOPS*PalM 5400 29600 PFLOPS*1350.4 PFLOPS*ChatGPT 2023 1 6.16 4874.4 PFlops GPT3 175B)3640 PFLOPS 35000 A100 1024 A100 13000 A100 433 A100/INT4/8 13B 50B,A10 V100 GPU 1024/8*17 2200 920/4000/Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM GPU8 2022 GPT 1746 GPT-3 DGX A100-80GB 1T 384 8 DGX-A100 GPU NVLink NVSwitch 8 200Gbps InfiniBand GPU=GPU/GPU GPU A100 32,3072 GPU 44%,52%GPU()T Token P n A100 X GPU X GPU GPU FP16 A100 312 TFLOPS GPT-3 T=3000 P=1750 A100 n=1024 1536 GPU 45%A100 312 45%=140 TFLOPS GPT-3=.=.（s）GPT-4 SemiAnalysis GPT-4 1.8 GPT-4 A100 GPT-4 SemiAnalysis OpenAI GPT-4 FLOPS 2.15 1025 25000 A100 90-100 GPU 32%-36%SemiAnalysis Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM GPU9 Similarweb GPU 10 2023 8 OpenAI Similarweb 2023 8 ChatGPT 14 7 1 ChatGPT 10 Token 1000 2 ChatGPT 24 3 GPU 45%1 1 GPT-3 1750 Int8 A100 624 TOPS A100 Token=月度访问 量 平均提问数 问题加回答 数=1.4 109 10 1 000=.=参数量 每月推理侧 数量 单个 计 算量=1.75 1011 1.4 101 3 2=.=30 24 60 60.(TOPS)A100A100=A100 GP U=1.9 101 86.24 1014 45%()Similarweb GPU ChatGPT11 2023 9 25 ChatGPT GPT-4 GPT-3.5 ChatGPT 2023 8 860 ChatGPT 70 ChatGPT A100 70 1 SemiAnalysis GPT-4 2800 2 Token 30000 3 ChatGPT 20 300 emiAnalysis Similarweb A10012 ChatGPT 2023 9 21 GPT-4 Copilot Copilot 2023 9 26 Windows 11 Clipchamp Microsoft 365 Copilot 2023 11 1 Teams Outlook Word Excel Loop OneNote OneDrive AI Microsoft 365 Chat GPU Copilot13 Word+Copilot Outlook+Copilot Teams+Copilot 2023 5 Windows 10 Windows Copilot 15%-80%Copilot Token 500-5000 A100=每 秒推理 侧算力 需求A 100 GP U=日活 Cop i lot 使用 率 每日 Toke n 人均 输入输 出 数量 推理参数 量 224 60 60 6.24 1014 45%30%Token 2000 Windows Copilot A100 1.80%Token 5000 A100 9 GPU Copilot+Windows14A100 Copilot15%30%50%80%Token500 1731 3462 5771 9233 1000 3462 6925 11541 18466 2000 6925 13849 23082 36932 5000 17312 34623 57705 92329 Windows Copilot A100 FY20Q1 Office 365 Microsoft 365 2 15%Microsoft 365 3 Microsoft 365 Copilot 15%-80%Copilot Token 2000-30000 30%Token 10000 Microsoft 365 Copilot A100 2 80%Token 30000 A100 15 GPU Copilot+Microsoft 36515 Microsoft 365 Copilot A100 A100 Copilot15%30%50%80%Token2000 2077 4155 6925 11079 5000 5193 10387 17312 27699 10000 10387 20774 34623 55397 30000 31161 62322 103870 166192 16 AI A100&H100 A100 A100 H100 OpenAI GPU OpenAI Andrej Karpathy 1000 H100 A100 A100 GPT-4 10000-25000 A100 Meta 21000 A100 7000 A100 Stability AI5000 A100 Falcon-40B 384 A100 H100 a)OpenAI 50000 Inflection 22000 Meta 25000 30000 Azure AWS Oracle Lambda CoreWeave 10 CoreWeave H100 35000-40000 Anthropic Helsing Mistral Character 10000 b)Inflection GPT-3.5 3500 H100 GCP 25000 H100 Azure Oracle 10000-40000 H100 50 A800 GPT3 1,000 50,000 GPT4 3,000 150,000 10 2C 1 A800 2C 100,000 1,000,000 A800 500 A800 400 200,000 CONTENTS CONTENTS 02030401 AI 05 18 GPU 1 CSDN 19 GPU Meta OPT-175B 40 24 3 2.8 2 1.5 2021 9 3 5 1750 1024 80G A100 FAIR NLPgroups(Microsoft/NVIDIA/OpenAI)Meta 992 A100 33(0)3000 Token BF16 20 AI GPU LLM-AIGC-1024/512 A100(SCCGN7ex)4 A10(GN7i)30-65B 8*A10(GN7i)8*V100(GN6e,32GB)8*A100(GN7e,80GB)30B 4*A10(GN7i)8*V100(GN6V,16GB)3-10B A10(GN7i)3B A10(GN7i)3D 4 A100(GN7e)256/128 A100(SCCGN7ex)GPU AI1 SCC 2 3 21 AISCC SCC GPU NVSwitch RDMA VPC CPFS&GPU cGPU GPU GPU GPU 22 AI IaaS AI AIACC AI AIACC AIACC-Torch Pytorch AIACC-MLIR MLIR AIACC-HRT AIACC AIACC-Torch Pytorch PyTorch AIACC-MLIR MLIR MLIR Tensorflow AIACC-HRT AIACC 50%LLM 40%AIGC Finetune 40%50%LLM 80%AIGC AI AIACC AIACC-ACSpeed AIACC-AGSpeed AIACC-HRT AIACC-MLIRAI AIACC-TrainingAI AIACC-Inference SCC CONTENTS CONTENTS 02030401 AI 05 24AI1 CPU GPU GPU 32GPU GPU-4 IDC IDC GPU GPU IDC 25 AI AI LLM GPU NVIDIA NVIDIA LLM Kubernetes ML/AI Ops AWS Azure GCP Oracle Coreweave Lambda 6 4.5 4.5 3 2 1 26AI A800 150.0 5-7 21.4 kWh 6 75%kwh 3.9/kWh 0.5 2.0/2 5800 10.415060 60 6040 4030290050100150200250300350 A800 120-150/5-7 12/P/8 A800 60/FP16 A800 624 TFLOPS 8 A800 5P H800 1979 TFLOPS 8 H800 16PCONTENTS CONTENTS 02030401 AI 05 28 AI 2022 8 AI AI 2023 7 1000P A&H AI 9 1300P 2023 3000P GPU 2023 6 5120P 1280P 256;1280P/360 640P Orange-GPT 640P wind 29+AI NVIDIA A800 H800 2023H1 AI 2651 CAE+HPC GPU AI 30 AI 5000P AI 2023 5000P P+70%10 3000P 2023Q4 10000P 2 800-1000P A800 H800 9-10 2023 9 H800 960P 12/P/IB 3.456 2023 11 1 3 wind 31 2500P 2023 7 GPU GPU AI(GPU)GPU AIDC 8 3,20 8 29 80 9 14 2500P 9 26()1000P 32 AI 7000 2023 6 20 1 33 2 A 10.8 10.9 2 P 2 2533Pops Int8 43 Pflops(FP32)AI 2 1.92 0.03 0.03 0.02 80 H800 AI 1280P AI/H800 AI 31664Tops Int8)2533120Tops Int8)H800 AI 536TFlops FP32)42880TFlops FP32)324 TB 42772 TB/AI 8 H800 230 Tesla T4 10 324TB 75 CONTENTS CONTENTS 02030401 AI 05 35 GPU A800 H800 GPU 36 6-12 10%6-12-10%10%6-12 10%6-12 20%6-12 5%20%6-12-5%5%6-12 5%6-12 A 300 500 276 12 59 303 6008 10C 021-3858420937/QQ BBS 91610000719782242D
展开阅读全文