20231105_国盛证券_通信行业证券研究报告：用AI重新定义通信超算互联网时代的调度与调优_47页.pdf
|2023 11 05 AI AI GPU GPU 1./-/GPU AI AI buy more save more 2.-AI SaaS AI 3.-AI AI DCI OctoML DECI Databrick AI AI GPU AI BOSS AI S0680519010002 S0680519050002 S0680522050002 S0680522120003 S0680523070001 S0680521050001 1 2023-10-29 2 AI 2023-10-26 3 2023-10-22-16%0%16%32%48%64%2022-11 2023-03 2023-07 2023-11 3002023 11 05 P.2 EPS P E 2022A 2023E 2024E 2025E 2022A 2023E 2024E 2025E 300308.SZ 1.52 2.57 4.86 5.57 58.32 34.49 18.24 15.92 300502.SZ 1.27 0.93 2.04 2.75 25.20 34.41 15.69 11.64 300394.SZ 1.02 1.53 2.55 3.26 69.90 46.60 27.96 21.87 300570.SZ 0.78 1.01 1.88 2.21 41.21 31.82 17.10 14.54 688195.SH 0.45 0.64 0.93 1.38 68.71 48.31 33.25 22.41 603496.SH 0.24 0.41 0.72 1.04 112.13 65.63 37.38 25.88 603220.SH 0.32 0.66 0.82 1.02 114.13 55.33 44.54 35.80 Wind WUFUvMnQmPmNqOnOtOtMnMbRbP8OmOnNmOtQiNqRrQjMqQpR6MqQvNwMsOrOxNmNqO2023 11 05 P.3.6 1.AI.7 2./.9 2.1 AI.9 2.1.1AGI GPU.9 2.1.2DGX Cloud GPU.11 2.1.3GPU VS.14 2.1.4 GPU.14 2.1.5.16 2.1.6 AI.19 2.2.22 2.2.1.22 2.2.2.23 2.2.3 AI.27 2.2.4.29 2.2.5.33 2.3.35 2.3.1What-.35 2.3.2Where-.35 2.3.3How-.36 2.3.4Who-.37 3.38 3.1.38 3.1.1.38 3.1.2.38 3.1.3.40 3.2.40 4.43 5.46 1.7 2 Buy more Save More GH200.8 3 snowflake.8 4.9 5 CPU GPU.9 6 CPU GPU.9 7 h100.10 8 DGX A100 DGX H100.10 9 DGX Cloud.11 10 NVIDIA DGX.12 11 NVIDIA DGX IaaS.12 12 NVIDIA DGX IaaS.13 2023 11 05 P.4 13 10 23 SEC 8-K.14 14 13.15 15 atlas 900.15 16 AI.16 17.16 18 AI POD.17 19 AI POD.18 20 DGX GH200 256 GPU NVIDIA NVLink.19 21 AIGC.20 22.20 23.21 24.22 25 2019-2026.23 26 2019-2026.23 27.24 28 PCIe6.0.24 29 PCIe Switch.25 30 OpenCAPI.26 31 CXL.26 32 DGX GH200.26 33 RDMA.27 34 RoCE InfiniBand.27 35.28 36.28 37 OpenAI Kubernetes 7500.29 38 DPFR.30 39 Wireshark.30 40 DPI DFI.31 41 DPI.31 42 DPI.31 43.32 44.32 45 tensor flow.33 46 tensor flow Data FlowGraph.33 47.34 48.34 49 deepspeed.34 50&.35 51.36 52.36 53.37 54.37 55 TILT.38 56 ModularMojo notebook.38 57 OctoAI GPU.39 58 DeciLM Llama2 token.39 2023 11 05 P.5 59 Databricks Lakehouse AI AI.40 60 Databricks Lakehouse Monitoring.40 61 Databricks.41 62 snowflake.41 63 Deci.42 64.43 65.44 66.45 67.45 2023 11 05 P.6 AI AI BOSS 2023 11 05 P.7 1.AI AI IDC 1 AI buy more save more 2023 11 05 P.8 2 Buy more Save More GH200 AI 3 snowflake 2023 11 05 P.9 2./2.1 AI 2.1.1AGI GPU ICT IAAS PAAS SAAS 4 Kim et.al,EPSim-C:A Parallel Epoch-Based Cycle-Accurate Microarchitecture Simulator Using Cloud Computing AI CPU GPU open ai chatgpt AI CPU GPU GPU 5 CPU GPU 6 CPU GPU 2023 11 05 P.10 GPU openai gpt3 1750 117 GPU gpu gpu gpu cpu 7 h100 GPU cpu 400G H100 400G 800G 8 DGX A100 DGX H100 fibermall AI AGI 2023 11 05 P.11 1 2 3 4 5 2.1.2DGX Cloud GPU 2023 3 21 DGX Cloud AI AI DGX Cloud AI DGX Cloud AI NVIDIA DGX SuperPOD NVIDIA 9 DGX Cloud DGX Cloud Base Command Base Command AI SaaS AI AI Base Command DGX Cloud AI AI DGX Cloud AI AI DGX Cloud AI DGX Cloud DGX Cloud laaS AI DGX Cloud NVIDIA Base Command Platform GPU DGX Cloud NVIDIA AI Enterprise 2023 11 05 P.12 10 NVIDIA DGX DGX Cloud IaaS DGX Cloud NVIDIA DGX DGX Cloud DGX Cloud IssS 2-3 GPU 11 NVIDIA DGX IaaS GPT-3 400 3000 token GPT-3 160 NVIDIA A100 gpu DGX Cloud IaaS 2-3 DGX Cloud H100 20 IaaS 3 DGX Cloud 2023 11 05 P.13 12 NVIDIA DGX IaaS DGX Cloud AI GPU AI AI aws azure AI AI AI 2023 11 05 P.14 2.1.3GPU VS 255 GPU DGX 10+AI 9 2 AI AGI 13 IDC A100 A800 H800 910B GPU 2.1.4 GPU 2023 10 23 SEC 8-K 4800 A100 A800 H100 H800 L40S 10 17 30 13 10 23 SEC 8-K BIS 13 BIS 13 2023 10 17 2023 11 05 P.15 14 13 BIS AMD intel AI 910B A100 15 atlas 900 2023 11 05 P.16 AI AWS Azure AI AWS Azure nvda A100 gpu 16 AI azure AI AI AI AI rtx 4090 2.1.5 17 2023 11 05 P.17 GPU Memory Wall W_m ADAM 16G AI 20+AI All Reduce AI AI Ring-AllReduce 4K 1 68 AI openAI AIPod 3 CLOS Leaf LF Leaf Spine SP Spine SuperSpine SSP GPU Allreduce 5GB/s 70%90%GPU AllReduce 20GB/s GPU 400G 18 AI POD 2023 11 05 P.18 AIPod 8 8 8 GPU 8 Leaf 8 Leaf Group 512 GPU 8 Leaf 8 Spine Leaf fullmesh 16K GPU 19 AI POD SuperSpine Spine AIPod AIPod 1:1 12.8T 25.6T 51.2T AI GPU AI AI DGX A100 DGX H100 DGX GH200 GPU AI GPU 2019 Mellanox GH200 NVlink 900GB/s 450GB/s 8 3.6Tbps 28.8Tbps NVlink Switch NVlink Switch 9.6Tbps 800G 9.6T/800g*2=24 72 800G GPU 800G 1:9 DGX H100 GPU 800G 1:3 DGX A100 400G 2023 11 05 P.19 20 DGX GH200 256 GPU NVIDIA NVLink GPU GPU GPU 2.1.6 AI 1 UGC UGC AIGC AI AIGC UGC AIGC 2023 11 05 P.20 21 AIGC 2 BitInfoCharts 2023 10 31 476.73EH/s 2009 5MH/s 14 89 AI ChatGPT AI 3 4 AI OpenAI 2012 2018 AI 3-4 30 7 10 AI 2017 AIGC 2021 ELMo 9400 2020 GPT-3 1750 2021 MT-NLG 5300 22 Substack 2023 11 05 P.21 3 2009 1 CPU CPU GPU FPGA ASIC 23 36 AI 2012 Alex Krizhevsky ImageNet GPU GPU GPU 2017 2022 3 H100 2017 3 GTX 1080ti FP32 12 134/11.34 AIGC FP64 192 68/0.35 A I D C G PU G PU AI GPU AI A 100 A I G C 2023 11 05 P.22 2.2 2.2.1 AI GPU AI 24 5G AI 5G AIGC IDC 2026 111.3EFLOPS 1271.4EFLOPS 5G AI 2023 11 05 P.23 AI GPU 25 2019-2026 26 2019-2026 IDC IDC AI 5G AI B C 2.2.2 1 GPU AI GPU GPU AI GPU 2600 80G A800 10-20 2600 100-200 GPU GPU GPU 2023 11 05 P.24 27 GPU PCIe RDMA PCIe NVLINK/CAPI/GenZ/CCIX/CXL PCIe 2001 PCIe PCI-Express PCIe CPU GPU SSD 2003 PCIe1.0 6.0 64GT/s 16 256GB/s 28 PCIe6.0 viavisolutions PCIe PCIe Switch PCIe PCIe PCIe Switch PCIe Switch PCIe PCIe PCIe Switch PCIe 2023 11 05 P.25 29 PCIe Switch GPU PCIe NVLINK/CAPI/GenZ/CCIX/CXL AIGC GPU GPU GB/S PCIe GPU GPU PCIe Switch PCIe Switch PCIe PCIe GPU CAPI IBM Open CAPI I/O IBM CAPI GenZ GenZ AMD ARM IBM Nvidia Xilinx GenZ GenZSwitch CXL 2019 Intel CAPI 2021 GenZ 2022 Open CAPI CXL CCIX ARM GenZ 2023 11 05 P.26 30 OpenCAPI 31 CXL NVLINK GPU PCIe NVLINK 1 2 GPU GPU 3 CPU GPU GPU NVSwitch ASIC NVLink GPU GPU 2023 5 29 AI DGX GH200 NVLink 和 NVSwitch 256 GH200 GPU 100TB 32 DGX GH200 IB RDMA IB TCP/IP IO RDMA CPU RDMA Infiniband RoCE iWARP 2023 11 05 P.27 Infiniband RDMA IB IB RoCE RDMA RoCE iWARP TCP RDMA TCP RoCE iWARP TCP iWARP 33 RDMA 34 RoCE InfiniBand 2.2.3 AI AI 1-1750 2 GPU 100 8 30 2-Allreduce GPU Allreduce 5GB/s 70%GPU Allreduce 20GB/s 90%GPU 400G 3-GPU+GPU GPU 99.9%60%8 99%GPU 99.99%8 50%checkpoint checkpoint 2023 11 05 P.28 35 1*N 16000 10 MTBF 1 1000 MTBF 4+36 2023 11 05 P.29 openAI checkpoint OpenAI checkpoint checkpoint checkpoint checkpoint 37 OpenAI Kubernetes 7500 InfoQ 2.2.4 AI IB InfiniBand packet DPFR Data Plane Fast Recovery DPFR DPFR 2023 11 05 P.30 38 DPFR Wireshark TCP/UDP Fiddler HTTP/HTTPS tcpdump&windump solarwinds nast Kismet Wireshark MAC Wireshark 39 Wireshark techlockdown 2023 11 05 P.31 CPU Wireshark DPI DFI DPI Deep Packet Inspection DPI IP TCP UDP DPI DFI Deep/Dynamic Flow Inspection/DPI DFI 40 DPI DFI DPI/DFI/DPI SaaS/PaaS DPI DPI DPI/41 DPI 42 DPI 2023 11 05 P.32 SaaS/PaaS DPI 43 100G/IB 200G/IB 100G IB 44 2023 11 05 P.33 2.2.5 GPU GPU GPU AI Tensor Flow PyTorch AI GPU GPU 45 tensor flow 46 tensor flow Data FlowGraph Tistory freecodecamp N N N N 2023 11 05 P.34 47 48 colossalai Chen et.al,EFFICIENT AND ROBUST PARALLEL DNN TRAINING THROUGH MODEL PARALLELISM ON MULTI-GPU PLATFORM Deepspeed DeepSpeed GPU GPU GPU 49 deepspeed aheader Deepspeed Deepspeed 2023 11 05 P.35 GPT AGI GPU 2.3 2.3.1What-50&2.3.2Where-2023 11 05 P.36 51 2.3.3How-CPU GPU 52 2023 11 05 P.37 2.3.4Who-AI/53 54 2023 11 05 P.38 3.3.1 3.1.1 LLM NVIDIA NeMo MetaLlama LLM snowflake AI Snowflake Applica TILT Text-Image-Layout-Transformer 55 TILT Arxiv snowflake 3.1.2 Modular Mojo Modular Mojo Python MILR AI Python 56 ModularMojo notebook Modular 2023 11 05 P.39 octoML OctoML StableDiffusion OctoML OctoML GPU Llama2 3 GPU GPU KV OctoAI 57 OctoAI GPU octoML Deci DeciLM Transformer Grouped-Query Attention GQA Transformer attention DeciLM Transformer 60 DeciLM PyTorch 70 Llama2 58 DeciLM Llama2 token Deci 2023 11 05 P.40 3.1.3 Databricks GPU Databricks snowflake Databricks snowflake Lakehouse AI Databricks Lakehouse Monitoring LLM 59 Databricks Lakehouse AI AI 60 Databricks Lakehouse Monitoring Databricks Databricks 3.2 AI AI AI AI/databrick databrick SaaS Databrick SaaS SaaS 2023 11 05 P.41 61 Databricks Databricks snowflake cloudflare Snowflake AI AI AI AI 62 snowflake snowflake 2023 11 05 P.42 DECI octoML DECI octoML AI AI 63 Deci Deci AMD GPU GPU GPU GPU GPU 2023 11 05 P.43 4.AI AI 64 PE 2023E 300608.SZ 36 ICT 688316.SH 26-14.0 ICT 002657.SZ 60+AI 9 603110.SH 64 688229.SH 20 ONE 603220.SH 123 55.3 603019.SH 545 28.0 600941.SH 20034 14.6 601728.SH 4886 15.6 600050.SH 1415 16.9 Wind PE Wind 2023 11 3 2023 11 05 P.44 65 PE 2023E 300308.SZ 712 34.5 2022 300502.SZ 227 34.4 300394.SZ 282 46.7 300570.SZ 74 31.7&688195.SH 40 48.2 Lumentum 688205.SH 55 31.7 301205.SZ 88 142.0 800G 000988.SZ 288 25.8 400/800G 688498.SH 131 163.2 A 603083.SH 96 800G 002902.SZ 52 800G Wind PE Wind 2023 11 3 2023 11 05 P.45 66 PE 2023E 000063.SZ 1258 11.1 ICT 000938.SZ 567 23.4 301165.SZ 212 31.9 IT 688702.SH 200-519.7 A 301191.SZ 63 29.0 EoC PON 601138.SH 2930 12.1 688256.SH 454-72.8 AI ASIC 688418.SH 41 90.4 Wind PE Wind 2023 11 3 AI AI 67 PE 2023E 603496.SH 86 65.8 688292.SH 42 61.0 002912.SZ 48 28.9 1675.HK 81 300047.SZ 56 75.6 BOSS 300166.SZ 117 Wind PE Wind 2023 11 3 2023 11 05 P.46 5.1.AI AI AI AI 2.AI 3.2023 11 05 P.47,6 A 300 500 15%5%15%-5%+5%5%10%-10%+10%10%8 7 7 100077 868 One56 1 10 200120 021-38124100 1115 330038 0791-86281485 100 24 518033
