|2023 10 17 昇 AI 1 昇 昇 AI CANN AI 2 昇 310 910 AI 昇 910 Int8 0.6 P 300W 3 昇 910 310 AI 昇 Atlas Atlas 1 AI ChatGPT AI 2021 2 ADS 2.0 2023 NCA L2 10 TOPS L3 100 TOPS L4 1000 TOPS 昇 MDC 3 AI 昇 AI AI 2023 7 6 昇 AI 25 4 昇 1 昇 910 2019 AI AI 2 3 昇 昇 ISV S0680518090001 1 GPT-4V 2023-10-15 2 2023-10-10 3 100 AI 2023-10-09-16%0%16%32%48%64%2022-10 2023-02 2023-06 2023-10 300 2023 10 17 P.2 昇 AI.4 1.1 AI.4 1.2+.6.12 2.1 AI.12 2.2.16 2.3 AI.19 2.4.21.23 3.1.23 3.2 Huawei Inside+.29 3.3 ISV.30.33 1.4 2 昇.5 3 昇 310 昇 910.6 4 昇 310 昇 910.6 5 GPGPU/AISC.7 6 昇 AI.8 7 昇 AI.8 8 Atlas 200 AI 3000.8 9 Atlas 300T 9000.8 10 昇.9 11 Atlas 500.9 12 Atlas 500 3000.9 13 Atlas 500.9 14 Atlas 800 3000.10 15 Atlas 800 9000.10 16.10 17 Atlas 900 AI 9000.11 18 Atlas 900 AI.11 19 AI.11 20.12 21 NLP.13 22 CV.13 23.14 24.14 25.14 26.15 27 Nature.15 28.15 0YFUvMrMmPnMrPmPsPtMnM7N9R6MmOrRsQoNfQpOqQfQoPuM9PmNtQvPnQqNwMrMtM 2023 10 17 P.3 29.15 30 GPT-3.16 31 LLM GCP TPU v4.16 32 HUAWEI ADS 2.0.17 33 HUAWEI ADS 2.0.17 34 HUAWEI ADS 2.0.17 35 2022-2023.18 36.18 37 Octopus.18 38 ResNet50 GPU.19 39 昇 AI.19 40.20 41.20 42.20 43.20 44 2022 10 7 4.21 45.22 46：昇 腾计算 产业生 态.23 47.25 48 AI.26 49.27 50.28 51.28 52 HI.29 53 HI.29 54&-.30 55 IT.30 56 HiHope.31 57 HarmonyOS/OpenHarmony.31 58.31 59 昇 AI.31 60.32 61.32 62.32 63 Atlas 200.32 2023 10 17 P.4 昇 AI 2019 昇 昇 AI 1+昇+x86+GPU AI Atlas 900 AI Atlas 300 AI Atlas 800 2 AI 1 1.1 AI 昇 昇 HUAWEI Ascend AI CANN Compute Architecture for Neural Networks AI 2023 10 17 P.5 2 昇 昇 昇 AI 昇 1 昇 昇 AI 昇 2 昇 CANN MindStudio AI MindSpore 昇 昇 昇 laas Paas SaaS 昇 AI 2023 10 17 P.6 1.2+昇 HUAWEI Ascend 310 FP16 16 TFLOPS INT8 8 TOPS 8W AI AI AI 昇 HUAWEI Ascend 910 AI 3D Cube AI 昇 910 FP16 320 TFLOPS INT8 640 TOPS 310W 3 昇 310 昇 910 4 昇 310 昇 910 昇 HUAWEI Ascend 310 昇 HUAWEI Ascend 910 3D Cube 3D Cube 16 TOPS INT8 8 TOPS FP16 640 TOPS INT8 320 TFLOPS FP16 8W 310W 12nm FFC N7+昇 910 A100 GPGPU/ASIC 昇 910 INT8 A100 PCIe 80GB 2023 10 17 P.7 5 GPGPU/AISC FP64(TFLOPS)FP32(TFLOPS)INT8(TOPS)A100 PCIe 7nm 9.7 19.5 624 600GB/s 80GB PCIe 300W A800 PCIe 7nm 9.7 19.5 624 400GB/s 80GB PCIe 300W Tesla V100 PCIe 12nm 7.8 15.7-300GB/s 80GB PCIe 300W AMD Instinct MI250X 6nm 47.9 47.9 383 100GB/s 128GB PCIe 560W AMD Instinct MI250 6nm 45.3 45.3 362.1 100GB/s 128GB PCIe 560W AMD Instinct MI100 7nm 11.5 23.1 184.6 92GB/s 32GB PCIe 300W 100 7nm-37 295 64GB/s 32GB PCIe 250W 100P 7nm-240 1920 448GB/s 64GB PCIe 550W 104P 7nm-256 1024 192GB/s 32GB PCIe 300W Google TPUv4i 7nm-138 300GB/s 8GB-Google TPUv4 7nm-275 1200GB/s 32GB-昇 310 12nm-16-8W 昇 910 7nm-640-310W T20-32 256 300GB/s 32GB PCIe 300w MLU370-X4 7nm-24 256 307.2GB/s 24GB PCIe 150W MLU3370-S4 7nm-18 192 307.2GB/s 24GB PCIe 75W TechPowerUp Unixcloud AMD NextPlatform 昇 910 310 AI 昇 腾 计算 产业 在硬 件方面 坚持“硬 件开 放”策略，通过自有硬 件和 合作 硬件 相结 合的方 式为 客户 提供 多样 化选择。Atlas 昇 AI Atlas 200 Al Atlas 200 DK AI Atlas 300 Al Atlas 500 Atlas 800 Al Atlas 900 Al Atlas Atlas Atlas 200 Al AI Atlas 300 AI AI OEM/ODM 2023 10 17 P.8 6 昇 AI 7 昇 AI 昇 AI 1 昇 Atlas 200 Al 3000 昇 22TOPS INT8 AI Atlas 300I 3000/3010 昇 88 TOPS INT8 80 2 Atlas 300T 9000 昇 320 TFLOPS FP16 AI 8 Atlas 200 AI 3000 9 Atlas 300T 9000 2023 10 17 P.9 10 昇 Atlas 200 AI 3000 Atlas 300 3000 Atlas 300T 9000 AI 昇 310 昇 310 昇 910 AI 22TOPSINT8 16TOPSINT8 8TOPSINT8 88TOPSINT8 280TFLOPSFP16(Pro)220TFLOPSFP16 PCIe PCIe 4 Gen3.0 PCIe x16 Gen3.0 PCIe x16 Gen4.0 4GB:5.5W/8GB:8W 67W 300W 2 昇 昇 Atlas 500 11 Atlas 500 Atlas 500 智能 小站（型号：3000）AI 芯片 昇腾 310 AI 算力 22/16 TOPS INT8 内存规格 LPDDR4X，8GB/4GB，最大 51.2GB/s 编解码能力 支持 H.264 硬件解码，16 路 1080P 30 FPS 等 接口 网络：2 个 GE RJ45 其他 I/O：1 个 HDMI 接口，1 对 3.5mm 立体声输入输出接口；2 个外部和 1 个内部 USB2.0 接口（Type-A）典型功耗 无盘配置：25W；有盘配置：40W 环境条件 无盘配置：-40 70；有盘 配置：-40 60 结构尺寸 无盘配置：45mm*235mm*220mm；有盘配置：45mm*355mm*220mm 12 Atlas 500 3000 13 Atlas 500 3：昇 Atlas 800 Atlas 500 Pro 2023 10 17 P.10 Atlas 800 3000/3010 昇/Intel AI 8 AI 704 TOPS INT8 OCR Atlas 800 9000/9010 昇 Intel 8 昇 2.24P FLOPS FP16 Atlas 500 Pro 3000 昇 AI 4 AI 352 TOPS INT8 Al 475mm 600mm 14 Atlas 800 3000 15 Atlas 800 9000 16 Atlas 800 3000 Atlas 800 9000 Atlas 500 Pro 3000 2U AI 4U AI 2U AI CPU 2*920 4*920 1*920 CPU 32 DDR4 3200MT/s 32 DDR4 RDIMM 3200MT/s 16GB/32GB/64GB 4 DDR4 3200MT/s AI/8*昇 910/AI 8 Atlas 300I/4 Atlas 300I AI HBM/32 GB 1228GB/s/AI 704 TOPS INT8 2.24 PFLOPS FP16 1.76 PFLOPS FP16 352 TOPS INT8 4 Atlas 900 Al 昇 HCCS PCIe 4.0 100G RoCE 昇 256P 1024P FLOPS FP16 50 PC AI 2023 10 17 P.11 17 Atlas 900 AI 9000 18 Atlas 900 AI 2023 7 6 昇 昇 AI 4000 16000 AI 30 19 AI 昇 2023 2023 10 17 P.12 2.1 AI ChatGPT 2021 4 AI AI L0 L1 L2 20 L0 NLP CV NLP AI 2023 10 17 P.13 21 NLP C+Java pytho API/+LLM NLP CV 22 CV 10 AIGC 80%5 3D 2023 10 17 P.14 23 24 3D 昇 Transformer 25 2023 10 17 P.15 26 AI AI 27 Nature 28 29 AI 1-7 operational IFS 0.25 X0.25 1 13 CADD AIDD AI L1 L0 L1 L2 Oneflow 2023 10 17 P.16 GPT-3 140 LLM 1120 1 token OneFlow GPT-3 139.8 token 6N 2N N LLM FLOPS 46.2%TPU v4 PaLM 5400 30 GPT-3 OneFlow 2 LLM 2800 Gopher 5400 PaLM 200 1200 31 LLM GCP TPU v4 GPT-3(OpenAI)Gopher(Google DeepMind)MT-NLG(Microsoft/Nvidia)PaLM(Google Research)Model Parameters 175B 280B 530B 540B FLOPs/Token/Model Parameter 6 TPUs/Machine 4 Peak FLOPS/TPU 275T FLOPS Utilization 46.20%Cost/Machine/Hour(1-year reserved)$8.12 Seconds/Hour 3600 Training Cost/1000 Tokens$0.0047$0.0075$0.0141$0.0144 Train Tokens 300B 300B 270B 780B Training Cost$1,398,072$2,236,915$3,810,744$11,216,529 OneFlow 2.2 2023 9 12 AITO M7 1 M7 HUAWEI ADS 2.0 HUAWEI ADS 2.0 BEV Bird Eye View GOD 2.0 99%2023 10 17 P.17 2 M7 1 3 11 12 2023 12 32 HUAWEI ADS 2.0 HUAWEI ADS 2.0 2023 4 5 2023 4 2023 HUAWEI ADS 2.0 M7 HUAWEI ADS 2.0 AITO M5 11 S HI 33 HUAWEI ADS 2.0 34 HUAWEI ADS 2.0 2023 4 HUAWEI ADS 2.0 AI 1000 km 2023 9 NCA MPI Miles Per Intervention 200km 99%L2 L3+36 L2 2022 L2 L3 35%9%2023 51%20%1 B 2 2023 10 17 P.18 35 2022-2023 36 L2 10 TOPS L3 100 TOPS L4 1000 TOPS MDC MDC 昇 SoC AOS VOS AutoSAR MDC 36 37 Octopus Octopus MDC OS+AI MDC 昇 910 AI MindSpore AI ResNet50 GPU 2 TensorFlow PyTorch 24%35%9%2%11%51%20%11%00.10.20.30.40.50.6L1 L2 L3 L42022 2023E 2023 10 17 P.19 38 ResNet50 GPU 2.3 AI AI 昇 AI AI 39 昇 AI 昇 2023 2023 7 6 昇 AI 25 15 6 7 昇 AI 2021 5 31 2022 2 7 200P 2023 10 17 P.20.LuoJiaNet.LuoJiaSet 300 2023 9 30 400P 2021 5 9 9 AI 300P HPC 8P 2022 170 150 40 41 昇 昇+昇 10 14 2400 2021 10 21 100P AI 300P AI 2022 60 30 10 90%昇 2022 5 40P 4 95%2022 12 200 150 42 43 昇 昇 昇 AI 300P 85%140P 400P 300P 400P 2023 10 17 P.21 昇 AI AI 2.4 2022 10 7 FDP 4 28 2023 3 31 23 2023 6 30 ASML ASML 9 1 DUV 44 2022 10 7 4 2023 2023 2 23 2023 2 27 2023 5 28 2023 10 17 P.22 50 45 AI 2023-2024 2023 8 22 I G I CPU Intel G CPU I 2198 G 1977 47%2023 PC 2023 7 13 2023 PC ARM ARM PC PC 2023 9 9 2023 10 17 P.23 昇 昇 昇 昇 OEM ODM IHV C&SL ISV XaaS 昇 昇 46：昇 昇 3.1 Powered by Ascend 昇 昇 1 2023 7 PKS AI 2 2023 10 17 P.24 1992 2019 2022 3 1958 60 2020 6 18+昇 AT800 Model 3000 昇 310+920 8 Atlas 300I/V Pro AI AT800 Model 9000 昇 910+920 AI AT500 Pro Model 3000 2023 10 17 P.25 47 AT800 Model3000 AT800 Model3010 AT800 Model9000 AT800 Model9010 AT500 Pro Model 3000 2U AI 2U AI 4U AI 4U AI 2U AI AI 2 920 1 2 IntelXeonSP Skylake/CascadeLake 8*昇 910 8*昇 910 1*920 AI 1120 TOPS INT8 980 TOPS INT8 2.56/2.24/2/1.76 PFLOPS FP16 2.56/2.24/2/1.76 PFLOPS FP16 420 TOPS INT8/205W 5.6 kW 5.6 kW Atlas 300I Pro 72W 昇 4 2020 AI KunTai A222 2U AI/KunTai A722 2U AI 920 昇 Atlas AI ARM Al KunTai A924 4U AI 920 昇 Atlas AI ARM AI 2023 10 17 P.26 48 AI KunTai A222 2U AI KunTai A722 2U AI KunTai A924 4U AI 1*920 24 2.6GHz 2*920 32 48 64 2.6GHz 4*920 48 2.6GHz/64 3.0GHz AI 3 Atlas 300V Atlas 300I Pro Atlas 300V Pro 8 Atlas 300V Atlas 300I Pro Atlas 300V Pro 8*昇 910 200G RoCE AI 420 TOPS INT8 1120 TOPS INT8 512Tops Int8 256Tops FP16 5 2017 4 昇/昇/昇+昇 昇 昇 AI AI+AI AI 昇+2022 9 昇 11 2022 昇 RA2300 AI 920+昇 AI Atlas 300I Pro Atlas 300V Pro AI RA5900 AI 920+昇 910 CA9900 AI 920 昇 910 AI AI AI AI AI 2023 10 17 P.27 SA300 920+昇 AI 2U AI SA200 49 RA2300 AI RA5900 AI CA9900 AI SA300 SA200 2U 4U 47U 2U/AI 昇 310 8*昇 910 64*昇 910 昇 310 昇 310 AI Atlas 300I Pro Atlas 300V Pro/3 Atlas 300I Pro/300V Pro/AI 1.12 POPS INT8 2.56/2.24.2 PFLOPS FP16 20.48 PFLOPS FP16 420 TOPS INT8 384 1080P 30FPS 22/16 TOPS INT8 11/8 TFLOPS FP16/5.6kW 43kW/25W 40W 昇 6 昇 2020 5 G220K V2 昇 310 2U2 Al 8 Atlas 300I Al G440K V2 昇 910+920 AI AI G210K V2 2023 10 17 P.28 50 G210K V2 G220K V2 G440K V2 2U AI 2U AI 4U4 AI 352 TOPS INT8/2.24/2 PFLOPS FP16 AI 1*920,24 2.6GHz 2*920,24/32/48/64 2.6GHz 2*920 AI 4 Atlas 300I 8 Atlas 300I/4 DDR4 2933 MT/s 32 DDR4 2933 MT/s 32 DDR4 2933 MT/s 32/64GB RDIMM 昇 7 2008 昇 S500K2 920 24C 2U 51 S500K2 2U AI AI 14 Atlas 300I 3000 AI 352 TOPS INT8 12*3.5 SAS/SATA 4 DDR4 3200MT/s 64GB 86.1mm*447mm*475mm 昇 2023 10 17 P.29 3.2 Huawei Inside+Huawei Inside HI HI 1 5 AR-HUD HI HI OS AOS HOS VOS OS 52 HI 53 HI 2C 2020 11 BU CEO M7 1 2023 2021 2023 2 2019 12 MDC 2022 11 X8 PLUS HOL-Pilot 3 2021 HiCar 4 2023 8 BE22 2023 10 17 P.30 BE22 5 LUXEED S7 3.3 ISV 1 CTSP Cloud Transformation Service Provider HarmonyOS Connect/HarmonyOS HarmonyOS Connect PCB HarmonyOS SDK UX H5/FA 13 30%20%54&-55 IT 2 HiHope HarmonyOS/OpenHarmony HarmonyOS/OpenHarmony HarmonyOS Connect WiFi/BLE UX UX APP/FA HarmonyOS Connect H5/FA 2023 10 17 P.31 56 HiHope 57 HarmonyOS/OpenHarmony 3 IT SI IT 昇+昇 169 2022 4 21 昇 昇 2023 58 59 昇 AI 4 200 IT 2023 10 17 P.32 2023 6 6 AI AI 2023 9 26 L1 60 61 2022 5 2019 AI Atlas200 AI 2019 Atlas 2020 7 2021 7 昇 2021 WAIC 昇 2022 6 昇 2023 昇 62 63 Atlas 200 2022 2023 10 17 P.33 昇 昇 910 2019 AI AI 昇 昇 ISV 2023 10 17 P.34,6 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