重新定义网络借助思科唤醒认知人工智能思科呈现.pdf

1、Redefine Networking,Awaken Cognitive AI with CiscoOCP 2025 Executive TalkOctober 2025Rakesh ChopraSVP and Fellow,Hardware Architecture,Cisco 2025 Cisco and/or its affiliates.All rights reserved.Unlocking AI potential|Scaling AI cluster GPT-2|20191.5B parameters1e21 FLOPs10s-100s GPUs36%MMLU performa

2、nceGPT-3|2020175B parameters3e23 FLOPs1,000s GPUs43%MMLU performanceGPT-4|20231T parameters1e25 FLOPs10,000s-25,000 GPUs86%MMLU performanceGPT-5|2025 2-10T parameters1e26 FLOPs50,000s-100,000+GPUs92%MMLU performanceStatistics generated by ChatGPT4.1Baseline graph from Epoch AI,September 2025,added G

3、PT5 point manuallyFLOPs|Floating point operationsMMLU|Massive Multitask Language Understanding(measures general knowledge&reasoning)2019-2025ParametersFLOPsGPUsKnowledge&ReasoningMMLUTraining Compute(FLOP)GPT-2GPT-3GPT-42018 2019 2020 2021 2022 2023 2024 2025GPT-5Other major models like Llama,Gemini

4、,Grok,others have excellent performance.Only using GPT to simplify trends.2025 Cisco and/or its affiliates.All rights reserved.Data Center evolutionScale-Up100s GPUs504xBandwidthScale-Out50K GPUs56xBandwidthFront-EndCompute/Storage7xBandwidthWAN/DCIEnd-User1xBandwidthGPT-5|2025 2-10T parameters1e26

5、FLOPs50,000s 100,000+GPUs92%MMLU performanceTraining models are at the limit today!50K GPUs50,000s-100,000+GPUs 2025 Cisco and/or its affiliates.All rights reserved.Scaling intelligence|Power is everythingScaling based on 51.2T switch|8 Rail|800GE GPU2-TierNetwork3-TierNetworkCluster size unlocks in

6、telligence|Limited by power16K128K512K33M22MW185MW690MW44GW800G between Leaf/Spine100G between Leaf/Spine33M800G GPUs44GWPowerPower Consumption:92%GPU,4%Switch,4%OpticsPower Consumption:88%GPU,6%Switch,6%OpticsPowerGPUs(Log)Todays switches enable massive AI cluster scaleScale=Intelligence 2025 Cisco