AMD公司发表了题为“利用开放系统构建未来人工智能数据中心”的演讲.pdf

1、BUILDING THE FUTURE OF AI DATA CENTERS WITH OPEN SYSTEMSOctober 2025Krishna Doddapaneni,Corporate Vice President of Engineering,Networking Technology and Solutions Group,AMD2|PublicAI is Driving Massive Compute DemandParameters(Log Scale)2014201620182020202220232025AmoebaNetB557 millionResNet5026 mi

2、llionBERT-large330 millionChatGPT-41.8 trillionGPT-3175 billionPaLM540 billionGPT-21.5 billion2x per yearImage and speech recognition models14xper yearLanguage+recommender models2024Llama 4 Behemoth 2 trillionHigh Performance AI NetworkingSolving the Most Critical Challenges for Scaling AI and HPCHi

3、gh Performance AI NetworkingSolving the Most Critical Challenges for Scaling AI and HPCOpen ProgrammablePerformant4|Advancing AI 2025Coming in 2026AMD“Helios”Rack-Scale ArchitecturePowering AI of the FutureSCALE-UPSCALE-OUT&ACROSSAI INFRASTRUCTURE SOFTWARE5|Advancing AI 2025Network Challenges for AI

4、 WorkloadsAt an average 30%of training cycle timeis elapsed in waiting for networkingCommunication accounts for 40%-75%of timewith Training and Distributed Inference Models2%of time spent in networking40%75%AI Model 1AI Model 235%57%18%38%AI Model 1AI Model 2AI Model 3AI Model 435%57%18%38%40%75%Sou

5、rces:1)2022 OCP Keynote by Alexis Bjorlin,VP at Meta,2)Computation vs.Communication Scaling for Future Transformers on Future Hardware,https:/arxiv.org/pdf/2302.02825.%of time spent in networking40%75%AI Model 1AI Model 235%57%18%38%AI Model 1AI Model 2AI Model 3AI Model 435%57%18%38%40%75%6|Open In

6、teroperable Solutions for AMD Instinct AI NetworkingPollara 400 AI NIC Product OptionsRoCEv2,UEC-Ready RDMA&Custom TransportLeadership Performance;P4 ProgrammableTrusted by Industry-leading OEM/ODM PartnersAMD InstinctMI3XX GPUPollara400-1Q400P(PCIe Gen5)Pollara400-1Q400P-OCP(OCP 3.0)7|AMD Pensando