规模化还是停滞不前：开放生态系统中专用连接的重要性（Astera Labs 报告）.pdf

1、1Scale or StallThe Case for Purpose-Built Connectivity in Open EcosystemsAhmad DaneshAssociate VP,Product Management October 14,20252Audience Poll What are the biggest challenges with scaling AI infrastructure?Proliferation of AI accelerators&platforms from multiple vendorsRapid rollout of platforms

2、 to keep pace with emerging AI workloadsEnsuring AI infrastructure operates at maximum utilizationAll of the above?33Purpose-Built Connectivity for Rack-Scale AISoftware-Defined,Intelligent Connectivity PlatformHQ San Jose|Austin|Bengaluru|Haifa|Ho Chi Minh City|Irvine|Seattle|Shanghai|Singapore Tai

3、pei|Toronto|VancouverTrusted by leading hyperscalers&AI platform providersScaling to Support AI Deployments WorldwideEmbedded SoftwareManagement APIsModulesControllersI/O ChipletsSignal ConditioningFabricBoardsConnectivity ICs,Boards&ModulesOpticalCopperNVLinkFounded 2017Nasdaq|ALAB4AI clusters requ

4、ire high-throughput and efficient data transfer across racks,leveraging multiple protocols to optimize performance and reduce latencyAstera Labs develops products that form the connectivity nervous system,enabling seamless communicationAI Infrastructure Demands Robust Connectivity at Scale5AI Infras

5、tructure Transformation:From Server-Scale to Rack-Scale ComputingThe Server is the unit of computeThe Rack is the unit of compute1.02.08 GPUs Scale-Up to look like one Big GPU=Scale-up performance limited by the server chassisComponents optimized for server form factor GPU,switch,power,cooling,footp

6、rint,etc.Scale-out multiple servers over a network72/288/576 GPUs Scale-Up to look like one MASSIVE GPU.Scale OutScale Up=Scale-up performance limited by the rackComponents optimized for rack form factorGPU,switch,power,cooling,footprint,etc.Scale-out multiple racks over a networkAI InfrastructureAI