实时人工智能推理模式：来自生产游戏工作负载的经验教训.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.I N D 3 3 1 6Real-time AI inference patterns:Lessons from production game workloadsKyle SomersHe/HimSr.Manager,Solutions ArchitectureAmazon Web ServicesJ.Aaron FarrH

2、e/HimCo-Founder&CTOJam&Tea Studios 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.AgendaAI use cases in gamesChallenges of real-time inference:validity,latency,costBackground of Jam&Tea Studios Jam&Teas approach to solving these challenges 2025,Amazon Web Services,Inc.or its affi

3、liates.All rights reserved.AI use cases in gamesGame DevelopmentCreating content to accelerate game production,facilitate communication,prototyping.Publishing OperationsAccelerate publishing functions like marketing asset creation,customer service,and community managementIn-Game Features “AI Native

4、Games”Improve player immersion,creativity and learning with AI-powered game features.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.To develop AI Native Games,we needValiditySpeedLow costMust return results the game engine can interpret while still being“creative”Must return resu

5、lts fast enough for real-time gameplayCost of inference needs to be within budgetLarge models excel at complex reasoning but are slower and cost more vsSmall models are fast and cheap,but can be inconsistently valid 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Player IntentAI R

6、easoningAI gameplay examplePlayer types/says:“Make a sandwich”Determine action typeUnderstand environmentNearby objectsCurrent game stateUnderstand game rulesDetermine difficultyDetermine outputsApply guardrailsRespond within timeAI Gameplay1.Remove the meatball from the world2.R