AWS re:Invent 2025 - AWS Graviton: The best price performance for your AWS workloads (CMP360)

AWS Graviton: The Best Price-Performance for Your AWS Workloads

Overview of AWS Silicon Portfolio

• AWS has over 10 years of innovation in silicon, including 6 generations of Nitro cards, 5 generations of Graviton chips, and Trainium and Inferentia for machine learning
• AWS has shipped millions of chips across these platforms and seen remarkable success

Why Graviton?

• Graviton delivers the best price-performance in EC2 instances
• Based on the Graviton chip with AWS innovation across the entire stack
• Over 300 Graviton-based instances available in 38 regions worldwide
• Used by over 90,000 customers across a wide range of workloads

Evolution of Graviton Generations

• Graviton 1 (2018): 16 cores, great for scale-out microservices
• Graviton 2 (2019): 64 cores, enabled mainstream workloads with up to 40% better price-performance
• Graviton 3 (2021): Even faster and more efficient, enabled compute-bound, memory-bound, container, Java, and machine learning workloads
• Graviton 4 (2023): Increased performance for scale-up workloads like large databases and EDA applications

Graviton 5: Doubling Down on Performance

• Doubled core count from 96 to 192 cores in the same chip size
• Built on a 3nm process with 25% higher performance per core
• Increased L3 cache by 5.3x to 192MB, over 600MB total cache
• Supports up to DDR5-8800 memory and PCIe Gen 6
• 33% reduction in latency and massive increase in bandwidth per core

Security Innovations with Graviton

• Graviton 2 was first to support DDR encryption, Graviton 4 first with PCI encryption
• Graviton 5 introduces the Nitro Isolation Engine:
  • Compartmentalizes critical interactions using formal verification
  • Proven properties like memory safety, confidentiality, and integrity

Performance Improvements with Graviton 5

• 35% improvement in PyTorch-based machine learning workloads
• 32% increase in requests per second for a Grails Java web application
• 47% performance boost for a large monolithic internal workload
• 27% improvement in Nginx web server performance
• 40% increase in Hammerdb database benchmark

Real-World Customer Results

• Snowflake saw over 30% performance boost on their virtual data warehouse
• Honeycomb observed 20-25% latency improvement, 36% better throughput per core
• Airbnb found the M9 instances 25% higher performance than other architectures
• SAP HANA saw 35-60% performance increase generation-over-generation

Simplifying the Migration to Graviton

• 90% of applications work on Graviton without code changes
• Mature OS ecosystem with official ARM64 builds
• Container and serverless support through ECS, EKS, Lambda, Fargate
• Managed services like RDS, Aurora, EMR support Graviton out-of-the-box
• AWS Transform tool automates compatibility checks and recompilation
• Graviton Savings Dashboard helps calculate cost savings

Recommendations for Specific Workloads

• Spark: Optimize shuffle partition size, use latest versions
• C/C++: Compile natively for ARM64, leverage SIMD vectorization
• Python: Use ARM64-native libraries, optimize math libraries
• Java: Rebuild native libraries, leverage latest JDK versions
• Machine Learning: Use ARM64-native frameworks, leverage AWS DLC

Getting Started with Graviton

• Explore the AWS Graviton Technical Guide for in-depth resources
• Try the free T4g Graviton instance to experience the benefits
• Leverage AWS tools like Transform, Aperf, and Savings Dashboard
• Sign up for the Graviton 5 M9g preview to test the latest generation