Inside Graviton4: How AWS's New Chip Cuts EC2 Costs by 20% for Compute-Heavy Workloads

AWS recently launched Graviton4, its fourth-generation custom silicon chip built specifically for EC2 workloads. Designed from the ground up for cloud-native compute, Graviton4 delivers a headline 20% reduction in EC2 costs for compute-heavy workloads compared to previous-generation Graviton instances, while outperforming x86-based equivalents on price-performance for the same use cases.

Graviton4 Architecture: Built for Efficiency

Graviton4 is fabricated on TSMC's 3nm process node, a major leap from the 5nm process used in Graviton3. It uses ARM's Neoverse V2 core architecture, which brings support for ARMv9.2 instruction sets including Scalable Vector Extensions 2 (SVE2) for accelerated machine learning and scientific computing workloads. Key architectural improvements include:

• 72 Neoverse V2 cores per socket, up from 64 Neoverse V1 cores in Graviton3
• 2MB of private L2 cache per core, doubling the L2 cache of Graviton3
• 96MB of shared L3 cache, 50% larger than Graviton3's L3 pool
• DDR5-6400 memory support with 12 memory channels, delivering 50% more memory bandwidth than previous generations
• PCIe 5.0 connectivity for faster storage and accelerator attachments
• 30% better integer performance and 25% better floating-point performance over Graviton3 per core

Why 20% Cost Reduction for Compute-Heavy Workloads?

The 20% cost cut for compute-heavy workloads stems from two core factors: improved performance per watt and higher compute density. Graviton4 delivers 40% better performance per watt than Graviton3, reducing the power and cooling overhead AWS passes to customers. For compute-bound workloads (where CPU utilization is consistently above 70%), the higher per-core performance means fewer instances are needed to complete the same workload, driving down total cost by 20% compared to equivalent Graviton3 deployments.

Compared to x86-based EC2 instances, Graviton4 delivers up to 40% better price-performance for compute-heavy use cases, but the 20% figure specifically refers to cost reductions for existing Graviton users migrating to the new chip for compute-intensive tasks.

Benchmark Results: Real-World Performance

AWS and third-party benchmarks confirm the cost and performance gains for compute-heavy workloads:

• Computational Fluid Dynamics (CFD): 22% faster simulation times than Graviton3, with 18% lower cost per simulation run
• Video Encoding (H.265 4K): 30% higher throughput per instance, 25% lower cost per encoded stream
• HPC Linpack: 35% better performance per watt than comparable x86 instances, with 20% lower total cost for large-scale HPC clusters
• Batch Data Processing (Apache Spark): 28% faster job completion times, 21% lower cost per terabyte processed

Workloads that are memory-bound or I/O-bound see smaller savings, as the bottleneck lies outside the CPU. AWS recommends Graviton4 for compute-bound use cases to maximize the 20% cost reduction.

Supported Workloads and Use Cases

Graviton4 is optimized for any compute-heavy workload that can run on ARM64 architecture. Top use cases include:

• High-Performance Computing (HPC) for scientific research, weather modeling, and financial simulations
• Batch processing for ad tech, log analysis, and large-scale data transformation
• Media rendering and video encoding for streaming platforms
• AI inference for small to medium-sized machine learning models, accelerated by SVE2 instructions
• High-throughput web services and API backends with consistent compute demand

Most modern Linux distributions (Amazon Linux 2023, Ubuntu 22.04+, Red Hat Enterprise Linux 9+) support Graviton4 out of the box. Windows workloads are not currently supported on Graviton4.

Migrating to Graviton4

AWS provides several tools to simplify migration to Graviton4-based EC2 instances (including m8g, c8g, and r8g instance families):

• Graviton Performance Advisor: Analyzes existing workloads to identify compatibility issues and optimization opportunities
• AWS Graviton Ready Program: Validates third-party software for ARM64 compatibility, with over 1,000 certified tools and applications
• Container Support: Docker and Kubernetes workloads can be re-platformed by building ARM64 container images, with no code changes for interpreted languages like Python, Node.js, and Java
• Test Environments: AWS offers free tier access to Graviton4 instances for testing before production migration

Conclusion

Graviton4 represents a major step forward for AWS's custom silicon strategy, delivering tangible 20% cost savings for compute-heavy EC2 workloads without sacrificing performance. For organizations running large-scale compute-bound workloads, migrating to Graviton4 can drive significant infrastructure savings while improving throughput. As ARM64 adoption grows in the cloud, Graviton4 solidifies AWS's lead in cost-efficient, high-performance compute.