Graviton5’s improved design increases speed and energy efficiency — beyond Moore’s law

2026-06-10 · Source: Amazon Science homepage · Field: Technology & Digital — Cloud Computing & IT Infrastructure, Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Advanced, short

Summary

The AWS Graviton5 processor, now generally available in M9g and M9gd EC2 instances, features a new four-chiplet architecture with 192 cores and custom die-to-die links providing up to 420 gigabytes per second of bandwidth. This design doubles the core count from Graviton4 and includes 192 MB of L3 cache, over five times more than its predecessor. Graviton5 supports DDR5-8800 memory and PCIe gen6 interconnects, utilizing a three-nanometer process for increased circuit density. Each core offers 25% better performance, and the Neoverse V3 core significantly improves branch prediction, leading to up to 30% better performance for real applications like databases. Overall, Graviton5 delivers up to 25% better computational performance than Graviton4, with specific gains of up to 35% for web applications and machine learning inference, and 30% for databases. It also introduces the Nitro Isolation Engine, a formally verified cloud hypervisor, enhancing VM security.

Key takeaway

For cloud architects evaluating compute options for demanding workloads, Graviton5-powered M9g and M9gd instances offer substantial performance and efficiency improvements. You can expect up to 25% better computational performance, with specific gains for web applications, ML inference, and databases. Consider migrating existing Graviton4 workloads or new deployments to these instances to benefit from the enhanced core count, faster memory, and mathematically proven VM isolation provided by the Nitro Isolation Engine.

Key insights

Graviton5's chiplet architecture and advanced components deliver significant performance and efficiency gains for diverse cloud workloads.

Principles

• Chiplet designs enhance scalability and bandwidth.
• Formal verification strengthens cloud security.
• Optimize for real-world application performance.

Method

Graviton5 employs a four-chiplet design with custom die-to-die links, DDR5-8800, PCIe gen6, and a 3nm process, integrating Neoverse V3 cores and a formally verified Nitro Isolation Engine.

In practice

• Deploy M9g/M9gd for general-purpose tasks.
• Use for agentic AI and ML inference.
• Leverage for database and web application hosting.

Topics

• AWS Graviton5
• Chiplet Architecture
• EC2 Instances
• Neoverse V3 Cores
• Nitro Isolation Engine
• Cloud Security

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Hardware Engineer, AI Architect, MLOps Engineer

Related on AIssential

• In another wild turn for AI chips, Meta signs deal for millions of Amazon AI CPUs — AWS secured a significant Meta deal for Graviton CPUs, intensifying cloud and custom AI chip competition.
• Apple's M5 Pro & Max Architecture is INSANE — Apple's M5 Fusion architecture uses asymmetric dual-die design for scalable performance and efficiency in portable devices.
• Helping data centers deliver higher performance with less hardware — Sandook intelligently balances SSD workloads to significantly boost data center performance and extend hardware lifespan without new equipment.
• Using Accelerated Computing to Live-Steer Scientific Experiments at Massive Research Facilities — GPU-accelerated computing enables real-time analysis and steering of exascale scientific experiments, drastically reducing data processing times.
• Deep Dive Into 4-Wave Interleave FP8 GEMM — The 4-wave interleave FP8 GEMM optimizes GPU performance by fine-grained software pipelining and memory access patterns.
• GigaDevice Expands GD25UF Series Density Empowering AI Computing with 1.2V Ultra-Low Power Storage — GigaDevice's GD25UF series expands SPI NOR Flash density, enabling low-voltage AI and compact, power-efficient devices.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Amazon Science homepage.