Skip to content

AWS Graviton4

The AWS Graviton4 is Amazon Web Services' fourth-generation custom ARM-based processor, introduced in 2024. Building upon the advancements of its predecessors, Graviton4 offers significant improvements in performance, memory bandwidth, and scalability, making it ideal for a wide range of cloud workloads.

Key Specifications

  • Architecture: ARMv9.0-A (64-bit RISC)
  • Core Count: Up to 96 cores based on ARM Neoverse V2 microarchitecture
  • Clock Speed: 2.7 GHz (192 cores) / 2.8 GHz (96 cores)
  • L2 Cache: 2 MB per core (192 MB total)
  • Memory Support: 12 channels of DDR5-5600 ECC memory
  • Memory Capacity: Up to 3 TiB
  • PCIe Support: 32 lanes of PCIe 5.0
  • Chiplet Design: 7-chiplet configuration with compute, memory, and I/O chiplets
  • Security Features: Always-on memory encryption, Branch Target Identification (BTI), and encrypted high-speed interfaces

Architectural Enhancements

1. ARM Neoverse V2 Cores

Graviton4 utilizes ARM's Neoverse V2 cores, offering improved performance per clock cycle compared to previous generations. These cores are designed to handle demanding workloads efficiently, providing a balance between performance and power consumption.

2. Advanced Memory Architecture

The processor supports 12 channels of DDR5-5600 ECC memory, providing up to 160% more memory bandwidth compared to Graviton3. This increase in bandwidth is crucial for data-intensive applications, ensuring faster data access and processing.

3. Enhanced Floating-Point and Cryptographic Performance

Graviton4 delivers up to 60% higher processing power compared to Graviton3, with improvements in floating-point and cryptographic workload performance. These enhancements are beneficial for applications requiring intensive mathematical computations and secure data processing.

4. Machine Learning Optimizations

The processor includes support for bfloat16, a format commonly used in machine learning models. This optimization enables more efficient processing of AI workloads, leading to faster inference times and reduced energy consumption.

EC2 Instance Families Powered by Graviton4

AWS offers several EC2 instance families powered by Graviton4 processors:

  • R8g: Memory-optimized instances ideal for high-performance databases, in-memory caches, and real-time big data analytics.
  • X8g: Memory-optimized instances with up to 3 TiB of DDR5 memory, suitable for memory-intensive workloads such as SAP HANA and Electronic Design Automation (EDA).
  • C8g: Compute-optimized instances designed for high-performance computing (HPC), batch processing, and scientific modeling.
  • M8g: General-purpose instances suitable for a variety of workloads, including web servers and small to medium-sized databases.
  • I8g: Instances optimized for I/O-intensive workloads, offering high throughput and low latency for applications like AdTech and real-time analytics.

Performance and Efficiency

Graviton4 processors offer up to 40% better performance and up to 29% better price-performance compared to Graviton3-based instances. These improvements are achieved through architectural enhancements, increased memory bandwidth, and optimized power efficiency.

Industry Adoption

Several organizations have adopted Graviton4-based instances for their workloads:

  • Epic Games: Found Graviton4 instances to be the fastest EC2 instances they have ever tested, enabling improved performance for their applications.
  • Honeycomb.io: Achieved more than double the throughput per vCPU compared to the non-Graviton-based instances they used four years ago.
  • SmugMug: Observed 20–40% performance improvements using Graviton4-based instances compared to Graviton3-based instances for their image and data compression operations.

Summary

The AWS Graviton4 processor represents a significant advancement in custom ARM-based silicon, offering improved performance, memory bandwidth, and scalability for cloud workloads. Its architectural enhancements and optimizations make it a compelling choice for organizations looking to optimize their cloud infrastructure.