AWS Graviton Processors: An In-Depth
Look at AWS’s ARM-Based Evolution

by Barış Balta

Feb 10, 2024

7 min read

SHARE ON SOCIAL MEDIA

AWS Graviton processors, custom-designed by Amazon Web Services (AWS) on ARM architecture, have marked a major shift in cloud computing, offering high-performance processing that’s also cost-effective and energy-efficient. AWS launched its first Graviton processor in 2018, which was followed by the improved version Graviton2 in 2020, the powerful Graviton3 in 2021, and the latest Graviton4 launched in 2023, which has further enhanced capabilities to handle demanding workloads. In this blog, I’ll share my thoughts and experiences about each Graviton generation, their technical capabilities, and try to show the evolution of the performance and efficiency achievements so far.

Evolution of AWS Graviton Processors

AWS Graviton processors are custom-built to deliver specific optimizations for cloud-native workloads. Each new generation has introduced major improvements in terms of core architecture, process technology, power efficiency, and performance. Let’s dive into each of these generations to understand their technical specs and capabilities.

Graviton (1st generation)

The first Graviton processor, released in 2018, was AWS’s entry into ARM-based cloud computing. It primarily targeted general-purpose, low-power workloads and demonstrated AWS’s commitment to designing custom hardware solutions.

Core Architecture	ARM Cortex-A72 (64-bit)
Process Node	16nm
Core Count	Up to 16 cores
Memory Bandwidth	Moderate (DDR4 support)
Primary Use Cases	Lightweight web servers, simple microservices, dev/test environments
Thoughts	Not ideal for heavy compute or data-intensive applications; limited floating-point and cryptographic performance

Graviton2 (2nd generation)

Launched in 2020, Graviton2 brought a major improvement in performance, efficiency, and workload versatility. Built on the ARM Neoverse N1 architecture, it expanded Graviton’s reach to handle more demanding applications.

Core Architecture	ARM Neoverse N1 (64-bit)
Process Node	7nm (smaller, more efficient than 16nm in Graviton1)
Core Count	Up to 64 cores with simultaneous multithreading (SMT)
Memory Bandwidth	High (DDR4-3200 support, ~205 GB/s)
Primary Use Cases	Web and application servers, in-memory caches, containerized workloads, batch processing
Thoughts	7x performance over first generation Graviton, 2x floating-point performance, optimized for cloud-native workloads

Graviton3 (3rd generation)

AWS introduced Graviton3 in 2021, optimized further to handle even more compute-intensive applications, including AI and ML tasks. Graviton3 built on the ARM Neoverse V1 architecture, adding specialized capabilities and greater energy efficiency.

Core Architecture	ARM Neoverse V1 (64-bit)
Process Node	7nm (enhanced process compared to Graviton2)
Core Count	Up to 64 cores, with 25% faster core performance than Graviton2
Memory Bandwidth	Very high (DDR5-4800 support, ~300 GB/s)
Primary Use Cases	Machine learning inference, scientific simulations, high-performance computing (HPC), financial modeling
Thoughts	Hardware-accelerated cryptographic and floating-point operations, support for bfloat16 for ML workloads, 2x performance for ML inference over Graviton2, Up to 60% lower power consumption compared to similar x86-based instances

Graviton4 (4th generation)

Graviton4 represents the latest iteration which was announced in late 2023, building on Graviton3’s strengths and adding enhancements aimed at maximizing efficiency in large-scale, demanding workloads while further reducing power usage.

Core Architecture	ARM Neoverse V2 (64-bit)
Process Node	Advanced 5nm (more energy-efficient than Graviton3’s 7nm)
Core Count	Up to 80 cores (increased to support parallel processing for data-intensive tasks)
Memory Bandwidth	Very high (DDR5-5200 support, up to 350 GB/s)
Primary Use Cases	Real-time data analytics, multi-model AI, high-performance cloud computing, big data processing
Thoughts	Enhanced ML support (3x improvement over Graviton3), accelerated network performance, and encryption/decryption capabilities, achieves up to 30% reduction in power per workload over Graviton3

Key Technical Advancements in Each Graviton Generation

Each generation of Graviton processors has introduced technical enhancements to meet the evolving needs of cloud-native workloads. Some areas where each generation has improved include core architecture, memory bandwidth, process technology, and workload optimization. Starting from ARM Cortex in Graviton to ARM Neoverse V2 in Graviton4, each generation introduced more efficient and powerful cores. Progressing from 16nm in Graviton to 5nm in Graviton4, allowing for higher transistor density and reduced power consumption. Enhanced DDR memory bandwidth from DDR4 in Graviton and Graviton2 to DDR5 in Graviton3 and Graviton4, enabling faster data access and greater concurrency. Graviton3 and Graviton4 introduced hardware acceleration for cryptographic operations and ML tasks, helping to reduce latency and power usage for such workloads.

Detailed Performance Comparison: Graviton vs. Graviton2 vs. Graviton3 vs. Graviton4

Each Graviton processor generation has been optimized for specific workloads, offering improved performance and cost-efficiency.

Graviton (1st gen): Ideal for low-power, lightweight applications such as basic web servers and simple backend services. (t type instances)

Graviton2 (2nd gen): Expanded usage to include data processing, containerized applications, and scalable web applications due to higher core count and memory bandwidth.

Graviton3 (3rd gen): Designed for compute-heavy workloads, such as machine learning inference, scientific simulations, and real-time processing. It provides hardware-accelerated cryptographic operations and floating-point calculations, making it ideal for AI/ML workloads.

Graviton4 (4th gen): The latest in the series, Graviton4 is tailored for large-scale, high-performance computing needs, including multi-model AI, big data processing, and complex analytics.

The following table provides a side-by-side comparison of the Graviton processor generations, showing improvements in key areas:

	Gravilton	Gravilton2	Gravilton3	Gravilton4
Core Architecture	ARM Cortex-A72	ARM Neoverse N1	ARM Neoverse V1	ARM Neoverse V2
Process Node	16nm	7nm	7nm	5nm
Max Core Count	16	64	64	80
Clock Speed	Up to 2.3 GHz	Up to 2.5 GHz	Up to 3.1 GHz	Up to 3.3 GHz
Memory Type	DDR4	DDR4-3200	DDR5-4800	DDR5-5200
Memory Bandwidth	~80 GB/s	~205 GB/s	~300 GB/s	~350 GB/s
Floating Point Performance	Moderate	2x Graviton	2x Graviton2	1.5x Graviton3
Cryptographic Acceleration	Basic	Improved	Hardware Accelerated	Enhanced Hardware Acceleration
Machine Learning Support	Limited	Basic	Bfloat16, 2x ML Inference	3x ML Inference over Graviton3
Energy Efficiency	Basic	2x Graviton	60% better than x86	30% lower power per workload
Primary Use Cases	Lightweight web apps	General-purpose compute, HPC	Machine learning, HPC, AI	Big data, real-time analytics, AI

Why Choose AWS Graviton over others?

Graviton processors offer a competitive alternative to traditional x86 processors, especially for businesses looking for an optimized cost and performance for their cloud-native applications. Graviton processors provide major price/performance benefits, energy efficiency, and are continuously optimized to meet evolving cloud requirements. Whether you are running basic web applications, data analytics, machine learning, or real-time processing, the Graviton family offers a suitable solution with a better price/performance ratio compared to typical x86 instances.

Graviton4's improvements in core count, memory bandwidth, and power efficiency demonstrate AWS’s commitment to pushing the boundaries of ARM architecture, making it a future-proof choice for the teams targeting to scale their cloud environments efficiently. As cloud computing evolves, AWS Graviton continues to set the standard for cost-effective, high-performance computing.

If you are looking for a new generation, high performance and low-cost computing solution; then compute solutions based on graviton CPUs are the right choice for your workload.