The Story Behind AWS's 2026 Graviton4 Launch: Neoverse V2 and EC2 Optimizations

In 2026, AWS officially launched Graviton4, the fourth generation of its custom Arm-based processors, marking a major milestone in cloud compute efficiency. Built on Arm’s Neoverse V2 platform, Graviton4 delivers significant performance and efficiency gains over its predecessor, Graviton3, with deep optimizations for Amazon EC2 to serve diverse cloud workloads.

Neoverse V2: The Architectural Backbone of Graviton4

Graviton4 is the first AWS custom processor to adopt Arm’s Neoverse V2 microarchitecture, which complies with the Armv9.2-A instruction set. Key Neoverse V2 features integrated into Graviton4 include:

• Support for Scalable Vector Extensions 2 (SVE2) for accelerated high-performance computing (HPC) and machine learning (ML) workloads
• Native Bfloat16 and INT8 data type support for 2x faster ML inference compared to Graviton3
• Up to 64 Neoverse V2 cores per socket, with 1MB L2 cache per core and a shared 64MB L3 cache, doubling L2 cache capacity over Graviton3
• DDR5-6400 memory controllers, delivering 50% higher memory bandwidth than Graviton3’s DDR4-3200 interface
• Enhanced security features including Arm Memory Tagging Extension (MTE) v2 and hardware-enforced control flow integrity

AWS worked closely with Arm to customize Neoverse V2 for cloud workloads, adding proprietary logic for EC2-specific use cases like container orchestration, serverless compute, and managed database services.

EC2 Optimizations Tailored for Graviton4

AWS paired Graviton4 with a suite of EC2 optimizations to maximize performance per watt and reduce total cost of ownership (TCO) for customers. These include:

• New EC2 Instance Families: Launch-day availability of M8g (general purpose), C8g (compute optimized), and R8g (memory optimized) instances, with plans for X2g (memory intensive) and I8g (storage optimized) later in 2026.
• Nitro System v6 Integration: Graviton4 is tightly coupled with the sixth generation of AWS’s Nitro System, offloading network, storage, and security processing to dedicated hardware to free up core cycles for customer workloads. Nitro v6 delivers up to 200Gbps of network bandwidth and 80Gbps of Amazon Elastic Block Store (EBS) bandwidth per instance.
• Memory and Storage Tuning: EC2 instances for Graviton4 support up to 12TB of DDR5 memory, with sub-microsecond memory latency optimizations for in-memory databases like Amazon Aurora and ElastiCache. Local NVMe SSD storage options deliver 30% higher IOPS than Graviton3-based instances.
• Workload-Specific Optimizations: AWS pre-optimized popular open-source tools including Docker, Kubernetes, PostgreSQL, and TensorFlow for Graviton4, with 40% faster container startup times and 25% higher throughput for OLTP databases.

Projected Performance and Efficiency Gains

Internal AWS benchmarks project Graviton4 to deliver:

• 40% higher integer compute performance over Graviton3 for web and application workloads
• 2x faster ML inference for BERT and ResNet-50 models, thanks to Neoverse V2’s Bfloat16 and INT8 support
• 35% better energy efficiency per core, reducing carbon footprint for sustainability-focused customers
• 50% higher memory bandwidth for data-intensive workloads like big data analytics and real-time streaming

Ecosystem and Availability

Graviton4-based EC2 instances launched in 16 AWS regions globally in Q1 2026, with support for Amazon Linux 2026, Ubuntu 24.04 LTS, Red Hat Enterprise Linux 10, and Windows Server 2026. AWS also expanded its Graviton Ready program, with over 500 independent software vendors (ISVs) certifying their applications for Graviton4 at launch.

Conclusion

The 2026 Graviton4 launch represents years of collaboration between AWS and Arm, combining cutting-edge Neoverse V2 architecture with deep EC2 optimizations to deliver industry-leading cloud compute performance and efficiency. For customers migrating from x86 or older Graviton generations, Graviton4 offers a compelling path to lower costs and higher performance across nearly all cloud workloads.