Specifications
Core Count
192 Arm Neoverse V3 cores per socket
Architecture
Armv9.2-A (Neoverse V3 'Poseidon' class core)
Package Design
Four-chiplet design in a single socket — no cross-socket NUMA hop
Process Node
3 nm
Frequency
3.3 GHz
L3 Cache
~180 MB shared L3, roughly 5x Graviton4
Memory
12-channel DDR5-8800
Memory Speed Jump
DDR5-8800 vs DDR5-5600 on Graviton4
Host Interconnect
PCIe Gen 6
Inter-core Latency
Up to 33% lower than the two-socket Graviton4 layout
EC2 Instance Families
M9g / M9gd (general purpose), C9g (compute optimised), R9g / R9gd (memory optimised)
Compute vs Graviton4
Up to 25% faster
Database / Web vs Graviton4
Up to 30% faster databases, up to 35% faster web applications
Machine Learning
Up to 35% faster ML workloads vs M8g
Network / Storage
15-20% more network and storage bandwidth
Availability
M9g and M9gd generally available June 2026 (preview from December 2025); R9g family launched 2026
Generation History
Graviton3 (Neoverse V1) → Graviton4 (Neoverse V2) → Graviton5 (Neoverse V3)
Overview
AWS Graviton5 is the fifth generation of Amazon's in-house Arm server processor and the densest Armv9.2 CPU in the Graviton line. Instead of pairing two 96-core dies as Graviton4 did, Graviton5 places 192 Neoverse V3 cores on a single socket built from four chiplets, which removes the cross-socket memory hop and cuts inter-core latency by up to a third.
Each package is fabricated on a 3 nm node, runs at 3.3 GHz, carries roughly 180 MB of shared L3 cache — about five times Graviton4 — and supports 12 channels of DDR5 at 8,800 MT/s with PCIe Gen 6 connectivity. For inference-heavy cloud services the generational jump in memory speed and cache capacity matters more than core count alone: it lifts database, web-serving and machine-learning throughput by 25-35% over the previous generation.
Graviton5 ships across the M9g/M9gd general-purpose family, C9g compute-optimised instances and the R9g/R9gd memory-optimised family, all sharing the same silicon with different memory-to-core ratios. AWS has also stated that the majority of new compute capacity it adds now runs on its own silicon. QS Compute supplies Arm-based cloud and on-premises AI infrastructure — contact us for configuration and volume pricing.
Key Benefits
Single-socket density: 192 Neoverse V3 cores with no NUMA penalty. Memory bandwidth: 12-channel DDR5-8800 for bandwidth-hungry inference. Cache headroom: ~180 MB shared L3 for large working sets. Proven software: full Armv9.2-A support across mainstream Linux and ML frameworks.
Applications
Cloud-native AI inference and model serving, large-scale databases and caching tiers, web and application hosting, container platforms and microservices, data analytics, and cost-optimised Arm instances for agentic AI control planes.
Request a Quote — AWS GRAVITON5 — 192-CORE ARM SERVER PROCESSOR FOR CLOUD AI
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →