Amazon keeps the pressure on Intel, AMD with 192-core Graviton5 CPU

Steven

Amazon keeps the pressure on Intel, AMD with 192-core Graviton5 CPU

Summary

AWS has unveiled Graviton5: a single-socket Arm server CPU packing 192 Neoverse V3 cores on TSMC’s 3nm process. The chip brings a larger 192MB shared L3 cache, faster memory (7200 MT/s with 8800 MT/s DIMMs planned), and native PCIe 6 support. Amazon pairs Graviton5 with its Nitro 6 smartNIC to double network bandwidth to 100 Gbps and offload system services, and says M9g instances based on the chip deliver about 25% higher performance than Graviton4-based M8g instances.

AWS argues the move to a denser single-socket design reduces inter-socket latency seen in multi-CPU systems and benefits workloads such as gaming, high-performance databases, EDA and analytics. Graviton5-based instances are in preview now, with compute- and memory-optimised variants due next year. The launch further cements hyperscalers’ shift to custom Arm silicon, putting Amazon in closer contention with AMD, Intel and other cloud vendors building their own chips.

Key Points

  1. Graviton5 ships with 192 Arm Neoverse V3 cores on TSMC 3nm, consolidating prior dual-CPU setups into one socket.
  2. The chip includes a 192MB shared L3 cache and an improved memory subsystem (7200 MT/s, with 8800 MT/s support planned).
  3. AWS claims M9g instances based on Graviton5 deliver ~25% higher performance than Graviton4 (M8g) equivalents.
  4. Graviton5 is the first server CPU to support PCIe 6 out of the box.
  5. Nitro 6 smartNICs double network bandwidth to 100 Gbps and offload storage, networking and virtualisation work.
  6. Single-socket design reduces inter-core latency vs linked multi-socket CPUs, improving performance for latency-sensitive workloads.
  7. Instances are in preview now; compute (C9g) and memory (R9g) variants will follow, using the same chip with different memory/compute ratios.
  8. The launch highlights the broader industry shift to custom Arm silicon, with Microsoft, Google and others also rolling their own designs.

Content summary

Amazon’s Graviton5 represents a significant step in the company’s in-house chip programme. By packing 192 cores into a single socket and increasing cache and memory bandwidth, AWS aims to reduce the performance penalties that come from inter-CPU communication in multi-socket machines. The chip’s native PCIe 6 support and pairing with Nitro 6 smartNICs signal AWS’s focus on both raw core scale and system-level efficiency.

AWS positions Graviton5 as a versatile, high-utilisation part that can be deployed across a wide range of instance types simply by changing memory-to-core ratios — a strategy that helps keep costs down and utilisation high. The rollout comes at a time when other cloud providers are accelerating their own Arm CPU projects, making Arm the de facto architecture for hyperscaler custom silicon.

Context and relevance

This launch matters because it advances the economics and performance profile of hyperscaler-owned silicon. For cloud customers, denser Arm instances can mean better price-performance for many server workloads and reduced reliance on x86 vendors. For the wider industry, Graviton5 emphasises the momentum behind Arm in the datacentre and the competitive pressure now exerted on Intel and AMD by cloud providers designing vertically integrated stacks.

Organisations planning future deployments — especially for scale-out services, high-performance databases, gaming backends or EDA tasks — should evaluate Graviton5 instance pricing and performance as part of their procurement and optimisation workstreams. Developers and ops teams will also need to keep an eye on ecosystem support (OS, compilers, libraries) as Arm variants proliferate.

Author’s take (Punchy)

Big, fast chip. Same AWS playbook: build versatile silicon, run it at high utilisation, pass the savings to customers. If you’re running cloud workloads at scale, this isn’t background noise — it’s a direct lever on cost and performance. Read the details if you care about squeezing latency or budget out of your infrastructure.

Why should I read this?

Short version: if you run or buy cloud compute, this will affect price-performance choices. Graviton5 is a proper power move by AWS — denser cores, faster memory, less inter-socket faff. That means cheaper, faster instances for many workloads and more pressure on Intel/AMD. It’s worth a quick look to see whether your stacks can benefit.

Source

Source: https://go.theregister.com/feed/www.theregister.com/2025/12/04/amazon_graviton_5/