While AMD and Intel continue to struggle with rising TDP, chiplet cascades and supply chain issues, Amazon Web Services (AWS) is taking a radically different approach with its new Graviton5 CPU – quietly but decisively. At its own “Re:Invent”, AWS presented its new ARM processor with 192 cores, manufactured in a 3 nm process at TSMC. A monolith instead of a double socket compromise. More cache, faster cores, increased efficiency. And a hidden ace up its sleeve: A formally verified hypervisor that makes everything else on the market look old in terms of security.

But let’s start with the hard facts: Graviton5 accommodates 192 ARMv9 cores on a single die for the first time, no multi-socket trickery like its predecessor. It is based on ARM’s new Neoverse V3 cores, which clock at 3.1 GHz and, according to AWS, are around 11% faster than Graviton4. But a pure clock speed increase is not the story. The decisive step lies in the system design: 12 DDR5 controllers with support for 8400 MT/s and eight PCIe 6.0 controllers give an idea of where the journey is heading. Although the theoretical memory bandwidth of 806 GB/s per die is slightly lower than with two Graviton4 chips (1.07 TB/s), there is almost triple the L3 cache, a whopping 192 MB for all cores.
And this cache is not a nice extra, but a targeted architectural policy. Because AWS has recognized this: In distributed workloads, memory latencies are often the real bottleneck. Graviton5 aims to break this with a monolithic structure, coherent interconnect and improved memory subsystem. Customer benchmarks show that the concept works: AWS speaks of a 25% to 60% increase in performance in real workloads compared to its predecessor. The next noteworthy point is the cooling. Graviton5 comes without a heatspreader, which is otherwise only common for OC CPUs in the enthusiast segment. However, AWS goes even further and relies directly on water cooling in its racks. This reduces the thermal resistance chain and enables a higher performance level with improved energy efficiency at the same time. In line with hyperscaler logic: less infrastructure loss, more load per square meter of data center.
But hardware alone does not make a quantum leap. AWS combines Graviton5 with its own “Nitro” accelerators, dedicated chips that are responsible for networking, security and virtualization. And this is where it gets exciting: the new Nitro hypervisor is written in Rust, but that’s not all. It has been formally verified, i.e. proven mathematically correct. This is more than just marketing speak. The Isolation Engine reduces the functional scope of the hypervisor to a minimum, deliberately removing critical rights and thus enabling an architecture in which attack vectors have been radically minimized. This is a game changer for security-critical cloud customers, such as public authorities, financial institutions or pharmaceutical companies.
From a strategic perspective, Graviton5 is a clear challenge to AMD and Intel. While both are still wrestling with hybrid CPU designs, chiplet validation and CPU-GPU mergers, Amazon delivers: a vertically integrated CPU package of its own design, its own infrastructure and a target group that AWS already controls. Over 50% of newly installed CPU capacity in AWS data centers is now based on Graviton chips, according to the company, and the trend continues to point upwards. This is no longer a tech experiment. This is a major assault on x86 supremacy.
Of course, not everything is gold: the memory bandwidth is not quite on par with the dual Graviton4 variant, the tooling around ARM in the cloud environment is not yet consistently on par with x86, and developers still have to relearn, both at low level and in debugging, performance analysis and deployment processes. But AWS is moving fast. The EC2 M9g instances with Graviton5 are already running in preview, with variants for storage-intensive (R9g) and compute-intensive (C9g) applications to follow in 2026. In a market where x86 providers are struggling with “AI PC” buzzwords and PowerWalls, AWS is focusing on throughput, integration and security. Quietly, but unmistakably.
Conclusion:
Graviton5 is not a revolution in a single component, but a precisely orchestrated evolution of the entire server stack. Monolith instead of chiplet proliferation. Cache instead of a bandwidth trick. Formal security instead of security promises. What is happening here is the slow but systematic transformation of the cloud architecture, away from generic x86 goods and towards customized ARM technology with a controlled pipeline. And if AMD and Intel can’t counter this, Graviton will soon be more than just a footnote in the cloud business. It will become the new standard.
| Source | Key message | Link |
|---|---|---|
| AWS re:Invent – Graviton5 announcement | Presentation of the new Graviton5 CPU with 192 ARMv9 cores, production at TSMC in 3 nm process, cache structure, DDR5 and PCIe 6.0 subsystem | https://aws.amazon.com/blogs/aws/introducing-aws-graviton5-the-most-powerful-aws-designed-processor-ever |
| AWS EC2 – M9g instance preview | Technical details on the new EC2 M9g instances based on Graviton5, performance values and target groups | https://aws.amazon.com/ec2/instance-types/m9g |
| AWS Nitro System Documentation | Description of the Nitro hypervisor, formal verification, security architecture and role of the Nitro accelerators | https://aws.amazon.com/ec2/nitro |
| ARM – Neoverse V3 architecture | Technical basics of the Neoverse V3 cores used in Graviton5 | https://www.arm.com/products/neoverse/v3 |
| TSMC – 3 nm process family overview | Information on the N3 generation, manufacturing process, efficiency and technical parameters | https://www.tsmc.com/english/dedicatedFoundry/technology/3nm |
| AWS Infrastructure Blog – Liquid Cooling | Classification of the new liquid cooling approaches in AWS data centers, background information on server cooling | https://aws.amazon.com/blogs/infrastructure/introducing-aws-direct-liquid-cooling |
| AWS – Customer Benchmarks & Performance Summary | Official measured values from real workloads with 25 to 60 percent performance increase compared to Graviton4 | https://aws.amazon.com/blogs/aws/aws-graviton5-performance-benchmarks |

































6 Antworten
Kommentar
Lade neue Kommentare
Urgestein
Urgestein
Urgestein
Urgestein
Veteran
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →