AWS recently announced the new Graviton5 processor and the preview of the first EC2 instances running on it, the general-purpose M9g instances. According to the cloud provider, the latest chip delivers up to 25% higher performance than Graviton4, introduces the Nitro Isolation Engine, and provides a larger L3 cache, improving latency, memory bandwidth, and network throughput.
According to the press release, the new Arm-powered EC2 M9g instances provide up to 192 CPU cores per instance. The higher core density reduces inter-core latency by up to 33% and increases bandwidth, improving scaling for workloads such as databases, analytics, application servers, gaming, and Electronic Design Automation (EDA).
Graviton5 adds the Nitro Isolation Engine to the Nitro System, a new engine that uses formal verification to prove that workloads are isolated from each other and from AWS operators. The engine has a small, verified codebase, and AWS will give customers access to the implementation and its proofs for review. Mohamed Mediouni, kernel/hypervisor engineer at AWS, writes:
A vendor hypervisor (because that’s what that actually means) – in practice, a replacement to KVM – is an interesting thing to have for sure. Let’s see where this whole story will actually lead.
The session “Introducing Nitro Isolation Engine: Transparency through Mathematics” was recently presented at re:Invent, and is now available on YouTube. According to Ali Saidi, VP and distinguished engineer at Amazon, “the Nitro Isolation Engine is harnessing Rust and formal verification to create a formally verified cloud hypervisor, pioneering a new standard for mathematically proven cloud security.”
Graviton adoption among AWS customers has grown significantly: according to the cloud provider, over 50% of new CPU capacity added by AWS is on Graviton, and during the latest Prime Day, Graviton powered more than 40% of the EC2 compute used by Amazon.com.
Graviton5 includes a 5× larger L3 cache than Graviton4, giving each core 2.6× more L3 cache, which reduces memory access latency. Memory speeds are higher than those of Graviton4-based M8g instances, improving performance for large and memory-intensive workloads. According to the announcement, network bandwidth is up to 15% higher on average, EBS bandwidth is up to 20% higher, and the instances support up to 2× the network bandwidth compared to prior generations. On Hacker News, user diath comments:
No benchmarks. No FLOPs. No comparison to commodity hardware.(…) “9 is faster than 8 which is faster than 7 which is faster than 6, …, which is faster than 1, which has unknown performance”.
On Reddit, comments have been mixed, with most concerns about the lack of details and the instances being only available in preview in a subset of regions. User Ill-Side-8092 writes:
While a natural incremental release, this is good to see and the Annapurna team is a rare exception of real innovation happening in AWS at the moment. There is also a reality that custom silicon is now more of a table stakes asset amongst big tech though with Google, Apple, and Microsoft all also having their own very respectable silicon offerings.
Other users highlight that region parity is still far from being achieved on AWS, with customers facing challenges due to parity differences across services and their features. User Rude_Walk comments:
Maybe finish the rollout of Graviton4 first? It is not even available in Singapore which is a major region.
According to AWS, early adopters include workloads at Adobe, Epic Games, and Pinterest. A signup page is available to request access to the M9g instance preview.
