惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

月光博客
月光博客
小众软件
小众软件
爱范儿
爱范儿
Y
Y Combinator Blog
博客园 - Franky
美团技术团队
博客园 - 【当耐特】
The Cloudflare Blog
罗磊的独立博客
Hugging Face - Blog
Hugging Face - Blog
Jina AI
Jina AI
IT之家
IT之家
人人都是产品经理
人人都是产品经理
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
大猫的无限游戏
大猫的无限游戏
Apple Machine Learning Research
Apple Machine Learning Research
博客园 - 聂微东
WordPress大学
WordPress大学
V
Visual Studio Blog
博客园_首页
阮一峰的网络日志
阮一峰的网络日志
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
有赞技术团队
有赞技术团队

The New Stack | DevOps, Open Source, and Cloud Native News

Agentic development hinges on verification. For cloud-native software, that is a runtime problem. AI agents need infrastructure: Why Europe’s regional cloud strategy matters Transform your AI coding agent into a deterministic Java Spring expert WeAreDevelopers is coming to the US to give unsung developers a bigger voice Cleaner AI training data, fewer bugs: Sonar’s SonarSweep explained Observability overload is drowning engineers Google’s DiffusionGemma is 4x faster than its other Gemma models Fable 5: Guardrails and burn rate are annoying users, who say it’s still better than Opus 4.8 The Anthropic leader who built Claude Code says he ditched prompting — now he just writes loops. Microsoft pulled 73 GitHub repos after malware attack — but still won’t say who’s compromised Databricks wants to kill the “email me a file” problem for AI agent skills Ramp bets forward deployed engineers can do what off-the-shelf finance AI can’t Git real: AI agents aren’t just for solo developers anymore Anthropic launches Claude Mythos/Fable 5, but you better try it soon Spring is 23 years old. AI just made it a security emergency. This AI agent startup ditched Anthropic for DeepSeek — and says it’s saving millions When your data model is the bottleneck: lessons from Medium’s feature store How long before we stop reading the code? The tokenmaxxing party is over, and Revenium is mopping up How AI is solving the memory crunch it created Microsoft’s pitch to enterprises: Ditch Azure Repos for GitHub, despite its rocky reliability record Claude Code’s biggest upgrade yet ran 5 agents at once — here’s what happened Why Anthropic just doubled Claude Cowork limits at no charge For years, Apache Cassandra handed this work to your team — 6.0 takes it back “A dangerous combination”: The 2 factors that can “corrupt” AI agent workflows With Foundry, Microsoft bets the enterprise AI battle is about reliability, not capability Microsoft unlocks Visual Studio for developers left behind by its own AI AI teams now deploy 1,000 times a month. Your pipeline wasn’t built for that. Microsoft just made the agent runtime free — and kept everything around it “Whoever builds the most joyous product wins”: The agent war begins
AWS can now mathematically prove your VMs are isolated
Frederic Lardinois · 2026-06-11 · via The New Stack | DevOps, Open Source, and Cloud Native News

On Wednesday, AWS announced that its Graviton5 processor is now generally available and powers two new Amazon EC2 instances: M9g for general-purpose workloads and M9gd for workloads that require high-speed local storage. These are the first Graviton5 instances since AWS previewed the chips at re:Invent 2025.

The new chips double Graviton4’s core count from 96 to 192, and AWS says the new instances will deliver up to 25% better compute performance than the previous generation. 

AWS says the new instances will deliver up to 25% better compute performance than the previous generation. 

What’s maybe more interesting, however, is that these new instances also run on the sixth-generation Nitro System, which now includes the Nitro Isolation Engine, a new, formally verified security component that AWS describes as “a separation kernel whose sole job is isolating virtual machines from each other.”

A kernel whose only job is isolation

The idea of a separation kernel has been around for a while. John Rushby coined the term in 1981. His core idea was that a standard OS kernel, even at the time, had become far too large to formally verify. A smaller, specialized separation kernel, however, would still allow formal verification due to its relatively limited complexity.

AWS launched its Nitro system and hypervisor in 2017, and it has enforced isolation in EC2 ever since. But Nitro also handles substantial business logic and manages device drivers and other AWS-specific features, meaning it was not designed for formal verification from the start.

With this new system, the Nitro hypervisor still handles policy, including VM creation, resource allocation, migration, and scheduling, but it is now somewhat deprivileged and must ask the Nitro Isolation Engine (NIE) to perform any operation that touches guest state, and the Isolation Engine checks every request before acting.

“Distilling the hypervisor’s security-critical isolation logic into a minimal component, the NIE, makes it small enough to verify and audit, giving customers unprecedented visibility into how isolation is enforced,” AWS’s Dominic Mulligan and Nathan Chong write. “We also wrote NIE in Rust, a language that lends itself more naturally to formal verification.”

How we got here

Some of the earlier work in this area includes Columbia’s SeKVM project, which was the first formally verified commercial-grade hypervisor in 2021. But it looks like this was mostly a research project that never ran in a commercial cloud. 

AWS itself credits seL4, a project that demonstrated that OS verification was feasible. 

To prove the kernel behaves correctly, AWS used the Isabelle/HOL proof assistant. AWS says the model and proof “comprise 330,000 lines of machine-checked mathematics,” making it comparable to the seL4 project. “However, unlike seL4, NIE is designed for a commercial cloud environment and ships on production hardware as an always-on feature for Graviton5 users,” AWS writes in its announcement.

192 cores: a substantial redesign

The chip itself got a substantial redesign. Graviton5 moves from TSMC’s 4nm process to 3nm, and while Graviton4 put all 96 cores on a single chiplet with PCIe and DDR controllers on their own chiplets, Graviton5 divides its 192 cores across four chiplets, each with its own controllers, which puts the memory controllers closer to the cores. 

AWS claims that, together with faster memory chips, this will allow web applications to run up to 35% faster and ML inference up to 35% faster than on Graviton4. 

Graviton and agentic AI

AWS positions Graviton5 as purpose-built for agentic AI. That pitch, which isn’t all that different from Google’s pitch for its ARM processors, seems to be working. Meta signed a multibillion-dollar agreement in April to deploy tens of millions of Graviton cores for its agentic AI work, and Snowflake committed $6 billion over five years in May. AWS says Uber is also deploying Graviton for agentic workloads.

In total, AWS says, more than 120,000 customers currently use Graviton, and that more than half of new CPU capacity added to AWS has been Graviton for the third year running. 

AWS may have a bit of an early mover advantage here. Microsoft’s Cobalt 200 is still in preview, Google’s latest Axion chips only went GA in January, and Nvidia’s Vera — maybe the most hyped of all of these chips — will arrive in the second half of the year.

YOUTUBE.COM/THENEWSTACK

Tech moves fast, don't miss an episode. Subscribe to our YouTube channel to stream all our podcasts, interviews, demos, and more.

Created with Sketch.