Amazon Web Services Inc.’s next-generation custom silicon is finally being made accessible to customers for the first time with the launch of the Elastic Compute Cloud M9g and M9gd instances.

They’re powered by the all-new Graviton5 central processing unit, which is designed to deliver superior compute performance for diverse cloud workloads, including artificial intelligence applications.

The company said Graviton5 is designed to support real-time AI reasoning, code generation, multistep task orchestration and complex, always-on workloads to enable “AI that takes action.” By that, it’s referring to autonomous AI agents that can be set to work performing tasks on behalf of humans with minimal supervision.

AWS announced Graviton5 during its annual AWS re:Invent conference in December, saying it represents a “massive architectural leap” for enterprise cloud infrastructure.

The chip is the most powerful in the Graviton family to date, purpose-built for the demands of agentic AI, which requires CPUs to run large numbers of concurrent computing environments and maximize the performance of graphics processing units and other AI accelerators. According to AWS, Graviton 5 delivers a 25% improvement in terms of overall compute performance compared to its predecessor, Graviton4. It does this while providing greater energy efficiency, allowing companies to reduce their infrastructure costs.

AWS said the new chip has already won over a lot of fans. For instance, Meta Platforms Inc. is planning to deploy “tens of millions” of Graviton5 cores to run its agentic workloads, and the cloud data warehouse company Snowflake Inc. has also signed up. Uber Technologies Inc. is another early adapter that has committed to deploying the processors.

The cloud giant believes there’s going to be massive demand for the M9g and M9gd compute instances now they’re generally available. More than 120,000 AWS customers globally are already running applications on earlier generation Graviton CPUs, and the new chips are the first in the family to support DDR5 memory and PCIe, packaging an ultra-dense 192 cores on a single chip.

Graviton5 also comes with a five-times bigger L3 cache, which is a high-speed memory buffer that’s designed to keep frequently accessed data closer to the heart of the chip. This means each Graviton core can access 2.6 times more L3 cache than Graviton4. Memory performance has also been enhanced, reducing the time the chips spend waiting for data to arrive. As a result, it enables faster application response times and runs memory-intensive processes much more efficiently, AWS said.

Elsewhere, the network and overall storage capacity have been increased. AWS said Graviton5 has 15% more network bandwidth and 20% greater Amazon Elastic Block Store bandwidth. This means that larger instances can have up to 100% greater bandwidth overall, which translates to more rapid data transfers, faster backups and enhanced performance for distributed applications.

In addition to the 25% overall compute performance boost, the M9g instances enable 35% faster web applications and machine learning inference workloads, and 30% faster databases. The M9gd instances differ because they’re targeted at workloads that require high-speed local storage, offering up to 11.4 terabytes of NVMe solid-state drive storage and 30% faster input/output operations per second than earlier generation chips.

AWS explained that the new instances work in tandem with AWS Nitro Cards to offload networking, storage and virtualization to dedicated hardware. Nitro is Amazon’s security and performance layer that’s designed for privacy-conscious organizations in industries, such as healthcare, financial services and government. By offloading these needs, Nitro can help the Graviton5 CPUs to run more efficiently while improving data security.

The instances are the first to come with support for the new Nitro Isolation Engine, which enhances Nitro’s security capabilities with the use of formal verification. It provides mathematical certainty that workloads are running in isolation from each other and other AWS operators.

Photo: SiliconANGLE