Amazon Web Services Inc. today announced that it’s rolling out the fourth generation of its most energy-efficient high-performance custom chip for cloud workloads with the Graviton4.

The Graviton family of Arm-based processors is used by AWS to deliver high-performance and reduced costs in an existing broad range of cloud compute workloads in the Amazon Elastic Compute Cloud. Speaking to SiliconANGLE in an interview, Rahul Kulkarni, director of EC2 product management for AWS, said Graviton4 is a significant upgrade over Graviton3, with 30% better computing power, 50% more cores and 75% more memory bandwidth.

The Graviton platform started in 2018 with Graviton1, which Kulkarni said involved the foundational technology behind the EC2 platform, called Nitro, a lightweight hypervisor that allows for the virtualization of compute, storage, memory and networking options. At the time, AWS wanted to provide a full compute platform stack and Graviton’s Arm-based architecture worked well.

“With each passing generation, we have broadened the scope of applicable workloads that fit on Graviton,” said Kulkarni. The first generation worked as a proof of concept for web applications, generation two increased in scope for scale-out workloads and generation three built-in floating point and machine learning capabilities and high-performance computing.

“We have all of our top 100 customers running production workloads in Graviton, and we have over 50,000 customers running Graviton in a meaningful capacity,” added Kulkarni.

Currently, Graviton processors are available in more than 150 different instance types across EC2, which represent different “shapes,” for the different properties of compute, memory and storage available for the purposes that may suit customer needs. According to Amazon, the company has designed, tested and deployed more than 2 million Graviton CPUs to date across 30 regions.

At launch, the Graviton4 chip will be available in the Amazon EC2 R8g instance, which allows customers to run improved execution for high-performance databases, with in-memory caches and big data analytics workloads at scale. The R8g instance shape supports 8 gigabytes of memory per virtual processor and can roll out up to 192 processors. Kulkarni said this will be the first instance type that will launch, but the plan is to roll out Graviton4 to correspond to every x86 instance family.

Big customers including SAP Inc., Epic Games Inc. and SmugMug Inc. have already noted significant gains from using R8g instances with Graviton4, Amazon said.

“The numbers that we have say we’re seeing about 25% to 35% performance improvement that varies between Graviton3 and Graviton4,” Liz Fong-Jones, field chief technology officer at Honeycomb, told SiliconANGLE in an interview. Honeycomb, a full-stack software observability platform, has been using Graviton chips since Graviton2, around 2021, investing early on making a switch to the Arm-based architecture.

Fong-Jones said that when switching from x86-based chips to Graviton, overall performance “tail latency,” or how long it takes applications to respond to the majority of requests, improved greatly. Then, with each successive generation of Graviton processors, it’s becoming possible to do bigger workloads.

“What we saw from Graviton2 to Graviton3 and now from Graviton3 to Graviton4 is tail latency continues to remain rock solid and performance remains very consistent. What we’re seeing is a large improvement to throughput and reduced CPU utilization to do a certain workload,” said Fong-Jones. “This means that you can throw more workload at a given quantity of instances of the same size.”

Given that the first Graviton chip came off the AWS assembly line in 2018, the company has produced custom silicon at a rate of one chip on average of under one and a half years. When asked if the company intends to keep this drumbeat, Kulkarni said: “It’s an absolute, resonating, resounding yes.”

“And it’s not just a forward-looking comment,” Kulkarni added. “It’s a forward-looking comment based on past history, and that is largely predicated on the maturity of our entrepreneur lab and silicon team that can deliver silicon of this quality at this predictable cadence.”

He explained that AWS cannot afford multiple iterations on any new generation as silicon pricing swings can mean cost structures going up. That could affect the company’s ability to pass on cost savings to customers, which is a core tenet of the Graviton family of chips with its price-performance value proposition.

Source: JLStock