reInvent: AWS unveils its latest chip, Graviton3E — here’s what you need to know
At reInvent 2018, four years ago, Amazon’s cloud arm, Amazon Web Services (AWS), unveiled the first generation of its in-house-designed chip, known as the Graviton processor. It was the first ever such chip within the cloud industry, and many believed it would be a game changer for the public cloud market. It was, after all, the first time Arm architecture was rolled out for enterprise-grade utility at such a colossal scale.
It all started in 2015, when AWS acquired Annapurna Labs, with a vision of changing how the cloud was delivered with the Nitro system. The AWS Nitro system is a virtualization infrastructure that serves as the underlying platform for the next generation of EC2 instances. Amazon EC2 (Elastic Compute Cloud) provides scalable computing capacity in the AWS Cloud.
That means that using Amazon EC2 eliminates a client’s need to invest in hardware up front, so they can develop and deploy applications faster. Ultimately, the goal of AWS Nitro is to accelerate AWS innovation, reduce customer cost, increase security, and deliver new instance types, including bare-metal instances where customers can bring their own hypervisor.
Now, Graviton has become an extension of that innovation, driving more choice and extending experiences across the entire infrastructure stack. Ultimately, the Graviton family of Arm-based custom processors was developed by AWS to give customers superior high-performance computing in EC2 at a reduced cost.
“Performance can be hard to achieve when you refuse to budge on things like security and cost,” AWS Utility Computing Senior VP Peter DeSantis said in his keynote session on the first night of reInvent. Throughout his keynote, DeSantis centered his message around the balance of performance, costs and security when designing solutions and services.
That said, at this year’s reInvent, AWS took the stage to introduce its latest chip, the Graviton3E, thus upgrading the current line of the Graviton family. The new Graviton3E, according to DeSantis, is made with noticeable performance improvements over the standard Graviton3, delivering over 35% better performance for vector-based workloads.
AWS has continued to play an important role in showcasing the benefits of Arm IP to the cloud computing market, from AWS Nitro to the first introduction of Graviton instances in 2018 to today. In his keynote address at reInvent this year, DeSantis shared how the latest chip performed better when used for life sciences and financial modeling, among other things.
He also noted that the Graviton3E chip is optimized for vector and floating-point workloads, which are common in higher-performance computing, especially research involving large-scale data modeling such as finance, weather prediction, life sciences, materials sciences and chemistry. The original launch of the Graviton3 provided up to 80% performance improvements over Graviton2 for some applications and even significant boosts for encryption and video encoding.
AWS shares at reInvent what Graviton3E will support
According to DeSantis, Graviton3E will support a whole new set of EC2 instances, including the upcoming HPC7G instances for high-performance computing workloads with 200 gigabytes of dedicated network bandwidth. It is apparently available in multiple sizes up to 64 virtual CPUs and 128 GiB of memory. It is, however, important to note that these instances will not be online until sometime next year.
The Graviton3E will also be available for C7gn instances for networking-intensive workloads such as virtual networking appliances — firewalls, virtual routers, load balancers and similar services — data analytics and tightly coupled computing clusters. DeSantis said they are capable of supporting 200gbps of network bandwidth and 200% higher packet performance.
He also explained that both of those new instances will take advantage of the new Nitro 5 hardware hypervisor, which was also announced at the reInvent. The new fifth-generation Nitro card almost doubles compute capacity on board, supports 50% more DRAM bandwidth, 60% more packets per second, 30% lower latency and 40% better performance per watt, DeSantis explained.
Along with the new Nitro, DeSantis said that C7gn offers 50% better packet processing performance at the lowest latency and the highest throughput. That, according to him, was made possible because the team had doubled the number of transistors on the custom Nitro chips.