Since the 2019 launch of AWS Graviton2, based on the Arm Neoverse N1 core, Arm and our partners have continued to push the performance of Arm-based designs. The AWS Graviton processor is now in its fourth generation and with the General Availability of AWS Graviton4, we wanted to reflect on the progress seen over the past few years.
At Arm, a critical stage of our silicon design process is RTL simulation. In this process, validation engineers take designs expressed in Verilog and use an RTL simulator (such as Siemens' Questa Advanced Simulator or Cadence's Xcelium) to show how the design will function. By defining the expected output given a particular input, it is possible to use an RTL simulator to validate that the design performs as expected. This is essential for ensuring a high quality product, but it is incredibly compute intensive. In the years since the launch of the Graviton2 we have been steadily migrating our simulations to Arm based compute, and now we perform well over half of our RTL simulations on AWS Graviton-based Amazon EC2 instances.
Graviton2 was the first platform to offer us performance better than x86. When Graviton2 launched, the Intel Xeon-based M5 instance was the most modern existing instance type in EC2. Compared to M5, the Graviton2-powered M6g instance gave us 20% higher performance at 20% lower cost per vCPU hour, delivering the same simulation results for 40% less cost. However, if raw performance was the key goal, users could use twice the number of M5 vCPUs, disable Symmetric Multi-Threading, and outperform the M6g.
At Arm, our regression simulations typically run overnight and consist of large numbers of independent tests. So raw performance was not an issue for these tests – they need throughput and so run with SMT enabled. However, our engineers also need to run some simulations during the day and in these cases a faster turnaround is desirable.
AWS Graviton3, implementing the Neoverse V1 core, improved on the performance of Graviton2 by up to 20-30%, depending on the workload. This brought the performance of a Graviton3 vCPU (one Neoverse V1 core) to about the same level as a traditional x86 core. This let us consider Graviton for performance sensitive simulation work where previously a non-Graviton instance with SMT disabled would have been used.
Now generally available, the AWS Graviton4 brings performance to a new level. Using the Neoverse V2 core, Graviton4 again delivers an up to 20% improvement over the previous generation for our EDA workloads. The graph below shows the performance on each of these generations of Graviton on a key RTL simulator we use. Per vCPU, Graviton4 is the fastest processor available in the memory optimized (R instance family) of Amazon EC2 instances and delivers outstanding price/performance. For the first time, Arm has the ability to turn off x86-based instances for some performance-sensitive workloads and instead use Graviton4.
Graviton4 gives access to a whole new range of EDA applications that were previously not feasible to run on Arm, extending beyond basic enablement to become the optimal platform for all classes of EDA workload.
Figure 1. RTL simulation performance for AWS Graviton generations using Siemens Questa Advanced Simulator
Each generation of Graviton has delivered large performance and price/performance benefits compared with its predecessor. Per vCPU, Graviton4 is nearly twice as fast as Graviton2 but is available in instances with up to 3x more vCPUs than previous generations of Graviton-based instances, resulting in a massive increase in maximum capacity per instance. Despite this higher performance, the price per hour has increased by less than 10% compared to Graviton3. The combination of price and performance resulted in Graviton4 being by far the cheapest instance type in AWS when measured by throughput.
Figure 2. RTL simulation cost for AWS Graviton generations using Siemens Questa Advanced Simulator
Arm continues to expand the use of Graviton-based instances to power the design of the next generation of Arm Architecture-based processors, ensuring this tradition of performance enhancements is maintained into the future.
Graviton4 is now generally available, powering the Amazon EC2 R8g instances and offering forward compatibility from previous generations of Graviton-based instances. It’s easy to get started with this class leading platform and we encourage you to test the performance for yourself.
Ready to adopt AWS Graviton? Check out AWS Graviton expert curated resources: https://aws.amazon.com/ec2/graviton/resources
More Neoverse Blogs
Great !That's the good way we have to take. For my part it's allready done since 2018 :).Of course for me it's more easy due of my archi-size.