At the start of the month, Amazon Web Services (AWS) announced the launch of their new Arm-based AWS Graviton2 processor, and associated 6th generation Amazon EC2 M6g, C6g & R6g instances. These chips have four times the number of compute cores with caches that are twice as large as AWS' first Graviton processors built on an earlier generation of Arm cores. According to AWS, they also have five times faster memory and 7 times the performance. With Amazon EC2 M6g instances, customers can optimize for both higher performance and lower cost per vCPU. M6g instances deliver up to 40% better price performance over current generation M5 instances.
Silicon design is an expensive business and getting a modern chip manufactured costs a lot of money. Arm must be sure that our designs are correct before delivering them to our partners. To achieve the necessary level of quality, our engineers depend on tools that simulate and exercise designs across a broad range of tests throughout the entire development process. The quantity of tests is vast, and results in an enormous amount of computer time being spent on running these simulations.
The first Arm design was simulated on a BBC Microcomputer made by Acorn, the company that would go on to become the ‘A’ in Arm. Over time, and tracking with ever more complex digital designs, simulations grew in complexity and required more powerful hardware than could be offered by Acorn or Arm based machines. Over the past few years, the capability of Arm-based servers has reached the point at which they aren’t just viable, they are compelling for running silicon validation tools.
Figure 1- Cadence tools running on an Arm platform during the Cortex A53 simulation effort.
Over the past few years we have worked closely with partners at leading Electronic Design Automation (EDA) software companies such as Cadence and Mentor to make this a reality. As a strategic customer of EDA tools, Arm helped drive the porting of key simulation packages including Cadence’s Xcelium and Mentor’s QuestaSim so they run well on Arm Architecture based servers. Our existing, mature toolchains made this a straightforward exercise, initially targeting Arm platforms using Marvell's ThunderX2 processor hosted in our own datacenters.
This has been a labour of love for us. There is a pride in being able to build your own tools but it has not only been about the strategic case for Arm to use Arm. We have also focused on the wider benefits available when using Arm-based hardware. For example, AWS Graviton2 processors deliver a major leap in performance and capabilities over first-generation AWS Graviton processors. Custom silicon design allows AWS to innovate on behalf of their customers and provide additional choices to optimize for performance and cost specific to our workload needs. For the Marvell ThunderX2 based servers we host ourselves, we achieve higher density in the datacenter, lower power consumption, and cheaper hardware compared to our existing engineering x86 datacenter platform. Thanks to Arm server standardisation, the exact same tool binaries run unmodified across a variety of Arm-based server platforms from different vendors including AWS and Marvell, both on premises and in the cloud.
Every month, Arm uses millions of hours of CPU time to simulate our designs, but the demand is highly variable depending on the phase of the project. This is a perfect use-case for Amazon EC2 where, as demand increases, we can start more instances and as demand lowers can stop them. Because of this, like many other companies, we are looking to take advantage of the flexibility that cloud service providers offer. With the introduction of the first AWS Graviton processor last year, we were able to begin investigating EDA in the cloud using Amazon EC2 A1 instances. We proved that our workflows could run on Arm-based instances, and identified some that fit the A1 instances very well. However, the available memory of the largest A1 instances was too small for other workloads, and for those we still required x86 instances, or could only use our on-premises Arm-based hardware.
The launch of the AWS Graviton2 and Amazon EC2 M6g instances changes the landscape considerably. With 4GB of RAM per vCPU, and up to 256GB RAM in total, the M6g instance type can now support many more of our workflows. The forthcoming R6g instance type with up to 512GB RAM will allow us to address even more workflows. The performance of AWS Graviton2 processors is a significant step up too, not only versus the original Graviton but also in comparison to those in the x86-based Amazon EC2 M5 instances which Arm had been using previously.
Figure 2- Total Runtime comparison between Amazon EC2 M5 and new M6g instances.
To test the M6g instance type we took real world validation jobs used to test the Cortex-A53 and executed them on an M5.24xl and an M6g.16xl, comparing the total CPU time taken. The M6g run time was only 2% slower, but the M6g instance had 33% fewer cores.
Figure 3- Performance comparison of Amazon EC2 M5 and new M6g instances.
For these EDA applications, the M6g instances provide 50% higher performance per vCPU, which is how Amazon EC2 capacity is priced and consumed. Furthermore, EDA software tools are frequently licensed by the number of processors used, so getting the same performance with fewer vCPUs reduces software licensing expense. Arm has been glad to partner with software vendors like Cadence and Mentor in porting and proving out the performance of these applications on the Arm Architecture. We expect our bottom line to benefit from the use of these tools on Arm based infrastructure as it reduces our engineering costs.
With new high-performance platforms from multiple partners being announced, this is an exciting time to be running on Arm-based platforms. It demonstrates the growth of the Arm ecosystem and all helps to make sure that the next Arm-based gadget you buy will work as intended.
Check out the new Amazon EC2 M6g instances here and sign-up for the preview to give your favourite software a try.
Find out more about the Amazon EC2 M6g