It has been about a year since we last wrote about our efforts to move Arm’s production EDA environment to the cloud, and specifically to Arm-based cloud instances.
In that time there were several major announcements related to Arm’s presence in the cloud. At Re:Invent 2021 Amazon announced the AWS Graviton3 CPU and the C7g instance family based on it. We have been evaluating C7g instances for our EDA workloads and share some results later in this blog.
The ecosystem of EDA tools available on Arm continues to grow. In the past year we have begun to move more workloads to AWS Graviton2-based instances, most recently deploying Formal Verification into AWS using Cadence’s JasperGold. And as SaaS providers have added support, we have adopted new Graviton-based instances like the R6gd. On our Databricks workload, R6gd has allowed us to move to Graviton and reduce our costs by 15%. While Databricks is a broadly applicable data analysis platform, it is essential to several of our engineering flows. It allows our engineers to interpret the outputs of their EDA simulation runs at scale.
The various different stages of silicon design all require software tools to enable engineers to deliver their projects. These different stages require very different tools, although they all come under the umbrella term of Electronic Design Automation, or EDA. Cadence provides Arm with a broad suite of EDA tools that allow us to simulate and formally verify RTL. These are products such as Xcelium and JasperGold, as well as circuit simulation (commonly known as SPICE) with Spectre and physical characterization with Liberate. At Arm, we use all these tools on Graviton powered instances from AWS and in our on-prem Arm and x86 clusters.
For this blog, we would like to focus on the performance and price improvements of the new AWS Graviton3 CPU and the C7g instance family for several of our key EDA workloads. To show the benefits of C7g, we compare it to the AWS Graviton2-based C6g instance. Why not x86?
While we do still use x86-based instances, we find the AWS Graviton family offers better performance and lower cost than x86-based instances for all the EDA workflows we have tested. Moreover, the pace of performance and cost improvements available from the Graviton family of instances is far outpacing what we see from x86. You can see this result in figure 1.
Figure 1 shows the vCPU hours and cost, based on list price, needed to complete an EDA workload. Within the Graviton family, moving from A1 to C6g resulted in a 52% performance speed-up. And moving from C6g to C7g resulted in a 21% performance speed-up. Contrast that to x86, where moving from C4 to C5 to C6i has produced very little performance improvement.
The cost (based on list price) needed to complete the same EDA workload shows a similar trend. The cost of Graviton has decreased noticeably with each new generation while the cost of x86 has improved, but very slightly. Comparing the latest generations of C7g and C6i, you can see the Arm-based AWS Graviton3 offers a 50% cost advantage over x86.
Figure 1. vCPU hours and cost to complete EDA workload
Looking at these same two metrics, performance and cost, for some of our front-end EDA workloads tells a similar story. Note: in each of these comparisons we used C6g.16xlarge instances to represent AWS Graviton2 and C7g.16xlarge instances to represent Graviton3.
On Cadence Xcelium Graviton3 improves runtime performance by 22% over Graviton2. And the total job cost improves by 12%. On Cadence JasperGold, Graviton3 runtime performance improves by 30% with a corresponding cost improvement of 18% over Graviton2.
Figure 2. Cadence Xcelium and Cadence JasperGold runtime and cost comparison between Graviton2 (C6g) and Graviton3 (C7g)
It is worth noting, in the past we have shown EDA workload results based on the M6g instance type. And for many of our EDA workloads we prefer the larger memory size to core ratio of the M6g vs. the C6g. We test with the C6g here to create a more apples-to-apples comparison with the C7g, as this instance is the only version of Graviton3 currently available. We expect a similar improvement comparing ‘M’ instances as we see here comparing ‘C’ instances.
We see an even greater improvement from AWS Graviton3 on our back-end EDA workloads, which use floating point mathematics heavily. On Cadence Spectre the runtime performance of Graviton3 improves by 35% over Graviton2, with a corresponding 22% cost improvement. And on Cadence Liberate Graviton3 performance improved by 33% and cost by 21% over Graviton2.
Figure 3. Cadence Spectre and Cadence Liberate runtime and cost comparison between Graviton2 (C6g) and Graviton3 (C7g)
If you are interested in hearing more about our results on using AWS Graviton for our EDA workloads, please come join us at AWS Re:Invent 2022. Arm is presenting two breakout sessions focused on moving our EDA workloads to the AWS cloud and the resulting benefits we are seeing.
The first session, CMP204: Build a cost-, energy-, and resource-efficient compute environment is part of the sustainability and compute tracks. Troy Gasaway, Vice President of Infrastructure & Engineering at Arm, will present on the business and sustainability benefits Arm has seen by moving to the cloud. These benefits include hardware selection and scalability efficiency. As well as the benefit unlimited cloud capacity brings to our project scheduling and throughput. And Troy will cover the cost benefits Arm has been able to realize through use of spot instances and AWS Graviton processors. He will finish with an overview of how we enabled a hybrid on-prem/cloud environment with some best practices and lessons learned.
Session two, CMP320: Accelerating semiconductor design, simulation, and verification is part of the compute track. Mark Galbraith, VP of Productivity Engineering at Arm, will deliver an update to the progress Arm has made since last year in moving our EDA workflows into the AWS cloud. Mark’s presentation will go deeper into some of the results presented in this blog. And he will discuss some of the supporting tools, like Databricks, IBM LSF, and FxX ONTAP that Arm uses in our hybrid environment.
In addition, Ed Miller, Senior Principal Engineer at Arm and AWS Community Builder will be presenting COM310: Conserving wildlife with serverless, ML, and citizen science as part of the AI & ML track. Ed’s talk will look at the BearID Project and how it is using machine learning to identify individual bears for their study and conservation.
[CTAToken URL = "https://www.arm.com/partners/aws" target="_blank" text="Explore AWS Graviton" class ="green"]