Breakfast Bytes Blogs

At CadenceLIVE Silicon Valley, Ahmed Elzeftawi of AWS and Dibyendu Goswami of Cadence presented Pegasus TrueCloud for Gigascale Physical Verification using Hybrid Cloud on AWS. Amed is a Senior Partner Solutions Architect for Semiconductors and EDA, and Dibyendu is a Product Engineering Architect for Physical Verification. Actually, Dibyendu couldn't make the event and so Saransh Jain stepped in to present for Dibyendu.

Ahmed started off by pointing out that Amazon is part of the semiconductor and electronics industry. They design their own datacenter infrastructure chips (see my posts HOT CHIPS: The AWS Nitro Project and Climbing Annapurna to the Clouds). They also build their own chips for consumer devices like Kindle and Alexa, for robotics in their warehouses, AI, Amazon Go, and more. As Ahmed put it:

We benefit from AWS in our own IC development. AWS Nitro System, AWS Inferentia, AWS Trainium, AWS Graviton2 and Graviton3...100% developed in AWS Cloud for AWS Cloud.

Ahmed also covered the different types of instances available. That's too much detail for this blog post, although some of the information was covered in my post EDA in the Cloud: Astera Labs, AWS, Arm, and Cadence Report. To dive deeper, AWS has white papers, blogs, articles, videos, webinars, and more on their resources site.

Cadence Pegasus TrueCloud

Saransh took over from Ahmed to talk about Pegasus.

Physical verification has several natural aspects to it that make it especially suitable for taking advantage of large numbers of CPUs, such as on AWS. Different rules are largely independent. Parts of the design that are far apart do not interact. On the other hand, the input is physical layout and, at least at the full-chip level, is huge.

The traditional way of running DRC on the cloud is:

• FTP designs and schematics to the cloud
• FTP foundry rule decks to the cloud
• FTP setup files
• Run the job

The Pegasus TrueCloud way is to compute operations on the cloud while all the design and the foundry data stay on-prem. The Pegasus build is FTPed to the cloud and can start running. The advantage of keeping data on-prem is that there is a single source for everything and no possibility that what is in the cloud gets "out of sync" with what is on-prem. Licenses are checked out from the on-prem license server rather than needing to set up a special server in the cloud. As detailed later in this blog post, even the number of CPUs used can be varied dynamically to optimize turnaround time versus cost.

There are two execution models:

• All cloud: All the resources are from the cloud, and 16 CPUs are used on-prem
• Hybrid cloud: Available resources on-prem can be used with the remainder on the cloud (which can be more cost-effective if machines are available).

In a little more detail (also in the diagram above):

• On-prem: 16 CPUs to read the design layout, schematic, and foundry decks. Part of the pre-preprocessing is done as the design is being read.
• Cloud: 1 CPU to launch LSF jobs, other CPUs are acquired through LSF for the computation, with no limit.

Pegasus Console

The Pegasus Console allows for monitoring the status of distributed jobs and works across all of Linux, Window, Mac, iOS, and Android. It lets resources be added or dropped and can execute interactive Pegasus commands.

Pegasus FlexCompute

One challenge with running big EDA jobs in the cloud is deciding how many CPUs should be included to achieve the best turnaround time for big signoff jobs. This can be a challenge both with on-prem datacenters (where the CPUs need to be available) and in the cloud where CPUs are (almost) always available but payment is by the hour. Too few CPUs result in greater runtimes, but too many results in worse CPU utilization and higher cost.

Pegasus FleCompute provides dynamic CPU management to avoid idle CPUs, so there is no need for the user to set any CPUs. CPUs reach maximum utilization automatically and are dropped if and when no longer required.

Summary

• AWS has opened up the door for designers to design and verify their designs on the cloud by providing instant access to thousands of CPUs
• Latest EC2 instances like X2iezn that have Cascade Lake processor with up to 4.5 GHz all-core turbo performance and up to 1.5 TiB memory, speeds time to tape-out
• Pegasus TrueCloud lets users launch physical verification jobs on the cloud without having to copy any design or foundry IP to the Cloud
• Pegasus Console offers a unique way of monitoring distributed Pegasus jobs on the Cloud; it lets designers update, analyze and debug Pegasus jobs
• Pegasus FlexCompute maximizes CPU Utilization and achieves fastest TAT. No need for the designer to guess how many CPUs are needed for the best run-time

Sign up for Sunday Brunch, the weekly Breakfast Bytes email