Introduction to semiconductor design workflows on AWS
Welcome to our first Semiconductor and Electronics blog post! This post, and additional posts that are coming soon, will help AWS customers and partners stay up to date with the activities of the AWS Semiconductor and Electronics industry team, as well as give prescriptive guidance for running your semiconductor and electronics workflows on AWS. We are a globally distributed team, dedicated to helping customers around the world accelerate their critical semiconductor workflows using AWS, from front-end design and verification, to back-end fabrication, packaging, and assembly. Our team is comprised of industry leaders, each with decades of experience in the semiconductor and electronics industry. We leverage this experience to guide customers from their initial AWS introduction, to running their first production workloads.
What should you expect to see in our Semiconductor and Electronics blog posts?
- Workload enablement using hands-on technical guides and AWS Reference Architectures
- Industry events that we will be attending, to include SEMICON, Design Automation Conference, Synopsys SNUG, and Cadence Live (CDNLive)
- Workshop information that will help you quickly deploy your environment on AWS
- Service and feature announcements that enable semiconductor workloads to scale
In this first post, we give an overview of running semiconductor design workflows on AWS. Using an AWS Reference Architecture, we explain a data migration path and introduce a few AWS services. For the purposes of this post, it is assumed that you are choosing to run a proof of concept workload prior to running production flows. Below is the reference architecture diagram that we will be using.
When you are new to AWS, we understand that translating what you are running in your on-premises data center to AWS services can feel a bit overwhelming. In the above diagram, we provide a hybrid on-premises and AWS architecture to illustrate how your current environment could map to AWS. We will walk you through the architecture, highlighting the services used and how the capabilities of the AWS Cloud provide a powerful, scalable platform for semiconductor and electronics design and manufacturing workflows. For the semiconductor use-cases, these workflows might include:
- Electronic Design Automation (EDA) including simulation, verification, and signoff workloads
- Computational Lithography including Optical Proximity Correction (OPC)
- Computer-Aided Engineering including electromagnetics, thermal, materials, and multiphysics
- Machine learning training and analytics, for example Yield and Failure analysis
- Supply chain and third-party collaboration, for example debugging problems or collaborating with a 3rd-party IP provider
- Software/firmware regression testing
Reference Architecture walk through
For each of the numbered items on the diagram above, we will provide guidance and highlight AWS services that customers in the semiconductor and electronics industry are using.
1: Determine which tool and what data will be needed for the proof of concept or test
Prior to running your proof of concept on AWS, you must decide which tool you would like to run and what data will be needed to run a successful POC. This first step has the potential of being a daunting task if you choose a tool or workload that has too many dependencies, or requires design data that is not easily decoupled from the entire workflow. In an effort to reduce dependencies, we recommend starting with a tool that has cloud-enabled licensing, and can use a standalone (or nearly) dataset. We frequently work with customers to help unwind decades of legacy dependencies, and often find that starting with either a new project or very small design is most effective for a POC.
Alternatively, AWS Semiconductor Solutions Architect team can host a one day workshop that launches an entire semiconductor design environment on AWS. This does not require the use of proprietary customer or 3rd-party data, but rather gives an example of how your flows can run on AWS. We’ll provide more details on this later.
2: Transfer data in to AWS using AWS Snowball, AWS Direct Connect, or another AWS service
Once you have determined what data you will need for your POC, you will need to transfer that data into AWS. The method of transferring will depend on how much data you will be moving. If you have a relatively small amount of static data and you have a fast, reliable internet connection, then you may be able to use your internet connection in to AWS. Another option and possibly looking beyond the initial POC, if you plan on frequently moving data both in and out of AWS, you should consider AWS Direct Connect. Using AWS Direct Connect, you can establish private connectivity between AWS and your data center, office, or colocation environment, which in many cases can reduce your network costs, increase bandwidth throughput, and provide a more consistent network experience than Internet-based connections. If you have large amounts of library, design, or simulation data that requires an initial one-time transfer, you should consider using AWS Snowball. AWS Snowball Edge supports up to 100TB and has a rich feature set that provides edge services and clustering abilities. For additional information, please see the When to use Snowball section of the AWS Snowball FAQ.
3: Data transferred in to AWS should initially be in an S3 bucket
When moving your data in to AWS, the best place to initially store data is an Amazon S3 bucket. Semiconductor flows require POSIX-compliant file systems, and EDA tools do not currently support object storage natively. That said, storing your data in S3 provides several features, including “11 nines” of reliability, 25Gpbs transfer to EC2 instances (per instance), cross-region replication, and data tiering (see Amazon S3 Features and Amazon S3 FAQs for more info). We’ll discuss this more in the Storage (#7) section, but there are AWS services that provide a POSIX-compliant file system linked directly to an S3 bucket. Alternatively you can deploy nearly any file system on AWS, and quickly transfer your data from S3 to the file system mounted on an EC2 instance.
Having the data in S3 also enables agility and fast failure. The high bandwidth connection to S3 allows for the fast transfer of data both to and from an EC2 instance. So rather than trying to preserve data or an entire NFS file system when a failure happens, just rebuild the entire environment in minutes and copy the data to the new environment.
4: User access through a remote desktop or command line using SSH
Now that your tools and data are on AWS, running a job is going to be very similar to how you would run in your own data center. Users access their environment through a remote desktop or command line login (SSH), just the same as they would be accessing resources in an on-premises data center. Engineering and physical design teams can use AWS NICE DCV, which is a secure remote desktop solution over varying network conditions. NICE DCV is provided at no additional cost, you just pay for the underlying infrastructure.
5: All of the infrastructure needed for semiconductor design workflows is available on AWS
Running semiconductor design tools on AWS will require similar specifications to what is being used in your on-premises data center. AWS has many services that provide the supporting infrastructure for the entire workflow. For example, license management is extremely important for semiconductor design flows. A License Manager can be deployed in minutes on AWS, either using licenses moved from on-premises or additional licenses purchased from your vendor. For job scheduling and orchestration you can deploy the solution you’re currently running, leverage a cloud native approach using only AWS services, or you can optionally build out a custom orchestration workflow using the flexibility of AWS. User management can be accomplished in multiple ways, for example leveraging our native services or moving on-premises user management solutions for legacy support.
6: AWS provides comprehensive options for compute
Before choosing your EC2 instances, you should determine the total number of physical cores and memory footprint your POC requires. AWS has over 200 EC2 instance types (Amazon EC2 Instance Types), and customers have successfully run semiconductor and electronics design workflows on many EC2 instances. We see many customers using our z1d, R5, C5, M5 and X1 instance types. The z1d instance family provides a 4-GHz sustained processor with a memory to core ratio of 16GB/physical core. The z1d allows customers to cost optimize by running jobs faster, resulting in increased license utilization. The R5 instance family provides a total memory footprint of up to 768 GB, and a memory to core ratio of 16GB/physical core. For jobs that require even more memory, for example physical design and timing closure, we have the X1 instance family and the X1e instance family. The X1e instances have up to nearly 4 TB of memory. Additionally, for compute intensive workloads that don’t require as much memory per core (e.g. front-end design), our customers use the C5 instance and the M5 instance families.
After you have chosen your instance types, we recommend limiting the size of the POC to test for functionality only. We have many flexible pricing options, for example Amazon EC2 Spot Instances, but rather than performing a large-scale test, we recommend running a successful POC and then determine which pricing options fit within your budget.
7: Storage options on AWS allow workflows to run without modification
As we briefly mentioned before, AWS provides several options for running jobs on a POSIX-compliant file system. To name a few, you can choose from Amazon EFS, Amazon FSx for Lustre, and you can also build your own file system on an EC2 instance. Depending on the I/O profile of your workflow, you should build a solution that has similar characteristics as the file system you are currently using. For example, you should determine if your tool requires a file system that has high IOPS performance, or high sustained bandwidth for large sequential reads, or if it is used just for temporary (ephemeral) data. Going into detail for each of the file system solutions is beyond this blog post, but we will cover this in a future blog post and you can also work with your AWS Solutions Architect to determine the best solution for your requirements.
8: Leverage additional AWS services
Once your data is stored in AWS, and more specifically in Amazon S3, you can leverage many of the additional services that the AWS Cloud provides. Many semiconductor design customers are building Data Lakes on S3 and using data analytics for log processing, license utilization, and job scheduling optimization. You can also take advantage of data tiering, with the S3 Intelligent-Tiering storage class. This storage class is designed to optimize costs by automatically moving data to the most cost-effective access tier, without performance impact or operational overhead. If you have teams located in other geos, use another Amazon S3 feature, cross-region replication (CRR). This allows you to automatically and asynchronously replicate data to a different bucket in another AWS Region. With your data in Amazon S3, you can start to explore and extend the capabilities of your current environment by leveraging the many AWS services.
9: Isolating environments leads to increased security
In the semiconductor and electronics industry, most designs require the use of a third party vendor or IP supplier. In traditional on-premises environments, access to the entire corporate network may be required to enable the necessary collaboration. Although safety and security measures can be put in place, the third party is still on the same network. On AWS, you can define a separate environment that only the parties involved with that project have access to. That is, you no longer have to open your network and then add restrictions. Instead, you setup an entirely new isolated environment on a project by project basis. Additionally, any communication with your fabrication facility can be done over a secure, optimized network connection. This includes not only transferring a GDSII file, but also yield analysis data that can be transferred back to the semiconductor design company to improve productivity.
10: Raise your security posture with AWS infrastructure and services
Using AWS, you will gain the control and confidence you need to securely run your business with the most flexible and secure cloud computing environment available today. As an AWS customer, you will benefit from AWS data centers and a network architected to protect your information, identities, applications, and devices. With AWS, you can improve your ability to meet core security and compliance requirements, such as data locality, protection, and confidentiality with our comprehensive services and features.
AWS allows you to automate manual security tasks so you can shift your focus to scaling and innovating your business. Plus you pay only for the services that you use. All customers benefit from AWS being the only commercial cloud that has had its service offerings and associated supply chain vetted and accepted as secure enough for top-secret workloads.
AWS has customers from all over the world designing everything from small ASICs to large SOCs with tens of billions of transistors, at the most advanced process geometries. We have only highlighted a few of the capabilities in this blog post; the AWS Semiconductor and Electronics industry team will publish more detailed posts expanding on these topics, and adding different subjects based on ongoing customer feedback. Please check back soon.
We look forward to helping you innovate faster!
Mark and David