Cadence Cerebrus in SaaS and Imagination Technologies Case Study

With the increasing demands of the latest applications such as 5G and AI, today’s design teams are adding more compute power and functionality to chips and moving to the cutting edge of process technology in a short time. In contradiction, today’s biggest team challenge is an industry-wide shortage of skilled design and implementation engineers. Current teams are overloaded, impacting their ability to market new products.

Fortunately, during the past few years, a key technology has become available to drive the next big leap forward in design productivity: Artificial Intelligence (AI) has made noteworthy progress and is now ready and available for electronic design automation. The Cadence Cerebrus Intelligent Chip Explorer utilizes AI—specifically, reinforcement machine learning (ML) technology—combined with the industry-leading Cadence digital full flow to deliver better power, performance, and area (PPA) more quickly.

However, this highly effective reinforcement ML employs many parallel and iterative training experiments using hundreds of CPUs and substantial amounts of memory to converge on the optimal solution. However, typical start-up and mid-size design houses do not have such machine resources in their on-prem setup.

Cadence Cerebrus in the Cadence-managed SaaS model offers customers nearly unlimited compute and software license capacity of the cloud to achieve PPA goals faster and accelerates time to tapeout.

The Cadence Cerebrus Intelligent Chip Explorer is a revolutionary, AI-driven, automated approach to chip design flow optimization. A typical design block optimization in Cadence Cerebrus employs between 10 and 20 parallel training scenarios in each iteration, requiring nearly 300 CPU cores. When optimizing multiple such design blocks in a complex SoC, the design team may need thousands of CPU cores and a substantial number of software licenses. Further, depending on the design block size, an optimal configuration may require 250GB or more RAM per machine and high-performance storage.

Building on the extensive experience of 300+ Cadence Cerebrus tapeouts and over 10 years of hosting cloud-based EDA design environments, Cadence has developed a Cerebrus-optimized, secure, and ready-to-use cloud setup. Additionally, this cloud solution offers Cadence Cerebrus in a cost-effective, pay-for-use SaaS business model aligned with the highly peaky compute and software license needs of Cadence Cerebrus.

Benefits of Cadence Cerebrus in SaaS

The Cadence Cerebrus in SaaS is built on a production-proven Cadence Cloud platform that has helped tape out over 100 production designs for customers worldwide. The platform provides a fully integrated environment for a full front-to-back design flow, a partial design flow, and a hybrid cloud-based peak capacity that augments an on-premises compute environment.

Traditional on-prem datacenters have many limitations—capacity restrictions, inconsistent turnaround times, and upfront capex investment—that hinder engineering productivity, project deadlines, cost-efficiency, and time to market. Companies wanting to set up their public cloud environments instead of an on-prem datacenter must hire unique IT, security, and EDA experts and expend months of effort to set up, secure, and manage an EDA-efficient public cloud environment.

Working in the Cadence-managed SaaS environment removes all such limitations. The Cadence-managed SaaS is already Cadence Cerebrus optimized, secure, and ready to use, and it gets the project up and running in a matter of days without any IT effort. Further, Cadence application engineers can easily partner with design teams and quickly debug issues in the cloud environment without the need to package up and provide a debug test case.

With the innovative SaaS business model, users can launch any number of Cadence Cerebrus scenarios in parallel, allowing Cadence Cerebrus reinforcement ML to train faster and converge on the optimal implementation quicker. The result is faster time to tapeout!

Cadence Cerebrus in a SaaS – Reference Architecture

The ready-to-use Cadence Cerebrus in SaaS is a comprehensive solution that comes built-in with

Every customer is provisioned a virtual private chamber (VPC), which only customer-assigned users can access. The cloud environment is ISO 27001 and 27017 security-certified and SOC2 compliant, fulfilling security requirements for IPs and PDKs from foundries. Figure 1 shows a reference architecture for the SaaS cloud environment.

Using nothing more than a standard web browser, users connect to the secure Cadence Cerebrus environment, and start the project by uploading and/or synchronizing design files, scripts, libraries, and PDKs. Once the necessary data is in place, the users launch and monitor jobs via Cerebrus GUI or batch mode, analyze results, and download the final implementation, including the trained Cerebrus AI model.

For Imagination Technologies, Cadence had set up the Cerebrus in SaaS environment on Amazon Web Services (AWS) cloud infrastructure. The ready-to-use SaaS environment in AWS used high-performance Amazon EBS storage and R5d instances of Amazon EC2 for Cadence Cerebrus jobs.

Imagination Technologies provides world-class CPU, GPU, and AI intellectual property (IP). As part of this IP development, Imagination creates Design Optimization Kits that are reference flows for designers to implement the IP using various foundry process nodes. Imagination wanted to investigate if Cadence Cerebrus could be used to improve the PPA of these Design Optimization Kits, and the Cadence Cerebrus in SaaS environment was an ideal way for Imagination to quickly utilize Cadence Cerebrus without any on-premises IT setup.

Imagination selected three 5nm blocks from the C-Series GPU IP, ranging in size from 500K to 4.3M cells. Cadence Cerebrus used the current Design Optimization Kit reference flow as a starting point and optimized the blocks with timing closure as the primary objective and power and area reduction as secondary goals. Cadence Cerebrus used 20 distributed computers per block, each with 16 CPUs, so the peak CPU usage was 3x20x16= 960 CPUs.

Cadence Cerebrus improved PPA for all the GPU blocks, as Table 1 shows.

As part of the Cadence Cerebrus AI-driven design optimization, a machine learning model is generated that can transfer training data between Cerebrus runs, significantly reducing run-time for further optimization. These pre-trained Cerebrus AI models are available to Imagination and Cadence customers as a warm start for their Imagination GPU design projects.

The Imagination technical support teams are located in many worldwide geographies. Using the Cadence Cerebrus SaaS environment, any of the technical support engineers could access the Cerebrus runs, regardless of time zone or physical location, enabling efficient team collaboration.

The AI-driven Cadence Cerebrus automatically delivers a productivity boost and significant PPA benefits to today’s overstretched chip design teams. In addition, the software-as-a-service (SaaS) model provides cost-efficient and rapid access to nearly unlimited parallel compute and software license capacity delivering faster turnaround time.

Today’s chip teams looking for quick-ramp to design productivity, higher PPA, and faster time-to-market, without the burden of significant Capex investment as well as costs and delays of setting up and managing a secure IT environment—Cerebrus in Cadence-managed SaaS is the ideal solution.

Learn more about Cadence Cerebrus and the Cadence Cloud platform.