The input-output ratio of chip design on the cloud can be increased by 100 times! Or it may be an opportunity for new entrants to surpass

Author | Bao Yonggang

In recent years, traditional industries are accelerating their migration to the cloud to promote digital transformation. Although chip design belongs to the traditional industry, the cloud has a long history of chip design. However, with the popularization of cloud computing and the improvement of hardware performance, designing chips in the cloud can increase the input-output ratio by several to hundreds of times, and therefore it is adopted by more and more chip design companies.

For emerging chip design companies, this may be an opportunity to surpass traditional large companies. Of course, mature chip design companies can also use cloud computing to better allocate resources.

Will mature companies that are used to traditional chip design methods be happy to accept chip design on the cloud? What changes will the chip design cloud bring to the chip industry?

A long time ago, Synopsys, one of the world's three largest EDA providers, has been actively exploring cloud-based EDA (Electronic Design Automation) tools to help chip companies design better chips.

Mr. Xie Zhonghui, Vice President of Synopsys China and General Manager of Chip Automation Business Unit, told Leifeng.com: "We started deploying EDA tool cloud projects many years ago, mainly cooperating with large chip companies to carry out internal cloud deployment. For example, we worked with TSMC to jointly deploy cloud design and chip manufacturing platforms, helping TSMC become the first foundry to achieve cloud design. We are very proud that the world's first chip fully implemented in the cloud was born on this platform. We also successfully implemented a new cloud timing signoff process on the Microsoft Azure platform with TSMC to accelerate the development of the next generation of system-on-chip."

Xie Zhonghui, Vice President of Synopsys China and General Manager of Chip Automation Division

Previously, the cloud technology ecosystem, customer acceptance, and hardware advancement were not enough to promote cloud-based chip design. However, in recent years, we have seen Amazon AWS, Microsoft Azure, Google Cloud, Alibaba Cloud, and Unisplendour Cloud all launch chip design clouds, and the acceptance of chip design clouds has increased accordingly.

"The biggest driving force behind migrating chip design to the cloud is the flexible optimization of resource allocation," Xie Zhonghui pointed out. "For small and medium-sized chip design companies, the number of large-scale projects each year is not large, and the cost of purchasing all the tools and computing resources for the entire chip design process is high. For large companies with strong financial resources and many projects, the simultaneous advancement of multiple projects or the temporary addition of projects will also require flexible computing power and additional design tools. Therefore, regardless of the size of the company, migrating to the cloud can meet the resource optimization needs of different chip design companies."

In particular, as chip design complexity continues to increase, it is increasingly difficult for a company's IT and design departments to rely on experience to effectively match computing power with tool demand calculations. Pre-deployment of computing resources will bring a huge cost burden, and insufficient computing power and tools will make it difficult to quickly meet sudden and fluctuating loads.

A very important reason for this is that the chip design cycle is very long and the computing power requirements of each process are different. The chip design process generally includes front-end design such as functional design, design description, and design verification, as well as back-end design such as synthesis, STA (static timing analysis), and PR (automatic placement and routing).

SoC design process, source: Alibaba Cloud Research Center & Synopsys

An industry insider told Leifeng.com that the front-end and back-end of chip design have different requirements for computing power. The front-end is single-threaded, highly concurrent, and mainly uses small files with intensive raw data, while the back-end design simulation is multi-threaded and uses large files. In addition, the more advanced the chip process, the higher the demand for computing power. The difference in computing power requirements between different process nodes can reach exponential levels.

Xie Zhonghui also said: "The entire chip design process generally takes 12 to 18 months, and the number and types of design tools used in each stage are different. Cloud-based chip design can help small and medium-sized chip companies make full use of limited funds and human resources, optimize tool and computing resource allocation, improve their own technological innovation, and catch up with relatively mature design companies."

It should be emphasized that the cloud-based chip design should focus more on the input-output ratio rather than the simple cost investment. Xie Zhonghui explained that, for example, the tools or computing resources that can be purchased for 1 million are very limited and may not be fully utilized. With the same 1 million invested in purchasing cloud resources, reasonable time-sharing optimization, the long-term accumulated computing resources and productivity obtained far exceed 1 million. Chip design companies should view chip design on the cloud from the perspective of input-output ratio. At the same time, using cloud resources and a large amount of design data and AI to optimize the design process can also reduce labor costs.

From the perspective of the chip design process, compared with other links in the design process, chip simulation verification is not only complex but also time-consuming. Data shows that the time spent on design verification of some chips is usually as high as 70% of the entire chip design cycle. Therefore, the high computing power and memory of cloud computing can greatly reduce the time of chip simulation verification.

"The infrastructure of many Synopsys tools can perform deep learning autonomously, and verification tools have also added AI capabilities. If you buy EDA tools traditionally, the amount of model data in the tool is fixed, and you need to rely on the experience of engineers to optimize and iterate to meet the project requirements of the chip."

Xie Zhonghui further stated, "On the cloud, EDA tools can learn based on existing data, and then optimize through more intelligent tools. The experience requirements for engineers to achieve chip design goals will be much lower, and efficiency will be improved accordingly. It may take 5-8 engineers two months to achieve the optimization point, but now only two or three engineers can achieve it in one or two weeks. I think this is one of the main driving forces for the cloudification of chip design and EDA tools in the future."

Just last month, Synopsys announced that its IC Validator physical verification solution running on Microsoft Azure completed the verification of an AMD Radeon Pro VII GPU (comprising more than 13 billion transistors) in less than nine hours.

So how many times can chip design be improved by cloud computing? Xie Zhonghui said: "It depends on the scale of the project and the degree of optimization. If the optimization is good, the input-output ratio can be improved by more than multiple times. If it is combined with the optimization of deep learning technology, the input-output ratio may be improved by more than several times, and it may even be dozens or hundreds of times."

He also said that for small and medium-sized companies, the same investment can get multiple times the input-output ratio, and the advantages of cloud computing cannot be ignored.

Even with all the attractions, it remains to be seen whether cloud computing can quickly popularize chip design. " There are other considerations for chip design companies to consider when deciding whether to go to the cloud. IP is the core asset of chip design companies, and its security is very important. In addition, there are differences in legal terms and intellectual property protection, and it is also necessary to consider whether third-party IP suppliers support cloud computing,"

Xie Zhonghui said, "But the value of cloud computing for chip design is also obvious. In addition to the many obvious values ​​mentioned above, such as flexible usage modes and better input-output ratio, there is also an invisible advantage that deserves attention - real-time response from expert support. In the cloud design environment, expert-level professional support and services are real-time and not subject to any geographical restrictions. This allows chip design companies to truly enjoy the necessary and timely professional support in their specific work without any worries."

IC R&D platform layered architecture, source: Alibaba Cloud Research Center

In addition, as a SaaS service, chip design cloud requires close collaboration among EDA tool providers, cloud service providers, chip manufacturers, chip design companies, etc. The security and capabilities of the chip design cloud are also closely related to the positioning of the cloud service provider.

But in any case, the greater computing power released by the chip design cloud can inspire engineers to generate more creativity, and combined with deep learning technology, it will bring orders of magnitude innovation.

In recent years, with the rise of AI technology, Internet companies with large amounts of data have entered the chip industry and designed dedicated AI chips with the help of the cloud, such as Google TPU. Xie Zhonghui said: "The chips designed by these companies are all related to data centers and AI. Compared with general chip companies, their advantage is that they have large amounts of data and algorithms. Since these chips are mainly used to meet internal needs, they can be vertically integrated according to the specific business, so they can be optimized to the extreme in specific scenarios."

As a result, these chip design companies that emerge with AI and cloud technology may bring certain changes to the chip industry's R&D model, technology trends, industrial chain, business model and even culture in the long run.

Xie Zhonghui believes that this will be a virtuous cycle. For those Internet and system companies that are determined to invest in chip design, they need to understand and accept that the chip investment cycle will be longer than software development. At the same time, they will bring new inspiration to traditional chip companies, allowing traditional chip giants to no longer be limited to chip performance and power consumption, but also need to closely integrate with user application scenarios and provide better service experience.

A typical example is that in the mobile phone industry, the self-developed SoCs of Apple, Huawei, and Samsung have all put pressure on third-party mobile phone SoC providers. To this end, if they want to provide chips that perform better than the self-developed chips of mobile phone manufacturers, they need to integrate more closely with user experience.

The next key question is, how many traditional chip design companies will be able to quickly accept the form of cloud design? What will be the competitive landscape between them and the new chip design companies born in the cloud era? Time will tell us the answer.