The Skies Over EDA Are Finally Cloudy

EDA companies have been talking for years about providing access to their tools in the cloud, including more articles than I can count with titles about the EDA forecast being cloudy, clouds on the horizon, and so forth. The title of this post continues the dubious tradition of cloud-based puns, but there’s no future tense involved. Recent announcements from several EDA companies make it appear that cloud support is finally here and poised to become a mainstream vehicle for design and verification.

What took the industry so long? That’s a topic I first explored in a blog post almost exactly five years ago. At that time, there was quite a bit of buzz about EDA in the cloud with almost no real business. Synopsys co-CEO Aart de Geus was quoted at an EDAC meeting as saying that “Synopsys had made $0 on it.” Periodic articles addressed the question of when (or if) EDA vendors and users would embrace cloud computing. Many cited IP protection concerns as the biggest hurdle to overcome.

It’s quite understandable why users would not want competitors to get their hands on RTL code, testbenches, assertions, etc. In theory, a design could be copied to produce clone chips, possibly tanking sales opportunities in the targeted markets. Of course, design and verification IP is precious and must be protected; there’s no argument about that. Is this form of IP more precious than customer lists, financial data, personnel records, and the many other types of critical data that many companies already store in the cloud?

To put it another way, would a CEO rather have the complete contents of a CRM system or the RTL code for the next product compromised? I pondered this question five years ago and I’m still not sure of the answer. This topic, among others, arose at a roundtable discussion held earlier this year and published in three parts by Semiconductor Engineering. There was a clear sense that users have become more comfortable with cloud computing and that EDA vendors have a better understanding of what tools and IP protection techniques work best in a cloud model.

Actually, OneSpin figured this out five years ago. In fact, its announcement of cloud support for formal verification was one of the spurs for me to write my original blog post. Two years later, I asked OneSpin to provide an update on the cloud initiative. This response made two important points that still hold true today. The first is that distributed verification processes are a good fit for clouds. It is easy to allocate a pool of virtual servers, spread jobs across them, and scale up or down as needed.

Formal techniques are a natural fit for distributed processes –– for example, analyzing each assertion on a separate server or running multiple proof engines on the same assertion in parallel. Historically, simulation ran in server farms or clouds by running a different test on each machine. However, in recent years, significant progress has been made in truly parallel simulation so that even a single test can be run faster. Many other EDA tools, including logic synthesis and place-and-route, can also benefit from running in the cloud.

The second point was that OneSpin addresses IP security by uploading to the cloud only an encrypted formal model that could not possibly be reverse-engineered to recover the RTL design. OneSpin goes further than simply running jobs on remote servers with encrypted models. Only the compute-intensive, distributed portion of the process is sent to the farm, with a master process running locally and collecting results. This provides further levels of security, since even if the results could be intercepted they would make no sense without the local process.

A final point, and perhaps an obvious one, is that batch processes such as formal verification or simulation regressions make the most sense in the cloud. Interactive tools are generally run on local machines to maximize response time. With automatic scaling of capacity up and down on a per user basis, typical batch verification jobs have the potential of becoming effectively interactive. If you can run 10,000 jobs for only three minutes and get a result, instead of running 30 jobs for 1,000 minutes, this is a huge difference in engineering productivity.

Given the guidelines of parallelism, security and batch operation, it is clear that computing can work for EDA and is being used increasingly. The clouds have moved from the forecast to horizon to right overhead. I am hesitant to make any sweeping predictions given the complicated history, but from the recent news, it does seem that the cloud has arrived and that its time has come. We will be watching as the industry continues to evolve. As always, your thoughts and comments are welcome.