Earlier this year Cadence acquired Forte Design Systems, a pioneer of high-level synthesis (HLS) and provider of the Cynthesizer SystemC-based synthesis tool. Sean Dart, Forte CEO since 2006, is now senior group director for R&D at the Cadence System-Level Design Group. In this interview he discusses the history of HLS, its worldwide adoption, who's using HLS today and why, distinctive features of Cynthesizer, and the advantages of the Cadence acquisition of Forte.
Q: Sean, let's step back into history for a moment. Forte started in 1998 as C2 Design Automation, and began selling HLS tools soon after that. Was there demand for HLS at that time, or was it a missionary sale?
A: It was a missionary sale. You had to find people who had a vision. We had customers back in those days who certainly had a desire for [HLS]. We found a few of those people, and they bought early on and helped shape the product. But it certainly was not widespread.
When we first came into the market, the main competitors we saw were Synopsys and Get2Chip. There was also a small company named Dasys that came out of Carnegie-Mellon. They were doing VHDL and Verilog behavioral synthesis. We ended up buying Dasys and this formed the backbone of our engineering team in Pittsburgh.
Q: Your initial products were based on an open-source C class library. Then, in 2001, Forte began selling Cynthesizer, which was (and is) based on SystemC. What was the importance of the SystemC standard in the development and adoption of HLS?
A: I think it's critical. There are still people on the market using proprietary C subsets, and I think that's kind of narrow-minded. Customers expect to have multiple tools providing solutions, and they want to use tools from different vendors. The only way to have interoperability is to have some standard language behind it.
Also, from a sheer language point of view, it's really important to model things at the right level of abstraction. You need to do module-level or block-level decomposition in a systematic way. That makes your verification more effective. I think SystemC is definitely superior to non-standard C++-based class libraries.
Q: Many of your early customers were large Japanese consumer electronics companies. Why were they drawn to HLS?
A: We got into Asia very early and we started in Japan on the consumer and semiconductor side. Then we expanded into other parts of Asia and started to develop marketing in the U.S.
The Japanese customers were willing to discuss the value of behavioral synthesis technology to them. They wanted to jump ahead and they had some geographical pressures. For example, in the consumer space, they wanted to move some of the development to cheaper regions in Asia, but they wanted to keep their key proprietary algorithms in Japan.
I think in general there's a different corporate culture in Japan that perhaps allows companies to push strategic initiatives like [HLS] down into their organizations more efficiently.
Q: Who's using HLS today, and in what geographies?
A: In terms of geography, it's pretty much everywhere now. U.S., China, Korea, Japan, Europe, India, Israel—it's in major design locations around the world. HLS has really broadened out in terms of the application space as well.
At Forte, we preferred to go after larger companies because there is more bang for the buck. Once you actually get in, it is a lot cheaper to continue selling to an existing customer. There is a paradigm shift and you have to attract people to a different way of thinking.
Q: For what types of applications are companies using HLS?
A: When Forte came into the market, HLS was still very datapath oriented. After a few years we made a lot of changes so that Cynthesizer became much stronger in control-dominated designs. Today designs can be datapath, control, or any mix of the two. That has really changed the game. Customers have moved to such things as DMA controllers, video platforms, bus interfaces, network switches, and camera sensors.
Q: What do your users see as the primary benefit of HLS? Is it time to market, easier verification, or the ability to generate and compare micro-architectures?
A: All of these reasons are there.
There is always a reason that people buy in the beginning and there is a reason they keep buying, and these reasons may not be the same. Some people have been surprised by the results they've managed to achieve. We often get people who buy for improved time-to-market on the front side of the design cycle, and they get that, but they also realize that their IP reuse is so much higher than it is at RTL. IP reuse is a key reason that people keep coming back.
Q: Today there are several HLS tools on the market. What's distinctive about Cynthesizer?
A: It's a combination of things. Cynthesizer has great QoR [quality of results] and is very focused on the designer. There are very good IP libraries around the product and there are mechanisms for communicating between blocks. Finally, we try to educate people, and we do a lot of work on collateral, training, and documentation.
Q: How does Cynthesizer compare to the Cadence C-to-Silicon Compiler?
A: From a marketing point of view, it is just about heads-on. We practically could have written each other's slides. That's one of the reasons that joining the teams made so much sense.
There are, however, some different approaches on the technology side. At Forte, we had to integrate with a lot of tools from different vendors. We didn't have our own Verilog simulators. The Cadence group working on C-to-Silicon didn't need to do that—they could focus on just the synthesis tool.
C-to-Silicon has some very unique features. One is an incremental HDL synthesis capability for ECOs, which is a very compelling feature for a lot of customers. Design visualization is also a high-productivity value that C-to-Silicon offers. Cynthesizer, on the other hand, offers unique scheduling and datapath optimization technology, as well as some pipeline implementation details that are very beneficial.
Q: What do you see as the advantages of the Cadence Forte acquisition?
A: From a company point of view, Cadence and Forte spent a lot of time building the same thing and marketing the same thing. Not duplicating that effort, and being able to capture all the dollars and all the business, is definitely a more profitable path forward.
Of course, this really only works if you retain your customer base. We believe that is extremely achievable. Customers are completely engaged and we expect to maintain that. From Forte's point of view, it is great to be in a position where we have access to a significantly wider sales channel and resources within the company.
Q: What product plans does Cadence have following the Forte acquisition?
A: We are heading down the path towards a single product. Until we get there, and we can successfully transition customers to that product, we will continue to actively support the existing products.
Q: What is the future of high-level synthesis?
A: I think it's extremely rosy. The business is growing very well, we're seeing a lot of demand, and the results are very compelling. We regularly see designs implemented in HLS that have smaller area and lower power than hand-coded implementations.
Customers are starting to use HLS on very large designs. When we began, people were doing 50K gate blocks and 100K gate blocks, and we are now routinely seeing 10M-30M gate projects. People are willing to commit large design projects to the HLS flow. If HLS didn't deliver, these products would fail. There is no back-up plan. Now that we can get that kind of commitment, we believe HLS will grow very quickly and be very successful.
Related Blog Posts
How Cadence Acquisition of Forte Boosts High-Level Synthesis
Forte and Cadence at DAC: How to Deploy High-Level Synthesis
DAC 2014: High-Level Synthesis (HLS) Users Share Advantages, Challenges
Interesting, but I'm not sure I see what changed in the last five years www.eetimes.com/document.asp or since this case study 4 years ago: www.eetimes.com/document.asp High-Level Synthesis is still the same old technology, so what can you really do about so called "control-oriented" code? And you and I both know that SystemC may well be a standard, but the subset of what subset of SystemC is accepted for synthesis is still tool-dependent, unless of course you stick to "synthesizable SystemC", which is basically RTL written in SystemC...