Making Trouble in Las Vegas

For years John Cooley has organized what is called the Cooley's DAC Troublemaker Panel. This year, the participants (troublemakers?) were:

• Naveed Sherwani, CEO of SiFive (a RISC-V implementation company)
• Joe Costello of Montana (creating a hardware accelerator for SystemVerilog and also merged with Metrics a year ago)
• Anirudh Devgan, President of Cadence (I'm just going to assume you who know they are)
• Joe Sawicki of Mentor (a Siemens business)
• Mo Faisal of Movellus (doing analog IP)
• Raik Brinkmann of OneSpin (a formal verification company based in Munich, which they proved by serving German beer on their booth from 10:00am each morning)

Note that there are two Joes on the panel, so I'll call them Costello and Sawicki to avoid confusion.

John Cooley was the moderator (perhaps he's the troublemaker). In fact, as Joe pointed out at one point in the panel:

It makes it interesting when the troll on the panel is the moderator.

One clever thing about the panel is that the questions are submitted anonymously by anyone on John's mailing list. So they don't have to beat around the bush, as you will see below.

John kicked things off with a softball question to Sawicki: "You acquired AI from Solido, can you put all those AI buzzwords into English?"

Joe admitted that they just put ML on the end of every product name and it solves all problems! But obviously there's been an explosion of funding from the VC space around chips implementing machine learning. Big ones for data center, and small ones for the edge. Machine learning is really useful algorithmically for large improvements in tool performance, set up, or anything pattern based. Solido is one example, OPC is another. You can get 10X or 100X increase in performance. In OPC, they model literally billions and billions of shapes. “I can’t even count how many projects we have internally using machine learning.”

John moved on to Anirudh. "JasperGold Smart Formal...are you just coding press releases?"

Anirudh said that the good thing about machine learning is that it is computational. Social media is not. Lots of matrix multiplication is the sweet spot of traditional EDA expertise. So we as an industry are well suited to doing machine learning. There's lots of collaboration with NVIDIA, CMU, and so on. Most of the complicated stuff is mapping your problem onto existing ML models. We have algorithms that are only good for certain things like Boolean analysis. But logic functional verification is a hard problem and is a beautiful area for machine learning. “Verification closure” can give you better feedback than today where you run for six months and still don’t know if you are finished or not.

John described this as "faith-based computing."

Anirudh said let's look at place and route. We already have multiple algorithms and usually, you just see if they improve the PPA. In case of Innovus with ML it improves PPA quite significantly.

John moved to Mo. "Why are you doing analog?"

Mo told us that the AI chip market could be $50B in five years. Movellus IP is in a lot of those chips. Edge AI is about minimizing the amount of data that has to be shipped to the cloud. We delivered a sub-microwatt PLL. At the other end of the scale there are 100s or 1000s of cores. A lot of process customization is going on and the digital algorithms aren't too good. People are exploring memory-based networks. Some of these chips will be 60% analog.

John moved to wonder why SiFive is in business, too, and asked why they have a chance against Arm.

Naveed said that, in three years, they already have 70% of all the cores that Arm has. In the last year and a half, they have 102 design wins. Of the top 10 semiconductor companies, six are engaged with SiFive. When they release a video about anything they are getting 3-6 million people watching it. In the entire history of the semiconductor industry, this has never happened. We are bringing new people in. Within the next three years, you'll see RISC-V-based cellphones. Unlike Arm, SiFive sells a reconfiguration tool so you can drop ports you don't need, switch to 32 bit. Some customers are getting 1/3 of the power (66% reduction). Naveed reckons there's a lot in any Arm core that is not needed for that particular application but there is no way to take it out.

Costello jumped in (since Montana is built on top of RISC-V). "RISC-V is going to become a huge force," he said, "and is only going to grow. We are taking techniques that have been successful in other industries, like open source, and applying it to our industry. Within a year, you will be an idiot for picking anything else. Maybe it’s already happened. Custom processors is the key thing for SoC design in the next decade. We're a small company using RISC-V and we can use it to do anything."

Mo said that it's not just the core, it’s the ecosystem, too.

Next, Cooley moved onto Montana: "Joe, we've been hearing about Montana for years and years."

Costello admitted that it was seven years. "We demonstrated an Altera version and now we're going with a customized RISC-V processor instead, which is a better solution. We get to make what we want and it was a no-brainer choice for RISC-V. We make our own core."

"Are you going to be another Rocketick when you're done?"

Joe said that Rocketick is only on Intel architecture, but you can only go so far. The reason Google has thousands of guys building TPUs is that you can’t get there with a general-purpose processor whether its Intel or NVIDIA. The only way to improve performance for the next 10 years is special-purpose processors.

Raik was next for Cooley's withering questions. "OneSpin is getting trashed by Cadence's JasperGold."

Raik denied this, since there are four areas where OneSpin doesn't overlap at all with JasperGold. "We are winning more than we are losing. Not just against JasperGold, all 3 of them. 50% growth rate every year over the last four years."

Sawicki apparently had issued an invitation to Anirudh to see AMD scale Calibre to 1024 CPUs.

Sawicki said that the cool thing is that cloud for EDA is running differently from other applications. It is focused initially on burst compute for physical and functional verification. People cannot build data centers big enough and fast enough. So now you can pull tougher 4,000 or 6,000 CPU clusters for 12 hours. "Now most of our customers are saying that they need to be able to scale out to the cloud. A giant chip scaling up to 4,096 cores for AMD."

Cooley: "Anirudh, do you regret doing Pegasus?"

Anirudh stated that it's a very important area, and a marathon not a sprint because the DRC market will change a lot in the next three years. "Pegasus is profitable. We have more than 100 people in R&D. More than enough to be a world-class tool. I think Pegasus will do well. It will get more and more profitable. We haven’t done this in the past, we want a complete solution. Pegasus is an important piece of that. The good thing about DRC is there’s only one player today. It’s hard for a startup to go after, but for us it’s a big area. We already make enough money in DRC and LVS, and we are in it for the long haul. But it's very different from competing with a startup. Last year you asked me the same question about Voltus, but you have to be patient, especially for signoff."

Sawicki chimed in: "You've got some engineers, may as well have them do something. EDA has changed a lot because of how hard process technology is at this point. Design kits are hard because processes are so hard."

Costello pointed out that there's not a lot of incentive for foundries to work with a new competitor if one came along.

Anirudh pointed out that DRC is not a numerical qualification like, say, Tempus. DRC is a 0-1 answer—did you give the right answer or not. It takes time, but DRC qualification is a lot simpler than static timing.

Cooley went back to Costello. "Joe, last year you announced you were merging Montana and Metrics. What advantage does Metrics have any more?"

"It was the best move we ever made," Costello said. "Metrics hasn’t really made the move to the cloud, we’re just starting. If you move to the cloud there are a lot of pieces. Business model, SAAS model. We’ve not done that in EDA. You buy what you want when you need it. One customer is a 2000 core company but most of year they use just 400. But during tapeout they really need 6000. So you’re never buying the right amount. Second thing, the public cloud is going to be much cheaper in the long run than their own data centers. On the technical side, it will be differently built and containerized from scratch."

Cadence announced Cadence Cloud last DAC so was an obvious target for Cooley's follow up: "Anirudh, can't you just knock out a simple business model change?"

"We do have a cloud offering," Anirudh said. "The first issue is acceptance moving to the cloud. In terms of running Xcelium in the cloud or whatever, you can do that. We have a regular flow, and we have the CloudBurst hybrid environment. It’s a delivery mechanism at the end of the day. By itself, it’s not a vector to disrupt."

Naveed agreed that delivering a tool through the cloud, and having a cloud business model are two different things. Selling 10 licenses to deliver n the cloud is not really a SaaS model.

Anirudh broke it down. "There are three parts to the cloud. One, having the tools available. Second, the business side. We have eDAcard and go down to one-week licenses, so, for now, we have granularity at one week. Third is the ecosystem: security issues, foundry support, and the like. That's where we’ve spent a lot of time in the last couple of years. Microsoft Office is $99 per year. So EDA isn’t that different. But whether it’s expensive or not expensive that’s a pricing issue, not a business model issue."

Sawicki agreed. "Our cost of sales as an industry is north of 20% and the idea this is going to be some massive cost saving is not going to work. Salesforce upgrades servers, but in EDA we can’t just change the software and hardware. There’s going to be a change with cloud, but it won’t be the Salesforce model."

Naveed didn't agree at all. "That's like telling me Uber doesn't exist. But owning a car when you only use it 10% of the time makes no sense."

Costello thinks that when you are in a high-compute-intensive thing, it should be by the minute. You can containerize, and do everything including the user interface in the cloud.

Anirudh emphasized that it's the customers' choice to do cloud versus not cloud, how much public versus private. "We have both Uber and cars that you own."

Cooley moved on. "Anirudh, you put $150M into Green Hills. Synopsys put $500M into Black Duck. Are you guys trying to get out of EDA?"

Anirudh said that customers want Cadence to move to the system level. "This is not a new trend but has picked up in the last few years. We want to see areas that are more synergistic with our core business. Two areas are computational software, and another is system analysis. We launched Clarity. Another big area is embedded software and especially embedded security. EDA and IP is our core business,  but as customers do more and more systems we want to move into those adjacencies."

He ran out the clock with a pitch for Clarity.

Clarity does three things. Meshing in Clarity is very good. Then the solver itself, it's new solver technology. Parallel technology. But even on same number of CPUs, Clairity gives a big speedup, up to 5X. If you look at finite element, it’s a well-known method dating back to the 1940s. You can model any 3D surface and the physics is well understood. But then you end up with a huge matrix and it runs very very slowly. We can solve it much more efficiently since we have so much computational expertise. 50B node matrices we can solve. But in system space, even 30 million is a big number.

With that, our time in the room was up. Tune in again next year.

Sign up to get the weekly Breakfast Bytes email: