Experts at the table, part 3: Resources get shared across groups, but some want their own; stats show high utilization rates; performance wins every time over energy costs.
Semiconductor Engineering sat down to discuss the changing emulation landscape with Jim Kenney, director of marketing for emulation at Mentor Graphics; Tom Borgstrom, director of the verification group at Synopsys; Frank Schirrmeister, group director of product marketing for the System Development Suite at Cadence; Gary Smith, chief analyst at Gary Smith EDA; and Lauro Rizzatti, a verification expert. What follows are excerpts of that conversation.
SE: Are companies buying emulators for use in emulation farms the same way they did in the past for simulation?
Kenney: In terms of a resource shared by multiple project, absolutely.
Schirrmeister: They’re what we call professionally, ‘a private cloud.’
Kenney: You’ll have an organization within a company that has the most emulators, and they would like everyone to give them dollars to buy these emulators, but satellite groups say they want their own emulators. So there is certainly central control of these emulators, with small groups wanting control of their own emulators so they don’t have to share it with a bunch of other project groups. We see 10 and 20 emulators shared across 100 engineers.
Borgstrom: One of the things that’s enabling that is the emergence of advanced use modes, such as transaction-based verification. It used to be that you’d have to set up a dedicated lab to do in-circuit emulation and have physical access to the box in order to do your work. What we see is a real trend where the environment around your design can be represented in a host machine. You don’t need physical access to the emulator itself. That facilitates physical access to the emulator being put into a data center from around the world.
Schirrmeister: There’s a bifurcation between project-based and centralized. There is a set of companies with a central organization, and then you have the satellites. They also want control and access. And there’s a new set of companies that buys an emulator as a resource for a project. What’s the balance? It’s hard to tell. The ones that move the needle the most are the big, centralized installations. We all compete fiercely on those. For the smaller ones, it’s a question of usability. Do you have the right use models and the versatility in use models? Can you do everything from acceleration to ICE to low power to gate-level to post-silicon. Where the biggest money comes from, time will tell.
Smith: Shared resources work fine until you find out the job won’t get done for three days. Then you want your own box.
Schirrmeister: We have presentations from big chipmakers using a central resource with complex load sharing. It’s not only when do you get the job. It’s also whether you set the triggers right to get the right sequence. If you miss your DUT you may have to wait two days to get your next run. It’s like execution units.
Smith: When I was at LSI Logic there was a very large processor company doing a five-chip set. They came back and said they don’t need all that stuff. I said, okay, the project will be a month late. Four chips would be done on time and the fifth one would be sitting on the shelf waiting for resources.
SE: Utilization rates for servers in data centers is considered optimum at 80% to 85%. What is it for emulation?
Schirrmeister: It’s 100% 365 days a year now. If you get to 80% you’re in trouble. If it’s down for 30 minutes, we get a call from the CEO. You’re driving 24 x 7.
Borgstrom: Reliability and uptime and mean time between failure are key metrics that emulation customers pay a lot of attention to. Emulation vendors pay a lot of attention to those numbers, too.
Kenney: We’ve even developed a program that sits between the standard queuing engines, like LSF and NetBatch. It manages all of the emulators. You just pile jobs on it and it runs. The other thing that’s happening is the interactive debug while you’re sitting on the emulator is being pushed off. You can’t afford to pay for resources like this and have someone poking at keys to figure out what’s gong on for four hours. You need to grab debug data from a batch during hardware or software, then debug offline to get the next one running.
Borgstrom: In terms of reliability, it’s not just being able to fire off a job. The larger the chip you use in the emulator, not only do you get a power savings, you also get a reliability savings. There are fewer wires to go bad, fewer solder joints, fewer cables. Over the past couple years we’ve seen about a 10X improvement in mean time between failures. You fire off a job, you don’t have to worry about whether the hardware is going to cause your job to fail.
Schirrmeister: That’s part of use-model versatility, and off-line debug. You have traces, and you give it to a larger number of users. That makes it applicable for more users. More software developers can get access and use it in parallel.
SE: How important is the power drawn by the emulators in this whole equation?
Borgstrom: We are moving and had to think about that internally for our own servers. For analysis purposes we priced out a co-location market in the Bay Area and what it would cost to run our emulators at a co-location site versus our own site. That’s increasingly an operating metric people look at. Do I need to build out a new data center with power and cooling? It’s not cheap.
Kenney: I’ve tried to sell based on power savings in the past. Nobody cares. When they’re considering the cost of the electric bill versus whether they can get a design done, they’ll pick the one that will get the design done. Power and cooling is way down the list in terms of decision-making.
Schirrmeister: I fully agree. You can simplify it to a power number to say that it will save you money. But what do you do with it? How many debug cycles can you run through? The higher-power consuming system may get you to results faster, like trace buffer depth. If you have half the power, but you need to run four times as often, it’s still worse power consumption overall. It comes down to the use model.
SE: Because emulation is so capable of many things, is it leading to incremental business involving new applications?
Rizzatti: It could be done, but I haven’t seen anyone saying anything about that.
Schirrmeister: All of us have been looking at it. Can we do things like banking? There are companies out there where it becomes more efficient to sell what you can do with the box than selling the box itself. Celoxica sold high-level synthesis. They figured out their particular engine could do banking algorithms really well, so now they sell boxes to bankers that are 3/10ths of a second faster in trading, which is worth millions. That’s something we have been looking at—how can we expand the reach. But what happens in the pure systems world? There are some customers, although it’s not a mainstream use model, looking at how to build a system based on emulators. They’re simulating a system of chips, not only a system on chip. An emulator fundamentally uses RTL, so you need access to the RTL netlist. But if you can arrange with your platform provider to use the RTL plus other things, that’s a system model that is very interesting. Can you do performance analysis? Do you have the right traffic going between those chips?
Rizzatti: The issue here is the compiler, not the emulation technology. The compilation technology has to go along with it. Today there’s no data structure to compile into FPGAs. If you want to convert data from medical, it’s available. But a compiler to describe for that particular application and produce an object that will map to FPGAs is not there. I’m not aware of anyone doing this research, either.
Schirrmeister: It has happened in search. I’m aware of one architect who over the weekend created something where in one clock cycle he could find every phrase in the Bible because of the way it was mapped and compiled. Is that useful for search? Potentially. There are attempts to look into these markets. But to add some realism, there are still markets in the traditional semiconductor space where emulators can grow quite a bit.