Survey: Who uses custom chipsFully two-thirds of our developers use custom logic of some sort, as Figure 1 shows; only 35% overall answered "none of the above." The averages varied quite a bit by industry, with automotive developers showing the highest usage of custom chips (only 13% answering "none of the above"); aerospace contractors were the second strongest in the category. Custom logic vendors have the hardest time penetrating networking companies; they answered negatively 57% of the time, the highest in our survey. Developers in the audio-electronics and power-generation industries came in second and third to last in this category.
Figure 1: Custom logic usage
On the flip side, aerospace developers were the biggest fans of FPGA chips, with 64% answering in the affirmative. Close behind were video (62%), military (59%), and automotive (58%) developers. The lowest FPGA affinity scores came from the security (24%), networking (27%), power-generation (29%), and government (29%) industries.
Interestingly, security was the only industry segment where FPGAs weren't the most popular form of custom logic. PLDs and CPLDs edged them out of first place by a few percentage points.
The custom logic categories are not mutually exclusive and many projects include more than one type of custom chip. That said, we can safely assume that the three choices for cell-based ASIC, structured ASIC, and full-custom ASIC are all mutually exclusive; very few developers would be working on two such projects at once. With that in mind, we can say that about 24% of our survey takers are developing a "real" ASIC of some kind.
That leaves 41% of developers overall who are using "softer" logic such as a PLD, CPLD, or FPGA. In fact, 41% of respondents said exactly that: they have one or more FPGAs in their current project. An additional 30% also said they're using one or more PLD (programmable logic device) or CPLD (complex PLD) chips in their project. There's obviously some overlap among FPGA and PLD/CPLD users, since the two categories together exceed 100% after we subtract "none of the above."
It's also likely that some of the ASIC developers are also using PLDs, although we'd expect less overlap than with the FPGA contingent. Once an ASIC program is underway it tends to include as much of the miscellaneous system logic as possible. There's little incentive to leave features out, so FPGAs and ASICs don't often go together. Inexpensive PLDs, on the other hand, can be found cohabitating with an ASIC in order to handle features that change frequently or that got added to the board at the last minute.
A little quick arithmetic shows that two-thirds of developers who are using any custom chips at all are using FPGAs. They're the overwhelming favorite even though they're more expensive than PLDs or CPLDs. What's surprising is that the percentage of "hard" ASIC developers is so high, given the exorbitant cost of ASIC engineering. With 24% creating ASICs, they're almost as popular as PLDs.
We asked the developers using FPGAs whether they include a processor in the chip, either a "hard" processor permanently embedded in the device by the manufacturer, or a "soft" processor that the developers included themselves. As shown in Figure 2, overwhelmingly, developers told us, "Nope, there's no processor in my FPGA."
Figure 2: Processor in your FPGA?
That's disappointing news for Actel, Altera, Xilinx, and the other FPGA vendors who have worked to include processors in their chips. Altera and Xilinx offer ARM and PowerPC processors, respectively, in some of their devices. These combo chips combine a popular 32-bit microprocessor with equally popular field-programmable logic. The idea was to create an all-in-one chip that could become the "platform" for different users' applications. With just under 18% of respondents saying they use such chips, these platforms have been slow getting off the launch pad.
The soft-processor alternatives did even more poorly. A little over 14% of FPGA users say they download a "soft" processor into their FPGA. Soft processors differ from hard processors in that they're not a permanent part of the chip. Instead, they're programmed into the device just like any other logic circuit, making them more flexible—and free. Xilinx and Altera practically give away the designs for their soft processors (MicroBlaze and NIOS, respectively) in the hope that FPGA users will adopt them. It's a compelling idea but one that hasn't gained a lot of traction among our survey takers.
The future looks brighter, though. When asked about future plans, 19% of respondents said they'd probably use a soft processor in their next project. Another 24% said they'd likely use a hard FPGA processor. That's better than 35% growth in both segments—not bad at all. Only 57% of current FPGA users said they still had no plans to include a processor in their next FPGA project.
Other soft hardware
Although few FPGA users embed processors in their chips they all download something into the device. We asked what kind of standard or standardized IP they use; the results appear in Figure 3.
Figure 3: Downloaded FPGA IP
As you can see, simple UARTs and other low-speed serial interfaces are the most popular; darned near half of our FPGA users download a UART into their chip. PCI (including PCI-X and other PCI derivatives) was used in one-third of FPGAs, with Ethernet and USB controllers nearly tying for third in popularity. After those top four, the numbers dropped off a bit, with cryptographic functions (encoding, decoding, or general security logic) leading the rest, followed by video interfaces such as VGA, storage interfaces for SATA or Firewire, standard buses such as VMEbus, PC/104, FutureBus, and so on. For followers of the bus wars, RapidIO showed up on twice as many FPGA users' surveys as HyperTransport.
The "other" category among soft IP included ATM, RocketIO (not to be confused with RapidIO), CAN, DSP functions, SerDes, HDLC, Raceway, and the troubling "I don't know."
When we queried users about signal processing in a separate question, 40% said they use their FPGA for DSP functions of some sort, with the other 60% stating they did not. Yet the number of FPGA users downloading an actual DSP processor core into their chip was quite small, suggesting that although the FPGA is used for signal processing, it's not used as a DSP, per se. "Signal processing" covers a lot of territory outside of well-defined DSP processors.
Vendors and reuse
You gotta buy your chips from someone, and our survey-takers buy from a lot of different sources. As Figure 4 shows, Xilinx and Altera ranked as the top two, largely because they're the biggest suppliers of the FPGA chips that scored so well in our survey.
Figure 4: Logic vendor
One of the criticisms sometimes leveled at FPGAs is that they're seen as too expensive to use in production. FPGAs may enjoy low development costs compared with "real" ASICs but their unit cost is often much higher. Although there are $3 options, $1,000 FPGA chips aren't unheard of, either. Our survey shows that developers who use FPGAs generally use the chips in production (43%) versus those who use them only for prototyping (20%). Figure 5 displays the breakdown. Those prototype devices presumably turn into ASICs, closely matching the percentages we saw earlier.
Figure 5: FPGA usager
Figure 6: Reuse hardware IP?
As Figure 6 shows, most custom-logic developers reuse some of the circuitry, or hardware IP, from a previous project, and that it's IP originally developed in-house. About one-tenth reuse circuitry they've purchased or licensed from a third party, and a tiny fraction reuse hardware IP they've found in the public domain. The remaining one-quarter of developers don't reuse any hardware IP either because this is their first custom-logic project or because previous projects were all unrelated to the current one.
The very low percentage of public-domain IP reuse suggests that "open-source" hardware isn't nearly as popular as its software counterpart. Developers are understandably wary of betting their expensive hardware project—potentially worth millions of dollars—on free circuitry. As with open-source software projects, free hardware IP is usually offered without any support. Unlike open-source software, hardware IP is usually difficult to modify and even tougher to verify. For most developers, it's well worth the cost to license commercially supported IP.
Back on the topic of processors, Figure 1 showed that almost 14% of all respondents are using a configurable or customizable processor. That's a category that wouldn't have existed a few years ago, but user-defined processors are slowly gaining momentum. The response rate closely parallels that of full-custom ASICs, which is probably not a coincidence.
ARC and Tensilica are the two best-known vendors of user-customizable processors, and both companies have over 75 licensees apiece. Their designs are not compatible with FPGA or CPLD implementation; they're intended for ASIC developers who want to add their own touches to an embedded processor. Configurable processors can yield big performance benefits, but they're a bit of work to customize, and they can be expensive. Hence, the correlation with full-custom ASIC developers.
In future months, we'll examine the survey data for different 8-bit, 16-bit, and 32-bit microprocessors and DSP to see what instruction sets are most popular and which are declining. There's plenty of good, rich data left to farm. Stay tuned for the next installment.
Jim Turley is editor in chief of Embedded Systems Programming magazine and is the director of the Embedded Systems Conference. You can reach him at email@example.com.