What’s a “one-man dog,” you ask? It’s a dog that works for one person and one person only. (It’s also the title of a 1929 movie about a dog that solves a murder mystery, but that’s a story for another day.) What does this have to do with embedded processing or DSP, you ask? Well, I use the phrase “one-man dog” to refer to a processor architecture that is captive to a single semiconductor vendor.
When we started BDTI in the early 1990s, it seemed that just about every digital semiconductor company had its own processor architecture. Many companies had several. In those days, the processor core occupied a large portion of the silicon area on a typical chip, and there was plenty of scope for chip vendors to differentiate their products based on having a unique processor core that was faster, cheaper, more energy efficient, or easier to use for a particular set of target applications.
A lot has changed in 20 years. These days, a single processor core rarely occupies more than a small minority of the silicon area in an embedded processor chip, ASIC, or ASSP. The bulk of the area is consumed by memory, buses, multi-core structures, coprocessors, and I/O interfaces. As the logic capacity of digital chips has soared, the complexity of those chips has exploded. Whereas the processor core was once the main attraction, now it’s just a building block. Chip architects these days have many more things to worry about—and many opportunities to differentiate their products—outside of the processor core.
At the same time, processor cores have become much more complex—particularly those targeting high-performance applications. And of course, with processor design, there is no rest: As soon as the current design finished, the race begins to design the next-generation core with higher performance, lower power, and lower cost. As a result, the investment required to maintain a competitive high-performance processor core has grown to vast proportions—tens of millions of dollars per year.
At first glance, the math here seems pretty simple: The cost of developing processor cores is growing, while the differentiated value of processor cores is falling. Therefore, chip companies should stop making their own processor cores, and instead make use of proven, licensable cores. And indeed, that’s the direction in which many chip companies are heading. For example, when Trident Microsystems recently bought NXP’s set-top box and digital TV product line, it did not buy NXP’s venerable TriMedia processor core. (NXP then set out to find another buyer for TriMedia, but failed to do so.)
But things often aren’t as simple as they seem. One significant complication for many chip companies is the “installed base”; that is, the software that has been developed and tuned for the existing processor core. Often, significant software investments have been made by the chip companies themselves, by their software-provider partners, and by their customers. The more “legacy” code that exists, and the more closely that code is tied to the existing processor core, the higher the switching costs for everyone involved. And there’s risk here: If system companies are faced with significant switching costs due to their chip vendors switching processor architectures, it’s likely they’re going to take the opportunity to see what competing chip companies have to offer.
Another complication is organizational inertia. If a chip company has been making its own processor cores, it has a team of processor core engineers: architects, RTL designers, compiler writers, etc. These people are often some of the most talented and respected engineers in the company. And understandably, they are usually not eager to give up the work that they’ve been doing—and for which they have been recognized.
Such complications mean that established chip vendors sometimes decide to continue developing their own processor cores even when a straightforward analysis suggests they should stop. Sometimes, that’s the right decision. For start-up chip companies, on the other hand (and for established companies entering new markets), the decision is easier, and buying rather than making a processor core is usually the right decision. So, while start-ups and new market entrants certainly face very real challenges, they have at least one significant advantage: they can avoid spending tens of millions of dollars per year reinventing the wheel.
Add new comment