Embedded processors of tomorrow - Embedded.com

Embedded processors of tomorrow


Goodbye binary arithmetic, instruction sets, and assembly language programming. Hello 4.5-billion transistor, 250GHz processors. O brave new embedded world!

There are 10 kinds of people in the world: those who understand binary and those who don't. Those of us who do understand binary (and who get the joke) will be fewer and farther between in a world 15 years hence. Like the ability to tell time on a clock with hands or tie shoelaces without Velcro, an understanding of binary arithmetic or assembly-level programming will be a lost and forgotten skill.

That's a good thing. Few mourn the loss of broadswords, 8-inch floppies, or phrenologists. The purpose of embedded technology is to blend into the background, to become an everyday part of our lives. Our grandchildren are not likely to become programmers in the sense that we understand it, any more than most of us became colliers, lamplighters, or mule-team drivers.

How far is fifteen years?
When predicting the future, it's pretty tempting to extrapolate from existing data points. Fifteen years ago was 1988. Cheers was still on TV, Roger Rabbit was in movie theatres, and Bobby McFerrin had the #1 record. Motorola's 68030 chip was brand new and not yet considered an embedded processor. Intel's 25MHz 486 was still a year away; 386-based PCs humming a 16MHz tune were at the top of the charts. The bizarre Inmos Transputer chip was in its fourth (and nearly final) generation. In the burgeoning RISC camp, the 88100 was new, shoring up the early 88000. Berkeley's (and later Sun's) SPARC dynasty had just begun; Stanford's MIPS project was only slightly older. National's 32000 architecture had circled the drain for the last time, proving that “elegance is everything” except what's needed to survive.

PowerPC hadn't been invented yet. That gleam didn't enter IBM's eye for another three years, and the first 601 chip was two years beyond that. Macintoshes were straining 68020s and '030s. England's ARM had yet to discover royalty as the basis for its business; the first ARM6 chips were still five years over the horizon. OS/2 adorned some IBM machines, to profound disinterest from customers. Steve Jobs' NeXT introduced its first black box with a CD-ROM, no floppy drive, and something called Display PostScript.

In 1988 we had yet to see a million-transistor microprocessor chip. That distinction was to go to Intel's 486DX the following year. Now 15 years on, a mainstream Pentium 4 processor contains 45 million transistors, but that's not half the complexity of graphics chips from nVidia and ATI, with 125 million transistors. In 2002, about 6 billion new processor chips were made and sold—one new processor for every man, woman, and child on the planet. In total, something like 60 million transistors were fabricated for every human. Semiconductor transistors are as plentiful as grains of rice, and almost as cheap.

Connecting the dots is easy. Drawing a curve from yesterday's 16MHz chips to today's 3GHz parts suggests we'll have 250GHz processors in fifteen years' time. The transistor trend line passes through 4.5 billion transistors per processor in 2018.

Swell, but what does it all mean?

Future processors
Extrapolating from 15 years ago suggests that some of 2018's most popular microprocessors haven't been invented yet. That's definitely true. New processors and instruction set architectures are being invented weekly. Some won't survive the next year or the next round of funding, but many will find a niche in embedded's ever-changing ecosystem. We need new processors because we have new embedded systems to use them.

Future embedded processors will be a combination of the unrecognizable and the all-too-familiar. Unrecognizable because they'll be festooned with coprocessors and accelerators, intrachip networks, elaborate value speculation, unfathomable branch prediction, enormous multilevel caches, and inscrutable instruction sets. All-too-familiar because old embedded processors never die. Expect to see 8051s, Z80s, and more descendants of the x86 line soldier on in low-end systems.

Multiprocessors and multiple processors per device will be the rule. Microprocessors themselves (that is, the CPU cores) are already way too small to fill up a normal chip. Today designers fill the rest of their chips with peripherals, caches, memory, and more microprocessors. (The average ASIC already has more than three processors if it has any at all.) There's plenty of room to make microprocessors fabulously complex without taking up too much silicon. Svelte silhouettes count for little in this business.

Turning superscalar processors into massively parallel machines won't have much payoff. Superscalar execution wastes time and power, almost by definition. Instead, future generations of ¼ber-processors will cooperate, sharing tasks among hundreds of processors on the same chip. Interprocessor communication and on-chip networks will take up most of the silicon. The processors themselves will take a technological back seat to the tiny networks that help them communicate.

And what of RISC and CISC? Instruction sets will become increasingly irrelevant in the sense that we'll never see them. Nobody will program in assembly language; mnemonics will be like so much Sanskrit. Besides, the “real” instruction set may not be published, identified, or knowable. Think of Transmeta's Crusoe processor, with its apparent x86 instruction set but with a different and undocumented set of internal hardware instructions. When you're programming in a high-level language, who cares?

The whole concept of instruction sets will be a quaint anachronism anyway. Processors will likely adapt their features over time, learning and morphing as they adjust to changing workloads. Mass-produced processors may all ship from the factory the same way, but they'll likely transmogrify into something different in the field. Like children graduating from elementary school, they'll have enough education to get by but their ultimate character will depend on life experience.

At the extreme end of this spectrum will be “soft” virtual processors, with almost no native instructions at all. Instead, they'll run emulation or binary-translation code as a kind of ultra-low-level operating system that enables them to execute software from any legacy processor. Itanium XXXVII, PowerPC G500, ARM42—it'll all be the same to these chips.

Thinking inside the box
There's a great scene in Minority Report in which Tom Cruise's cereal box plays 30-second commercials while he eats. That's just the kind of strange-but-true embedded system we'll all be accustomed to in 15 years. They'll be a combination of the familiar and mundane with the leading-edge and technological. All rolled up in a package that makes money for someone. I would expect that cereal box to be touch-sensitive and to track click-throughs, too.

Today the average American household contains around 40 microprocessors (not counting a few dozen per car and another 5 to 10 in personal computers). Figure on that number growing to about 4,000, most of them dedicated to entertainment. Video games, video terminals, multiple levels of wireless networking, media caching, and always-on access to friends, news, entertainment, and data will keep our homes humming and millions of MIPS flowing. With terabytes and petabytes of storage, you'll be able to store every book, every song, and every movie you've ever seen or ever want to see.

Toys drive technology. Most of the growth in embedded sales, and most of the advancements in embedded processors, will come from consumer electronics, toys, games, and entertainment—not computers. That's because computing problems in everyday life don't get much harder, but virtual reality problems do. We don't all need to do our own weather prediction or (one hopes) calculate missile trajectories, so the demand for “computer” performance isn't great. But we're a long way from perfect, photo realistic, real-time video. That's an area that can easily soak up an infinite amount of computing power and produce cool products we'll pay for.

Physical materials will still cost money but processing horsepower will be essentially free. Speakers, plasma screens, and headphones have a cost but spatial audio positioning, noise cancellation, and so forth will be easy and ubiquitous. Wholly synthetic actors will star in computer-generated movies produced in real time based on market demand or daily events. Many celebrities will be cybernetic.

With processing power cheaper than mechanical equivalents, even mundane devices get booted up to the next level. Rear-view mirrors with embedded image sensors will recognize an impending collision and warn both drivers. Tiny webcams with wireless connections that cost next to nothing and are no bigger than a coin will push the boundaries of privacy. 10GHz processors will be so cheap and ubiquitous that they'll be disposable. We'll throw away the equivalent of a Cray supercomputer with each week's trash. GPS receivers and wireless transmitters will be so common that every solid object of any value will know where it is and can tell you so. Cell phones (and wireless data networks) will become nonproducts: uninteresting by themselves, but an added feature on another product, like AM radios today. Even now, researchers can transmit sound waves through arm and hand bones; to talk on the phone you stick your finger in your ear.

Semiprecious Joules
Ancient computers required enormous power supplies. Today's handheld PDAs run for weeks on battery power while outperforming their primeval cousins. Power efficiency will continue to rise, and power consumption will fall. Fifteen years from now it'll almost be possible to run a microprocessor on the electron energy that decays from its own materials. Piezo-electric power from shaking or squeezing the chip will be enough to drive most low-end microprocessors, like a self-winding wristwatch. But our demand for processing performance will rise faster than power consumption will fall, so such a scheme won't be widespread. Moore's Law throws us transistors at a 38% compound annual rate but “Eveready's Law” does not, alas, keep up.

The far side
Fifteen years is a good long while, but not long enough for some of the stranger predictions to come into being. Maybe there won't be microprocessors at all, for example. Dynamically reconfigurable logic has captured the imagination (and investment capital) of many who see it as an efficient alternative to microprocessors. Instead of programming a fixed processor with variable instructions, why not just make the hardware vary over time to fit the requirements?

Reconfigurable computing promises chameleon-like hardware that changes itself on the fly. It's a sexy, alluring, and elegant new technology that has everything going for it—if you don't count the last few hundred years of human history. Technical elegance doesn't win market share, and in a world where computers are dominated by x86 hardware and DOS-derived software, sophistication seems downright counterproductive. Fate has a perverse streak.

Nor will we see Java processors in the future. Java will be remembered as a spectacular marketing success overshadowing a minor technical curiosity. It's fundamentally unsuited to hardware—any hardware—and will gradually be forgotten. Microsoft's C# knock-off will likewise edge its way toward the abyss. More likely, some as-yet undiscovered hardware-cum-software language will emerge as the design and programming tool of choice for new processors.

Inertia and momentum exert their influence on technology, just as they do on tides and bodies in motion. We don't do what's best for us; we do what's easy. Legacy software and familiar user interfaces have a way of sticking around in spite of “better” alternatives. The human animal can accept only so much change. We adapt slowly to the world that we ourselves are altering. Our fastest bullet trains run on rails 4 feet 8 1/2 inches apart because that's the width of a horse-drawn wagon, which in turn followed the ruts left by Roman armies of a thousand years ago. Backward compatibility colors and flavors a lot of today's technology, often against all intellectual reason.

It's tomorrow
Embedded processors make up 98% of all the processors sold. (PCs fill in most of the rest; workstations are statistically insignificant.) That ratio will get even more lopsided in 15 years—not because PC volume will drastically fall, but because embedded volume will drastically rise. Sales of 32-bit embedded processors have doubled and tripled in recent years; we're witnessing the proverbial “hockey stick” curve of rapid early growth. Much of that growth is in communications infrastructure.

Our phones, computers, traffic lights, and gas meters now collectively rely on millions of embedded processors that we never see and care little about. Closer to home our cars, kitchens, and cable TV rely on hundreds more embedded processors. A musical greeting card has more computing power than NASA's lunar lander did in 1969. We wear computers on our clothes in the form of pagers, e-mail terminals, PDAs, and mobile telephones. Thousands of people have embedded processors under their skin, as pacemakers or hearing aids.

“Any sufficiently advanced technology is indistinguishable from magic,” wrote Arthur C. Clarke. Today we speak into the air and a person across the globe hears us perfectly. Everyday we look into a glass mirror and see actors we've never met on a stage we've never visited. We twitch our thumbs and a sword-wielding hero battles dragons in an epic tale that has no ending save what we determine to give it. Magical, indeed, yet trifles in today's world. Magical indeed will be the treasures that await.

Jim Turley is an independent analyst, columnist, and speaker specializing in microprocessors and semiconductor intellectual property. He was past editor of Microprocessor Report and Embedded Processor Watch and has written several books, including the Essential Guide to Semiconductors and Advanced 386 Programming Techniques . For a good time, write to .

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.