TMTA: The Software Chip
Here's an upbeat article from MIT's Technology Review:
techreview.com
The Last Word, first:
So 10 years from now, we can have 3,000-bit-wide architecture or whatever the state of the art will be. We can make that change. Everyone is focusing on this first chip. That's not the point. The point is, for the next decade we have a technique that's going to let us move faster than anyone else does. It's a cliché right now, but we can work on Internet time.”
What is it about Transmeta CEO Dave
Ditzel that makes you want to believe
him? Maybe it's the way he
unabashedly uses words like “cool” and
“neat.” Maybe it's because he had the
audacity to build his upstart chip
company within view of Intel
headquarters. Maybe it's because he
never completes a sentence, so
enthusiastic is he about Crusoe, his
company's brand of microprocessors.
From last January, when Crusoe was
announced in a blaze of fanfare, until
mid-August, when the company filed to
go public, Ditzel made himself hoarse
pushing the Crusoe chip. Whether in
front of 200 engineers or a single
reporter, his message was unflagging:
Crusoe—the Intel-compatible chip with one-tenth the power requirements of a
Pentium III—is going to change the world of computing forever. “Crusoe is
low-power, it's compatible and it's high-performance,” he said in one of a series of
interviews held before the August filing. “That's our mantra.”
This summer, the company and Ditzel went silent for the quiet period that follows
every initial public offering. But by then the Crusoe message had developed a life
of its own: Not since the Apple iMac had there been such a fuss in Silicon Valley
like the one Crusoe has brought ashore. It's no surprise that Valley insider rags
Upside and Red Herring ran Transmeta as their cover stories last spring, but
before the quiet period began, Ditzel was also quoted in Time, USA Today and a
horde of other consumer publications. Transmeta's publicity efforts have fed in
part on the company's hiring of Linux author and open-source software guru Linus
Torvalds. Torvalds has been part of the software design team at Transmeta, and
has lately been working on a version of Linux that will complement Crusoe's
application in the exploding market for mobile devices.
The IPO itself is a dour, close-mouthed document that reveals little of Transmeta's
future design plans, and is instead full of warnings about what might go wrong on
Transmeta's road to profitability. Indeed, as of this writing, not a single Crusoe
product has shipped in appreciable volume. The company lost $41 million in 1999
and another $43 million in the first six months of 2000; the prospectus makes it
clear that investors shouldn't expect to see profitability in the near future.
But it would nevertheless be difficult to find a startup that began more auspiciously,
or with a better lineup of initial customers. Last May, America Online and
Gateway declared that Crusoe will power a new line of household appliances that
will feature wireless Net access. IBM, Hitachi, NEC and Fujitsu followed suit in
June, with announcements of Crusoe-based notebook PCs that will run all day on
ordinary batteries. Sony followed with an announcement in August that Crusoe will
power a future version of its Vaio PictureBook line of notebook computers. Not
bad for a chip company with no fabrication plant and no track record—and
whose main asset is, as Ditzel puts it, “a vision of a better way to build
microprocessors.”
“Starting now, the Transmeta approach will be the smartest, quickest, cheapest,
most reliable, most flexible technology to solve virtually any computing-related
problem,” says John Wharton, microprocessor design consultant, Stanford
professor, and former design engineer for Intel. “Fifty years ago, the most
sophisticated systems were built using vacuum tubes. Ten years ago, the state of
the art was complex, fully integrated megaprocessors like the Pentium and
PowerPC. I see Transmeta as representing the next breakthrough in fundamental
design technology.”
The trail that Transmeta is blazing will lead to chips that use significantly less
electric power. That's good news for anyone who uses a laptop computer or other
portable electronic devices. But more profoundly, Transmeta has found a way to
radically improve the ability of chip designers to make changes in their products
without alienating the huge libraries of software that have been written to run on a
particular piece of hardware. They have, in a sense, removed the pesky governor
from the engine of chip-making progress.
Architecture of Liberation
Although there's been plenty of press coverage of Transmeta as a new venture,
what often gets lost is the technology itself. Crusoe is a hybrid software/hardware
chip whose sole purpose is to run software designed for other microprocessors.
Much of what Intel and others accomplish in silicon, Transmeta has shifted to
software. The advantages? First, the chips themselves take less silicon, making
them cheaper to build. Second, a simpler chip consumes less power—a
paramount concern for portable computers. But perhaps the most far-reaching
impact is that in crafting Crusoe, Transmeta has come up with an innovative
approach free of many of the problems that have plagued chip design for the last
two decades.
Before Crusoe, every microprocessor ever built has come with its own published
“instruction set”—an explicit contract that spells out how the chip will work with
software. An instruction set promises that if developers write software that does
X, the resulting action from the chip will be Y—now and forevermore.
The problem is that once a new chip is
designed, it is locked in time. As software
inventory for the chip builds, it becomes next to
impossible to make improvements to the
instruction set. Software development is
hampered too, since any new program must
obey the laws of the chip's instruction set in
order to work. Microprocessor designers want
chips to run faster, but they must also make
them run on existing software. So they squeak
out increments of speed with tricks such as
resequencing instructions to the processor. But
implementing major changes is next to impossible. It's like a very bad three-legged
race, with software and hardware engineers tied at the hip—never able to move
fast toward adopting state-of-the-art products, so dependent are they on each
other's design choices and the choices of previous generations.
Ditzel himself has firsthand experience with the difficulty of making fundamental
improvement in a chip's initial design. At Sun Microsystems, where he worked
before founding Transmeta in 1995, he was in charge of changing the instruction
set for the company's SPARC brand of microprocessor. Although he embarked
on the project in 1990, it wasn't until last year that the new instruction set was
ready for use. “You need time for the industry to catch up, to get software out
there, to get applications converted,” Ditzel told TR before the company's IPO.
“It's a really big deal.”
Software Camouflage
To wrest their chips from outdated instructions, microprocessor designers will
periodically throw everything out and start over with an entirely new chip,
complete with a brand new instruction set. It's a process that Intel is struggling
through with its much-delayed Itanium microprocessor, which will be the
company's first chip that routes data around in digital swaths of 64 bits—that is, on
a 64-bit-wide “bus.” Freeing designers from the current generation's 32-bit bus
will result in a great leap forward in performance. But starting over also results in a
chip that will initially have no software to run on it, hardly an ideal state. Even if
software developers cooperate and begin to write code to the new instruction set,
this approach only works once: Then you're back where you started, with legacy
software and a years-long cycle to make any fundamental change.
Ditzel has tried the start-over approach to chip design more than once in his
career. Two decades ago, as a graduate student at the University of California at
Berkeley, he co-authored a paper entitled “The Case for Reduced Instruction Set
Computing.” This seminal work inspired an entire school of microprocessor
design; today, so-called RISC chips are everywhere.
After graduate work on RISC design at Berkeley, he moved on to design a
RISC-chip variation called CRISP at Bell Labs; CRISP, however, never gained
wide support from software developers. Ditzel then made a third attempt to design
a new microprocessor when he worked on a gallium arsenide chip at Sun that was
never produced. “It was like I was telling people: ‘Look! You can use this great
new microprocessor—all you have to do is throw out all your software and start
over!’” Ditzel said. “I've fought that fight for 20 years, and I've given up.”
But he didn't really give up. Instead, he found a way out.
While at Sun in the early '90s, Ditzel was influenced by the work of Russian
supercomputer expert Boris Babayan, with whom he had informally collaborated,
and whom he names as a key mentor in his developing thought about chip design.
At the time, Babayan and his company Elbrus were experimenting with a
technique known as dynamic binary translation and compilation (which Transmeta
has given the much more market-friendly name “code-morphing,” a term they have
since trademarked).
Writing code so that one kind of software can run on another kind of hardware is
an old idea: IBM, for example, did it back in the 1960s. Results of these attempts,
however, were always hopelessly sluggish. But chips were getting faster all the
time. By the early 1990s, designers were postulating that there might be a way to
translate from one instruction set to another so rapidly that performance would
barely suffer. Instead of being a static, one-to-one translation of each instruction,
the technique could be dynamic, examining the application for inefficiencies in real
time, correcting them and remembering the corrections.
It is counterintuitive to think that putting an additional layer of software between an
application and a CPU wouldn't slow things down—it's like saying a curved line
between two points is shorter than a straight line. But the relationship between
software and hardware is no longer a straight line: Because of the inefficiencies
caused by years of developing around the same instruction set, dynamic translation
could, in theory, improve performance. On the hardware side, the process of
jamming more and more circuits onto a chip to eke out the last performance gains
can actually backfire, slowing things down. Software, too, is rarely as efficient as it
could be out-of-the-box: Applications developers with an eye on a ship date will
freeze code when it works, not when it's perfect. Dynamic translation could
theoretically find the slack and tighten it.
Before Ditzel founded Transmeta, translation techniques had been used only to
make existing, noncompatible software and hardware speak to one another. Ditzel
and his co-founders made an intellectual leap: If an additional layer of software
could make applications run on noncompatible hardware, what was to stop them
from making radical changes in the underlying hardware itself, taking advantage of
the latest capabilities?
In 1994, Ditzel and co-founder Doug Laird were both coming off a project at Sun
designed to make Windows run better on Sun workstations, using dynamic binary
translation techniques. “We realized that if we could just add some features to the
hardware, we could actually make this thing go pretty fast,” says Laird. “It was a
cool idea,” he adds, recalling that Sun wasn't interested in changing its processor
design to make it better at running applications that had been written to run on
standard Intel chips. Ditzel and Laird struck out on their own. Ditzel recruited
Colin Hunter, a respected expert in emulation techniques, and Robert Cmelik,
who had been doing work in code optimization at Sun.
As is often the case in technology innovation, practice proved harder than theory:
Transmeta's first chip design ran so slowly that it took the chip half an hour just to
boot the operating system. But with each of four chip revisions, the team learned
more about binary translation. Five years of painstaking work—performed by a
brigade of 200 engineers backed by several hundred million dollars of venture
capital—produced a chip that ran fast enough to compare favorably with Intel
processors. In January of this year, Transmeta announced the first two hybrid
silicon/software chips in the Crusoe line. The first, called TM5400, is a
700-megahertz chip for the ultrathin, ultralight Windows notebook PCs. It runs
software written for Intel chips on a fraction of the power a Pentium consumes.
The second, the TM3120, is a 400-MHz chip designed to run Internet appliances
using a version of Linux that Torvalds developed for mobile devices.
Both chips present a face to software developers that is completely compatible
with the instruction set in Intel's processors. Underneath are VLIW chips, for
“very long instruction word,” an architecture with a 128-bit-wide bus that can
combine Intel chip instructions into longer strings and thus execute them faster.
Between the outward-facing instruction set and the underlying hardware is
Transmeta's code-morphing software, which translates the Intel-style instructions
into a form that Crusoe can handle, optimizes their execution and stores the
optimized executions in memory. The next time the chip encounters the same
operation, translation is no longer necessary. The code-morphing software (which
resides in a read-only memory chip) is the first program to launch when the
processor boots.
Because so much of Crusoe's functionality has been moved from hardware to
software, the chip is far simpler than a comparable Pentium processor and
requires only one-fourth as many transistors. A side benefit of fewer transistors is
that Crusoe uses far less power to run—hence Transmeta's decision to target its
first chips at the mobile market. Another advantage to the Crusoe approach is in
shortening the time needed to develop a new chip. With much of the design
residing in software, Ditzel says some customers have already asked for changes
in the instruction set and that Transmeta engineers could implement them in 24
hours. While this probably doesn't include time for any kind of bug testing, it's
nevertheless clear that Transmeta has found a way to drastically shorten the
development cycle.
Nick Tredennick, co-architect of the original Motorola 68000 (the processor that
powered the first Macintoshes), and now an independent microprocessor design
consultant, is not alone in concluding Ditzel is on to something. “When I first heard
about Crusoe, I thought it was just the latest fad, or a rehash of emulation, which
has never worked,” Tredennick says. But after hearing Ditzel speak, Tredennick
became a convert. Transmeta, he says, is doing something “fundamentally different
from what has been done since the invention of the computer.”
Transmeta's chips are “inherently simpler to design” than conventional ones, says
Stanford's Wharton. “You can make a software change, incorporate it into a test
version, run it and see if it works, all in one afternoon. In the hardware realm, the
turnaround time can be three to nine months. Intel may put 500 or 1,000
man-years into designing Itanium. The next Transmeta chip may require 10, or 20,
or 50. That's mouse nuts.”
The movement toward making chips that are hybrids of software and hardware,
rather than pure silicon, has caught on broadly. But Transmeta is likely to keep the
lead for the foreseeable future. That's because Ditzel was the first person to take
these ideas out of the lab, hire 200 employees to work on them, and build a chip
that worked. Along the way, he created at least two roadblocks that will slow
down his competition.
The first is the company's testing tools. Indeed, Transmeta's jewels are probably
not even the chips themselves, but rather the diagnostic software the company was
forced to create in the development process. The off-the-shelf tools that exist for
checking out conventional chips all assume that there is a static relationship
between software, a chip and a given instruction. Transmeta needed to solve the
problem of testing a microprocessor that changes dynamically in response to the
software it runs. Other companies will need to start from scratch to build their own
testing tools, which could easily take a year or longer.
A second obstacle is finding people with the
skills to design chips like these. While the
concept of a combined software-and-silicon
chip is sound, everyone at Transmeta appears
taken aback by just how difficult it was to pull
off the first time. “If you told me up front it
would have taken so long, I probably wouldn't
have joined Transmeta,” says senior engineer
Godfrey D'Souza, an early employee.
A key factor in Transmeta's success was
finding the people with the appropriate
cross-training in both hardware and software
development, in a world where computer
science degrees specialize in one or the other,
and where chip designers have always been
trained in hardware-only techniques. The lure of stock has allowed Transmeta to
raid other companies for people that have the experience—including six engineers
from Hewlett-Packard's rival Dynamo project alone. Now they have assembled a
team that is the only one to have produced a chip using these techniques. Other
companies will be hard-pressed to lure these engineers away. The presence of
open-source celebrity Torvalds has also helped Transmeta's recruiting efforts,
especially for young engineers.
Transmeta is coy about its plans for a next-generation microprocessor.
Speculation abounds on the Internet: Will it be Macintosh-compatible? Run Java?
Transmeta isn't saying. In past press conferences, Ditzel himself has acknowledged
his delight in planting deliberately misleading information about the company's
plans, a strategy he's not above using to throw off his competition. A stray
comment last spring about Transmeta's developing a chip for the mobile phone
market was widely reported and sent the stock of market leader ARM Holdings
plummeting.
One thing is certain: The company's patents baldly state that the technology is not
intended for Intel-compatible chips alone. And although Transmeta is going to find
it difficult to keep up with current orders, much less conquer new markets, there is
no reason to think that the Transmeta way of designing chips won't be widely
useful, or that Transmeta won't capitalize on its technology lead.
“We've learned the lessons from the past, and we won't get trapped again,” Ditzel
told us. “By having this software layer, we can change underlying hardware every
year. So 10 years from now, we can have 3,000-bit-wide architecture or
whatever the state of the art will be. We can make that change. Everyone is
focusing on this first chip. That's not the point. The point is, for the next decade we
have a technique that's going to let us move faster than anyone else does. It's a
cliché right now, but we can work on Internet time.”
|
