Chips are always harder to design than roadmaps
Processor power, complexity grow as Sun, Intel lay out their plans
"More power" is the rallying cry for both Sun's and Intel's microprocessor teams for the next four or five years. Intel's roadmap includes a new IA-64 architecture led by the much-anticipated, but now delayed, Merced, while Sun is keeping to a slow, steady, incremental development strategy. New processors from both giants will battle for the lucrative high-end workstation and enterprise server markets, while Compaq's Alpha RISC processor hopes to take the spoils. (3,000 words, including one sidebar)
un and Intel have both released roadmaps teasing their customers with powerful chips planned out past the year 2000. At October's Microprocessor Forum, IBM and Compaq also took the wraps off some of their next-generation processors. Visions of the future are simply faster and more powerful processors, which means Sun will have a fight on its hands to maintain dominance in the enterprise server and workstation markets.
"No amount of performance is enough," said Anant Agrawal, vice president of engineering at Sun Microsystems' microelectronics division. "Any amount of performance you put on the desktop or in the enterprise -- people will find a way to use it. Compare what we have on the desktop with supercomputers in the past. Five years ago, did we think we would have a supercomputer level of performance on the desktop and there would be applications to use it?"
Indeed, the high-performance Alpha processor is coming back strong following Compaq's purchase of Digital. Intel also has revealed plans for 64-bit computing with Merced and its follow-ons using a new architecture. Everyone is emphasizing highly parallel multiprocessing in their designs for even bigger and more powerful servers.
Intel in particular has said it intends to use its IA-64 architecture to strike at enterprise-level servers and high-end workstations -- a traditional stronghold for Sun.
Sun, however, isn't standing still either. It announced its own processor roadmap in August, describing its intention to retain its processor marketshare and continue expanding outward, especially in the desktop embedded markets. The next step will be the UltraSPARC III processor, expected to ship next year. The new UltraSPARC will feature a 600-MHz clock and scalability up to 1,000 processors in a single system. (See last year's article on the UltraSPARC III in SunWorld.) By 2002, the planned UltraSPARC V will have a clock speed of 1.5 gigahertz (GHz).
Sun: on the beaten path
The Intel and Sun roadmaps illustrate the two companies' very different approaches to their markets. Performance targets are comparable, and the underlying silicon technology is the same, but everything else, from the architecture to the strategy, is different.
Sun's roadmap emphasizes steady, consistent growth with proven architecture.
"The key to success is coming out with a solid strategy and not deviating from it," said Agrawal. "We feel confident with the product strategy we have come up with. We're solidly behind it and willing to tell the world."
The odd-numbered UltraSPARC releases will feature major architectural upgrades, while the even-numbered UltraSPARCs get their performance improvements from such things as shrinking feature size (which makes for faster processing), improved compiler technology, and relatively few architectural additions.
Sun will continue to leverage the UltraSPARC by using the same basic core in three different processor lines. At the high end, the "S-series" will stress enterprise computing adapted for multiprocessing. In the middle will be the "I-series," a desktop line without the emphasis on SMP, but with the computing speed and graphics capabilities needed to drive modern workstations. Finally, the "E-series" will be designed for embedded applications, foregoing many of the bells and whistles included on the other two lines. It will also be highly customizable, enabling OEMs to add features needed for different roles -- from printers to autopilots to the basic chip.
The S-series, with 450-MHz UltraSPARCIIs,a 50-MHz improvement over the the just-announced 400MHz UltraSPARCII (see announcement on both processors in this issue of SunWorld), is scheduled to arrive next year. The UltraSPARC IIIs, in 600- and 700-MHz versions, will appear in the late 1999 to 2000 timeframe. The UltraSPARC IVs will be shipped in a 1-GHz version about the middle of 2000, and in early 2002, the distant 1.5 GHz UltraSPARC V will follow.
The I-series includes the new 360MHz UltraSPARC-IIi (see announcement in this issue of SunWorld) with 400-, 440- and 480-MHz versions of the UltraSPARC III shipping next year, and 600- and 700-MHz versions of the UltraSPARC III in 2000 and 2001. Also scheduled to arrive in 2001 is the UltraSPARC IV-based I-series.
The E-series UltraSPARC processors are planned in 300- MHz (1999), 400-MHz (early 2000), and 500-MHz (late 2000) versions.
Sun's roadmap for the S, I, and E-series
Intel, by contrast, is making a break with its x86 architecture and taking a new approach to the way it divides up the processor business. Intel will maintain compatibility with the x86, and continue to make 32-bit parts for at least the next several years, though clearly Merced marks a new fork in Intel's roadmap.
Both Intel's 32- and 64-bit processors will be based on the IA architecture. The IA-32 chips are continuations of the Pentium line, 32-bit processors designed for desktops and low-end servers. The IA-64 processors are 64-bit monsters designed to dominate the high end of the processor business, especially enterprise-class servers. To further confuse the issue, in about 2002 Intel plans to split its IA-64 family into two lines, with presently undisclosed processors aimed at the pure performance market (code named Madison) and the cost-performance market (code named Deerfield).
The IA-32 family will begin later this year or early 1999 with the Tanner processor. It will be followed in about the year 2000 by Cascade and then by Foster in about 2001. The IA-64 family will kick off in 2000 with Merced, to be followed in late 2001 or 2002 by McKinley and then the Madison-Deerfield split. Each of these will be significant redesigns, much like the 486 was to the Pentium, or at least the Pentium to the Pentium II. Intel has yet to specify clock speeds planned for these models.
The hard fact is that Sun doesn't have the money Intel has for processor development. But it does have plenty of market share to lose if a competitor like IBM or Alpha hits the jackpot.
One of Sun's problems is that the price of poker has gone way up. When Sun launched the SPARC series of RISC processors back in the mid-1980s, RISC chips were much easier to design. In fact, you could do the beginnings of a commercially competitive RISC chip as a course project in a university computer science class -- which is where the original MIPS design came from. RISC was simple, clean, and hence inexpensive to design.
doesn't have the money Intel has
for processor development, but
it does have the marketshare
The demands of increasing performance have made RISC chips a lot less simple and not nearly as clean. Where once you could be competitive with a lot of registers and a carefully chosen instruction set, today's RISC chips have to execute more than one instruction at a time and run at speeds that turn them into tiny radio stations.
"People said RISC should be easier to implement and get more performance," Agrawal said. "As we try to squeeze in more performance, the implementations become more complex. I would still contend that that the complexity is less in RISC than in x86 [architecture], but you don't see the difference you saw 10 years ago."
Finally, as the feature size has shrunk from about 1.5 microns down toward 0.15 microns, the cost of building chip factories has skyrocketed. A billion dollars is about the minimum for a competitive chip factory using the latest processes, two billion dollars isn't unheard of, and some people are predicting three billion dollar factories by the turn of the century. Agrawal expects as feature sizes drop below 0.1 microns, there will be only a few superfabs left capable of producing the latest-generation chips.
Unlike Intel, however, Sun does not actually make chips. Sun's UltraSPARC processors are made by Texas Instruments, which means a large part of the cost of shrinking the feature size on UltraSPARC is carried by TI's other chip businesses, including its advanced line of digital signal processors (DSPs).
In any case, the money is only half the battle. The other half is that Sun must not merely equal the performance of Intel's processors, but exceed them. One misstep and Sun's market share could evaporate in a matter of months -- never to be seen again.
one of Intel's blunders, which could
mean another opportunity for SPARC
to gain ground.
This has happened to Intel competitors before, such as Motorola with its 68000 series of processors back in the 1980s. Motorola remained a serious competitor with Intel as long as the latest 68000 processor kept pace or exceeded Intel's offering. When Motorola faltered in the mid-1980s and couldn't deliver its next generation on time, however, Intel pulled inexorably ahead. The lag in getting the latest Motorola processor was one of the things that drove Sun to develop the SPARC family.
Then again, Intel also makes mistakes. Merced, which could shape up to be one of Intel's blunders, may represent another opportunity for SPARC to gain ground.
The mediocre Merced
"Initially, Merced sounded like it was supposed to come out and deliver industry leading performance," said Lindley Gwennap, the vice president for publications at MicroDesign Resources and a leading analyst in the microprocessor field. "Now they're saying it's not going to be any faster than their own x86 chip, Foster, much less some of the RISC chips that might be out by then."
Instead, Gwennap said, Intel officials are now talking about the follow-on to Merced, McKinley, as the chip that will be the breakthrough for the company. Part of the change in tune is that Merced will ship nearly a year late. That gives Sun and other competitors more time to catch up.
"A big part of that is the fact that Merced slipped nine months," Gwennap said. "Originally the part was due out in 1999 against different competition. Now it is coming out in mid-2000, and it will be in competition against much stronger parts."
Gwennap is a cynic when it comes to technology roadmaps and their predictions for performance, relying instead on manufacturers' track records.
"Alpha has been setting the standard for five years and I expect them to continue to deliver high performance," he said. "Sun has been lagging in performance, and I don't see any magic bullet on the horizon to bring them up to the place where Alpha is. Alpha will probably continue to be high performance, and Sun will continue to increase performance, but lag behind Alpha...It's looking like Merced won't be that great. It will reduce the gap [between Intel, Sun and Alpha], but it may not get to the Alpha level."
The real question, said Gwennap, is what IA-64 will do (see sidebar). "McKinley is looking pretty good when it comes out in late 2001. That will be Intel's real chance to match or exceed Alpha's performance."
The others: Compaq and IBM
Alpha's performance record is strong, in stark contrast to its marketing record under Digital. Alpha has suffered from the same marketing malaise as the rest of Digital's product line and as a result never achieved the presence its supporters hoped for. Compaq is now faced with reviving Alpha's disproportionate market share, and has firmly committed to using the Alpha in its high-end server products.
Compaq has assured its Alpha customers that it will continue to back the processor and has announced new servers from its Tandem division that use Alpha chips in place of the MIPS RISC chips used in earlier versions. Meanwhile, Compaq has not one, but two development teams (one in Silicon Valley and one in Massachusetts) working full-blast on the next two versions of the Alpha.
IBM and HP are also still in the race. At the recent Microprocessor Forum, IBM announced its "GigaProcessor" -- a new design built around a PowerPC core and designed for high-end servers. Like most of the others, IBM includes L2 cache and a full memory and I/O subsystem on the chip. IBM's plan not only includes multiple iterations of the POWER3 on the chip, it also adds a high-speed interface to make it easier to tie multiple chips together into multiprocessing architectures.
"Does any of this matter?" Gwennap asks. "Sun has never had the fastest processors, but they sure have a lot of customers." It may not matter much against the Alpha, but it does in the race with Intel against Intel.
If Sun cannot exceed the performance of the latest Intel chip by a reasonable margin, Sun will probably lose ground rapidly -- in software as well as hardware.
Faster and faster clock speeds have become critical to drawing customers, but present several technical problems. For example, higher speed is predicated on shrinking feature size and is therefore tied to improvements in fundamental process technology.
The next obvious approach to improving performance is longer word length. The top-of-the-line new generation processors are all 64-bit chips, but longer word length mostly improves shuttling data around. What's bottle necking processors now is execution latency.
This brings us to caching strategy. The conventional strategy for 64-bit processors is to use large caches as part of the processor (Level 1), and even larger caches tightly tied to it (Level 2). Here again, Intel is being extremely aggressive.
The Merced will have three different caches, two of them on the same piece of silicon as the processor. The Level 0 caches will be fairly small, blazingly fast and associated with each execution unit on the chip. The Level 1 cache, also on the chip, will feed the processor as a whole. The Level 2 cache will be on a separate chip and tied to the processor by a very fast bus.
Many of the new generation chips are seeking the biggest advantages in parallelism: doing more than one thing at a time. Intel, for one, will offer new chips featuring multiple execution units on the same piece of silicon, which lets the processors execute multiple instructions per clock cycle much more efficiently -- if the rest of the system can feed the instructions to the execution unit fast enough.
Intel is relying strongly on the techniques of speculation (loading the instructions and data which are most likely to be used in the processor, as early as possible), and predication (removing code branches that probably won't be executed) in the IA-64 architecture. In fact, there's a sense in which these techniques are at the heart of the design. The elaborate cache system and multiple execution units in Merced don't make any sense except as support for a highly massaged, parallelized instruction stream.
and multiple execution units in Merced
don't make any sense except as support
for highly massaged, parallelized
"The performance of a microprocessor is really a combination of how much work your processor is capable of doing and how much can be exploited in [the compiler]," said Agrawal. "You can always make tradeoffs that are at the extreme end. For example, doing all interlocks and scheduling in hardware makes the processor very complex. Or, like Intel in Merced, you can rely on the compiler solely to do all the scheduling. The whole thing is about making the right tradeoffs so you can hit the sweet spot of the business you're trying to address."
Exploiting these parallel features requires intelligent compilers that can analyze the code and arrange it so the hardware features can have their best effect. This means that compiler design becomes at least as important as processor design. It is here that Sun's dual identity as a hardware and software company can pay dividends. Intel is certainly no slouch when it comes to compiler design either, and Hewlett-Packard has a lot of experience on advanced intelligent compilers with its Precision Architecture. However as an operating system company and a processor company, Sun is at least their equal in designing compilers and, some would say, their superior.
There's another consideration here as well. Unlike goosing the clock speed, predication and speculation, and all the rest of these techniques rely on having opportunities in the software to apply them. That means that programs are likely to vary widely in how much benefit they will see from these features. A program that shuttles a lot of data around and uses a lot of loops, such as a database application or a graphics program, can see major benefits. Software that has few critical loops and is compute intensive rather than data intensive will benefit much less. At worst, these elaborate features can be nothing more than expensive overhead.
As a result, software vendors will have to consider recompiling their software to run on Intel's IA-64 chips. Staying with their 32-bit versions will mean users may realize few benefits from Merced and subsequent processors.
What it all means for Sun
Pure processor performance has never been Sun's thing. As Gwennap points out, historically Sun has lagged behind the Alpha in raw power even as it has led Intel. But Sun has been good at tailoring its SPARC chips to the needs of its customers.
IA-64 means Sun will have to fight even harder for hardware business, especially in the high-end workstation and midrange server market. Even if Merced doesn't walk on water, it can at least swim, and that's enough to put the pressure on Sun.
As things look now, Sun will get some breathing space from Intel until the real competition comes from McKinley in late 2001 or 2002. The new Alpha processors and the Windows NT servers built on McKinley will have a more immediate impact, the extent of which will be difficult to predict until Alpha`s future becomes more clear. If Compaq and other vendors can really market the Alpha, it could become a significant competition for Sun.
Sun has a major competitive advantage, ironically, with Windows NT. Even in version 5.0, Windows NT is well behind Sun and other major Unix vendors in offering the features needed to handle enterprise-wide computing. What's more, there are serious questions as to whether Windows NT can catch up even by 2002.
Meanwhile, Sun is hedging its operating system bets. It has announced that it will develop a version of Solaris for the IA-64 architecture. That means that customers won't have to choose between Solaris and Windows NT when they choose a server architecture.
The big question for Sun is how well it can deliver on its announced strategy. If it is going to stay competitive in the hardware business, it has to produce chips that will stay one step ahead of Intel. Roadmaps are one thing, but in this industry very little ever goes as planned.
About the author
Rick Cook is based in Phoenix, AZ, and divides his time between writing science fiction fantasy novels ("full of bad computer jokes") and writing about high technology. Reach Rick at firstname.lastname@example.org.
If you have technical problems with this magazine, contact email@example.com
Merced, for example, features a mode bit to switch from IA-64 to x86 instructions, a clumsy arrangement which means tha applications will have to be recompiled to take advantage of Merced's features. Otherwise, software may actually run slower than on the IA-32 chips that will be available at the same time.
In short, handling the switch to 64-bit architecture is a bit like having a bear by the tail.
"When I talk to people, I keep hearing how much Merced is in disarray and the schedule is not in control," said MicroDesign Resources' Lindley Gwennap.
Intel itslef admits it underestimated the scope of the problem.
"The simplest way to look at this is that this is the first time we're [designing a processor] with this instruction set," said Fred Pollack, Intel's director of measurement architecture and planning. "It's not only a new instruction architecture, it's a new type of instruction set. Several things took us longer to do than we originally envisioned."
A new microprocessor is always a challenge (witness the problems Intel has had in getting its Xenon version of the Pentium II into full production). A new feature size is a bigger challenge, and a completely new architecture is a bigger challenge yet. Intel has already delayed Merced once, and while the company claims it is on track to meet its new 2000 introduction, many observers are taking a wait-and-see attitude.
Even the Year 2000 problem comes into play here. Most IT organizations are going to have their hands full next year and in 2000 just coping with their millennium bugs. That makes it a bad time to launch a new processor, especially the first in a new architecture aimed at the enterprise. There has been some speculation that Intel might be better off delaying the shipment of IA-64 products until 2001 when all the fuss over Y2K has died down.
The key event for Merced watchers will be the tape-out: the delivery of the design to the fab so Intel can start producing silicon. "Once they get tape-out, everyone will feel a lot better and breathe a sigh of relief," said Gwennap, "but they're four to six months from tape-out."
Still, too much may be being made of Intel's current problems. No one doubts that Intel will get a 64-bit processors out the door and that when the IA-64 architecture comes, it's going to be a very powerful piece of work. And if Merced is further delayed, Gwennap speculates that Intel might skip directly to McKinley, never really producing Merced at all.
But Merced or McKinley, 2000 or 2001, Intel's going to have a powerful 64-bit processor built on the IA-64 architecture.