In-Depth

Is Sun Microsystems on the Road to Recovery?

A major accord last month between Sun and Fujitsu could fundamentally shake up the high-end server space

• By Stephen Swoyer
• 07/06/2004

Analysts say the deal—which enables Sun to effectively outsource the production of its SPARC processor line—is a win-win for the Unix giant. It lets Sun focus on engineering a new generation of multi-core, multi-threaded processors, and gives the Unix giant an opportunity to dump its aging UltraSPARC architecture for Fujitsu’s scorching SPARC64 chips. (Sun and Fujitsu have separately developed and marketed chip designs based on the SPARC architecture for some time, and Fujitsu also distributes Sun’s Solaris operating environment with its PrimePower systems.)

According to the terms of the agreement that the two companies announced last month, however, Sun and Fujitsu will collaborate on a series of new products, code-named the Advanced Product Line (APL), which will replace their respective SunFire and Primepower product lines.

Sun and Fujitsu plan to merge their two product lines and begin selling APL systems based on Fujitsu’s forthcoming dual-core SPARC64 VI microprocessor, along with Sun’s Throughput Computing “Niagara” and “Rock” multi-core, multi-threaded processors.

Even though Sun must share the fruits of its vaunted Throughput Computing engineering effort with Fujitsu, analysts say that the deal nevertheless amounts to a slam dunk. For starters, notes Gordon Haff, a senior analyst with consultancy Illuminata, Sun gets access to a Fujitsu SPARC64 architecture that, in recent years, especially, had raced past UltraSPARC in terms of both clock speed and raw performance. More important, notes Haff, Sun no longer has to finance two full-fledged microprocessor development efforts.

“I think one of the big questions around the Throughput Computing initiative … all along has been does this mean they now have to effectively have two processor lines, because—particularly in the near term—a Throughput Computing-type approach isn’t going to be ideal for every workload,” he observes. “This deal basically lets them concentrate on their throughput computing while Fujitsu continues to produce processors that will be produced in systems that will be a good upgrade path for sort of the current customers of both companies.”

Nathan Brookwood, a principal analyst with microprocessor consultancy Insight64, agrees. “What the processor designers at Sun are working on now is either the low-end throughput chip which they call Niagara, or the high-end throughput chip which they call Rock,” he explains. “They’re effectively done with UltraSPARC. Sun has cancelled the [UltraSPARC] V, but [these engineers are] also doing minor enhancements to the UltraSPARC III or IV architecture to keep it competitive for the next few years until the Niagara and Rocks arrive.”

What exactly does Sun have in store with Niagara and Rock? InSight64’s Brookwood describes Throughput Computing as nothing short of a radical departure from current trends in microprocessor design. The risks, he says, are salient, but the payoff could shake the industry.

“Even Sun would characterize it as a radical approach—a revolutionary approach. Throughput Computing is going to provide, on the low-end, eight cores with four threads each, so it looks like a 32-way system. They haven’t said yet how many cores are going to be in the high-end Rock processor, but I think ultimately it’s going to be even more impressive,” he notes.

Before either Niagara or Rock become industry-shakers, however, Sun must solve a host of vexing problems. “I think other people are going to be somewhat skeptical about whether you can really keep 32 threads on a single chip going with any kind of real performance, because the I/O requirements and the memory bandwidth requirements are rather staggering,” he says. “If it works, it’s going to be a very impressive chip.”

Industry heavyweights such as IBM Corp. and Intel Corp. have by and large embraced multi-core chip designs—IBM’s Power4 processor was the first dual-core microprocessor on the market, while Intel has multi-core designs of its own in the works—but neither vendor plans on taking the concept to so radical an extreme as Sun. “It’s the direction that everybody is going in to some degree or other, although Sun is perhaps looking at somewhat more radical Throughput Computing-type designs than others are,” says Illiminata’s Haff. “Intel is going multi-core, IBM is multi-core, and in the near-term, I think they’re both going to have fatter cores than Sun is looking to have.”

Both IBM and Intel have also experimented to some extent with multi-threaded designs: Intel was first out of the gate a couple of years ago with its HyperThreading (HT) technology, while IBM has developed a technology called Simultaneous Multi-Threading (SMT) for its Power5 chip, which contains eight logical chip cores. But no vendor has placed so great an emphasis on both multi-core and multi-threaded chip design as Sun.

“I think Sun has taken a more radical approach, and with the Rock was willing to shrink down the features of any individual core in order to fit more cores on a chip, and I haven’t seen anything from IBM that would suggest a similar kind of thinking, so unless IBM’s roadmap has changed lately, IBM is going for fully-featured cores with just a few threads each,” he comments.

In fact, says Brookwood, Intel actually plans to de-emphasize multi-threading in future iterations of its desktop and server chips. “In talking with Intel about their plans for multi-core products, they’ve gone on the record saying next year they’re going to have multi-core desktop and server products. They don’t seem quite so intent on enabling HyperThreading in their products,” he indicates. “Their view is that multi-core is a form of HyperThreading, so they may very well just pursue one thread per processor, two cores per chip.”

So what kinds of applications can best exploit so extreme a combination of multi-core and multi-threaded processing horsepower? “Essentially, it’s a large multiprocessor on a chip, so you know, really, any application that you see running in very clustered environments, or general applications designed to run across large multiprocessor configurations, are the types of things that could benefit from Throughput Computing,” says Illuminata’s Haff.

In the near-term, however, Fujitsu will carry most of Sun’s water. Later this year, the Japanese computing giant’s SPARC64 VI will arrive in the form of a dual-core chip that’s capable of hitting 3.0 GHz. This kind of horsepower—coupled with Fujitsu’s emphasis on mainframe-like capabilities such as “Instruction Retry”—could once again make Sun a contender in accounts (or for applications) where the aging UltraSPARC architecture had been written-off.

Given the notorious delays that attended the releases of Sun’s UltraSPARC III and even UltraSPARC IV chips, there’s some speculation that both Niagara (which is promised for early 2006) and Rock (promised for late 2006) will not ship on time. InSight64’s Brookwood doesn’t buy that.

“They had a huge [number] of problems with UltraSPARC III, no denying [it]. Even they would agree with that. But on UltraSPARC IV, I don’t think that was too late compared with what they had originally expected,” he argues. “The people who brought you UltraSPARC III on a delayed schedule are no longer in charge, and the people who finally got UltraSPARC IV out the door are in charge of everything, so I’m not ready to believe that there’s going to be a delay.”