Q&A: Sun gambles big on future chip design

Sun’s recent surprise decision to drop work on its UltraSparc V processor could be seen as either a desperate cost-cutting measure by a troubled company, or a vote of confidence for the next generation of multithreaded, multicore “throughput computing” processors that Sun has been readying since its 2002 acquisition of Afara WebSystems.

Sun’s recent surprise decision to drop work on its UltraSparc V processor could be seen as either a desperate cost-cutting measure by a troubled company, or a vote of confidence for the next generation of multithreaded, multicore “throughput computing” processors that Sun has been readying since its 2002 acquisition of Afara WebSystems.

The demise of UltraSparc V appears to have been hastened by Sun’s recent decision to lay off 3,300 employees — about 9% of its workforce — including some engineers working on the processor. But according to David Yen, Sun’s executive vice president of processor and network products, the product’s termination is an indication that Sun is betting everything on its throughput computing designs. UltraSparc V, code-named Millennium, was intended to be a stop-gap offering as Sun readied its first throughput computing processors, the first of which — a network-intensive chip code-named Niagara — is expected in early 2006.

IDG News Service recently interviewed Yen, asking him to explain the decision to drop both the UltraSparc V and Sun’s first dual-core processor, code named Gemini, which the company had been on the verge of shipping.

With the end of UltraSparc V, Sun seems ready to consider a world of new possibilities, including throughput chips built with Advance Micro Devices’ (AMD) processor cores, and even an end to the UltraSparc brand itself, according to Yen.

How did you arrive at the decision to terminate Gemini and UltraSparc V?

Well, the quick comment is: People should have no doubt about our belief and our vision in the throughput computing area. We believe in it so much that we wanted to focus all our resources on trying to expedite its development. We did spend quite some time working on Millennium and Gemini. The two processors look a little bit on the traditional side. We believe the new CMT (chip multithreading) processors have such great promise that we really would like to maximize the amount of resources we can throw behind them.

There was really nothing wrong with those processors. We actually taped out both of them, and Gemini even reached the point where the chip was fully working. But in the Gemini space, we have UltraSparc IIIi and UltraSparc IIIi+ that are doing a very capable job. By not doing Gemini, it also helps a little bit in the product positioning.

In the Millennium space, with the current UltraSparc IV, followed by the UltraSparc IV+, and then with the upcoming Rock and Niagara systems, we actually believe that this is probably a better road map. That’s why we made the decision.

So were the performance improvements with UltraSparc V and Gemini not what you’d anticipated?

These are all first generation, so to speak. Gemini has two UltraSparc II cores. We already have the chip working, so we even have benchmark results (showing that the) single-thread performance is very competitive, to some extent — close to what the UltraSparc IIIi offers. In the multi-thread, throughput type of applications, in certain cases, because of the dual core, it actually performs better.

But considering the amount of investment in terms of productization for both processor-based systems, and considering the delta, we decided to have a more clear positioning. And that’s why we didn’t go that way.

UltraSparc V, or Millennium, started even earlier, because it is a very big, very sophisticated processor design. But at this point, even though we’d pretty much completed phase one — namely we finished the design and we taped out the chip — when we were facing the task of going forward, and then we looked at the promise from the more aggressive throughput computing, the chip multithreading type of concept, we believe the kind of performance that UltraSparc V was expected to achieve could be better achieved with a different approach.

Does the decision to cancel Gemini and UltraSparc V move up the ship dates for Rock and Niagara?

It definitely helps, because we are moving a significant number of people who up to now had been working on UltraSparc V to work on the Rock and Niagara family.

When do you now expect those two to be out?

We have said Niagara 1 — that is our first Niagara chip — will happen probably at the very beginning of 2006, and the subsequent members of the Niagara family, and the Rock chip, will follow that.

But you said that last year. How has the road map changed?

It certainly will secure the schedule, if not help to move it up earlier.

How many engineers are you moving from the UltraSparc V and Gemini projects to your throughput computing chips?

We are moving more than a couple of hundred engineers from those projects to work on the new processor projects. By doing that, we don’t have to hire yet additional engineers in the upcoming fiscal year.

What does this mean for the future of UltraSparc, then? Is UltraSparc as we know it dead now, or will there be future UltraSparc processors?

All of these throughput computing processors are Sparc-compatible. Whether we will continue using the UltraSparc name or not, that’s a separate decision. But these are every bit Sparc processors.

Yes, we did cancel a project called Millennium. That is the one that originally we planned to label as UltraSparc V when it comes out. And now, without doing that processor, we’re doing these newer, throughput computing Sparc processors. When one of them comes out, we may label that one as UltraSparc V to continue the sequencing, if that’s still the way we want to name them.

You have this relationship with AMD now. Would you consider using another type of core architecture as you design new multicore processors?

We will be working closely with AMD with Opteron-based Sun systems. However, please understand that we have a more than $127 billion installed base there. It is Sun’s obligation to maintain binary compatibility. This is one contract Sun considers very seriously.

When Sun uses throughput computing, at least at this moment, it is 100% Sparc V9 compatible (Sparc V9 is the architecture on which UltraSparc chips are based). But you are right, the whole innovation does not necessarily tie to Sparc. In fact, that’s why we believe our competitors are heading in that direction.

So do you expect to base throughput computing processors or systems on AMD technology eventually?

It could happen but, in this particular case it involves AMD, therefore until we are ready and we both agree, I cannot comment on that.

It seems like a very risky move to put all your resources behind throughput computing. What convinced you that it was worth the risk?

This is not the first time we’ve done this type of thing. In about 1994, there was this prevailing belief that SMP (symmetric multiprocessing) was dying, and everybody was advocating NUMA — the non-uniform memory access architecture — and yet at the time Sun engineering believed that there was still so much basic engineering that we could do (with SMP). We continued the SMP vision because we believed that it was the right architecture, the most friendly, and easier for the programmers. So in 1996, when our competitors all went on to this NUMA type of architecture, we had no competition in the marketplace. We enjoyed, literally, a six- to nine-month vacuum in the marketplace. That was the moment that Sun’s server foundation was really established. One just has to believe in what one believes.

So is there risk? You bet there’s a risk. Every time a new innovation is proposed, it’s always risky because it’s different and initially it’s unproven. But that’s the kind of a life that people in high tech live. The important thing is, do you believe that you’re right, and are you good enough to deserve that kind of self-confidence? And we believe we are.