Sun Pulls Plug on Future Millennium, Gemini Chips

by Timothy Prickett Morgan

The rumors have been flying around since late last week, and they are true. Sun Microsystems has killed the "Millennium" kicker to the UltraSparc-IV processor and its "Eagle" servers, as well as the dual-core "Gemini" chip for blade servers. While the press and analyst community is having a field day saying that the UltraSparc-V processor is dead, what is probably closer to the truth is that Sun killed those projects so it can get the "Niagara" and "Rock" multithreaded processors and their servers to market as quickly as possible.

It would not be entirely surprising to see the UltraSparc-V moniker slapped on either Niagara or Rock designs (or both), which is why anyone saying that the UltraSparc-IV is dead is probably not going to turn out to be true in the long run. For the past three years within Sun, the company's engineers have been pursuing two different future server architectures, said David Yen, executive vice president in charge of Sparc chip development, in an interview with us this week. Sun had two competing Sparc designs, and one of them had to go.

On the one side of Sun are those people working on Millennium processors and Eagle servers, and on the other are those who, in the wake of the acquisition of Afara Systems, which was building an eight-core chip based on stripped down UltraSparc-II processor cores, decided that cramming as many cores and threads onto a processor would be, in the long run, a better strategy for Sun to take. In a world where Sun has to cut 3,300 employees and trim its research and development costs, something has to give. And despite the fact that both Millennium and Gemini have taped out and Gemini is already in silicon, Sun stared down the amount of further investment it would have to make to productize these chips, and decided that it had better concentrate on the throughput computing designs.

The Millennium chip had a very aggressive pipeline compared to the UltraSparc-III core that is at the heart of the current "Cheetah" UltraSparc-III and "Jaguar" UltraSparc-IV processors, Yen explained. Sun's goal when it started the Millennium project nearly five years ago was to have a chip that was better than Intel's Itaniums and IBM's Power processors. Specifically, Millennium, which would have been productized as UltraSparc-V, would be able to switch between two modes: one optimized for multithreaded applications (as the multicore Niagara and Rock processors are, by their very design); and one optimized for the kind single-threaded applications that dominate database and ERP applications today.

As Yen explained almost two years ago, Itanium chips use sophisticated compilers to look for parallelism in applications, and compile them in such a way to make them run efficiently. Even the Power5 processors only have four threads per chip (two cores with two threads each). Sun wanted to have it both ways with the Eagle platforms. The Millennium processors were to be built using a 90 nanometer copper/low-k process from Texas Instruments, Sun's fab partner, and run at clock speeds in the range of 1.8 GHz to 3 GHz or higher. Yen said that, at the time, the ideas embodied in the Millennium chips and the Eagle platforms were the right ideas, but that they had been superceded by the ideas embodied in the "radically multithreaded" Niagara and Rock designs that Sun has been touting so hard in the past year.

"We want to invest more aggressively in these throughput computing approaches," Yen said. "It was time for us to make a decision. If we do not do a new breed of processors, we believe that our competitors will." He would not elaborate on how much money Sun spent on the Millennium chips and Eagle servers, but he said that with a radically different design, as Millennium encompassed, it was going to take a lot of money to get customers and software partners to use its new features. He also said that lots of the intellectual property developed for Millennium and Eagle will make it into the multithreaded designs, including memory subsystems and error correction electronics across the server.

As for Gemini, which was a dual core implementation of the "Blackbird" UltraSparc-II processor, this chip was in essence a first pass idea to test the throughput computing approach. However, at a 1 GHz to 1.2 GHz clock speed, Yen explained, it did not have a significant performance advantage--although a slight one on multithreaded applications--compared to the current "Jalapeno" UltraSparc-IIIi processor, which has a large L2 on the chip to boost performance and which runs at the same clock speed. With Opteron machines, Xeon machines, and two Sparc machines all competing for the same edge of the network space, even though Gemini was essentially finished, Sun cannot afford to productize it and doesn't believe it is wise to considering the overlap.

While Sun gave the Rock moniker to its future radical multithreaded chip in February, it has not said much about what it looks like and when it might be delivered. The Rock processors will follow the "Niagara" processors, which will be aimed at entry and some midrange servers. In February, Yen said that in the late 2005-early 2006 timeframe, TI would have a 65 nanometer process to create the "Niagara" processors, which will pack eight of the four-threaded simplified UltraSparc-II cores onto a single piece of silicon. In Sun's roadmap, relative performance is being reckoned against the oomph supplied by the "Hummingbird" UltraSparc-IIi processor running at 650 MHz. Sun has said that it expects that the Niagara chips will provide about 15 times the performance of this baseline processor. Sun has said that the third generation throughput computing chips--now called Rock--will deliver 30 times the performance of this baseline. But raw performance is not the only thing of value with these future CMT designs, according to Sun.

In the meantime, Yen said that Sun can crank out some more clock speed on the new Jaguar UltraSparc-IVs using the 130 nanometer process and that Sun will be able to push performance further with one or two iterations of a 90 nanometer copper/low-k/strained silicon process, creating the "Panther" UltraSparc-IV+ processor. The new UltraSparc-IV is a dual core processor, just like IBM's Power4 and Power5 and Hewlett-Packard's PA-8800 processors.

With the Panther chips, Sun is adding microarchitecture improvements to the cores to tweak performance, but the shrinking in size of the chip core by half (in terms of area) should allow the chip to hit 2 GHz to 2.5 GHz clock speeds, and eventually go higher. The Panther chips, Yen said, would also bring the outboard L2 cache in the UltraSparc-III and UltraSparc-IV chips onto the chip as well as an L3 cache controller. Sun will be adding an L3 cache to the UltraSparc-IV+ packaging, much as it now does with L2 caches on current UltraSparc processors. This is the first time Sun is using an L3 cache. Putting the caches on this chip lets them run faster, and allows them to be made with less capacity to deliver the same performance. By the way, UltraSparc-IV+ was already in silicon in February, which means that Sun can drop it into the market more quickly than it could the Millennium chips and their new Eagle frames.

Yen also said that the company intended to keep pushing the UltraSparc-IIIi processors, and that the move from 130 nanometer to 90 nanometer processes would allow it to boost clock speeds (again, my guess is by a factor of two) and to quadruple the amount of on-chip L2 cache.

"We believe the current progress on UltraSparc-IV and UltraSparc-IIIi can keep us going to a while and keep us competitive," Yen said, defending Sun's decision. "This results in a better focus for us." And citing the criticism Sun has already faced in the press, he said that if Sun was going to phase out UltraSparc and just go with Opteron, there would be an easier--and presumably cheaper--way of doing it. "What we offer now is a much cleaner roadmap, and the internal troops are excited about it because we will be able to pursue our throughput computing goals without resource constraints." He said that all of Sun's executives knew that this was a pretty big bet that Sun is making. "We believe this is our opportunity to lose if we do not execute well." To that end, Sun's chip teams and systems teams are now working more closely together to reduce the time to market on products. This mirrors a move rival IBM just made as it merged its Systems Group, which designs servers, and its Technology Group, which creates its Power processors.