Fujitsu Draws Sparc64 Roadmap Past 2010
Published: February 23, 2006
by Timothy Prickett Morgan
With Intel and AMD in a bloody battle over control of the X64 processors used in most servers today and IBM making noise about high-performance Power chips and Sun Microsystems touting low-power Sparc T1 chips, it is hard sometimes to hear what Fujitsu, the chip designer and manufacturer in the Fujitsu-Siemens partnership, is up to. But as the fifth largest server maker in the world, what Fujitsu is doing matters.
This is particularly true for the many Solaris shops out there that have big batch jobs or relatively lightly threaded workloads and need a processor like the Sparc64 to run their applications. Fujitsu-Siemens has been a licensee of the Sparc architecture for as long as I can remember (it's so long ago that Google can't find it), and through 2005, Fujitsu-Siemens has sold well over 60,000 Sparc64-based PrimePower servers worldwide. This may not seem like a lot of servers in a world that consumes many millions of machines a year, but Fujitsu-Siemens, like IBM with Power, Sun with Sparc, and Hewlett-Packard with Itanium, is aiming its RISC box mostly at large enterprises that need the extra resiliency, memory bandwidth, I/O scalability, and SMP scalability that the PrimePower can deliver. Considering that the Fujitsu-Siemens clone Sparc line was basically unheard of outside of Japan and Germany before 2000, the PrimePower product line has come a long way. Fujitsu-Siemens emerged from its quiet marketing just as Sun's own UltraSparc-III product line was much delayed and its server prices were very high. Fujitsu-Siemens got into many Sun accounts by offering better scalability and better bang for the buck, and considering that the PrimePowers run Solaris--unlike IBM's AIX-based pSeries and p5 lines--it is surprising that Fujitsu-Siemens didn't eat a lot more business than it did at the big financial services, telecom, and service provider businesses that were using Sun iron in the early 2000s.
Fujitsu has a very long mainframe heritage, even before it invested and then eventually took over the Amdahl clone mainframe business, and it is this heritage that gives its Sparc64 implementation of the Sparc architecture one of its distinctive features. Fujitsu, like IBM, thinks like a mainframe vendor, which means it has very long development cycles and a view toward overall balanced system performance for very specific kinds of workloads. Fujitsu doesn't boast much about its chips or its future plans, unlike Sun, which was claiming back in 1997 that it would be able to lash together over 1,000 UltraSparc-III processors into a single system image. That WildCat memory interconnect never did quite pan out (except for a few supercomputer customers), and the UltraSparc-III chips were about 18 months late coming to market and significantly less powerful than IBM's Power4 processors. Sun was supposed to have its dual-core UltraSparc-IVs to market to meet IBM, and only now, five years later, is Sun starting to close the performance gap. By shifting to Opterons for entry machines--and soon, midrange machines--it has closed the gap. Sun's woes in delivering bang and bang for the buck for its Sun Fire server line are why, ultimately, Sun partnered with Fujitsu in June 2004 to bring a jointly developed Sparc product line, tentatively called the Advanced Product line, or APL, to market in mid-2006.
Depending on who you ask and how long your memory is, the APL products are right on time or running a little bit late. But some of us have long memories, and we also know that mid-2006 is not the same as the second half of 2006, which is when Sun and Fujitsu-Siemens now aim to get the APL products first into the field. That is not a huge slippage for either Sun or Fujitsu, largely because the Sun Fire and PrimePower machines they currently sell are very powerful machines by any measure. With PrimePowers scaling to 128 processors and Sun able to deliver up to 72 dual-core UltraSparc-IV+ processors, this is a lot of performance. On many workloads, IBM's 32-way dual-core Power5 and Power5+ machines can best these two Sparc alternatives, and on many, the Itanium-based, 64-way Integrity machines can meet or beat them, too.
What is more important to a lot of customers who are investing in big iron machines is a roadmap that shows a vendor is committed to a platform for a long time. The large enterprises who buy big Unix servers or mainframes don't just buy machines, they marry them. (They will cheat on them with X86 servers wherever possible, of course. . . . ) Customers, of course, get to see lots of processor and server roadmaps, but only occasionally do vendors let the public and the press see them. And with good reason.
With Sun talking about how great its "Niagara" Sparc T1 processors are--which pack eight four-core Sparc processors onto a single chip that burns only 70 watts of juice at 1.2 GHz--and talking up its future "Rock" high-end chips, which will be enterprise-class processors along the same lines as the T1s, Fujitsu-Siemens has a vested interest in talking about its plans and how it will differentiate itself.
Among those of us who speak English and American, Richard McCormack, senior vice president of marketing for Fujitsu Computer Systems, the Fujitsu North American arm, gets the job of telling us about Sparc64 and PrimePower roadmaps. And while Sun is only planning on picking up the APL product line for a few years until its Rock systems get to market and its Niagara machines mature, Fujitsu-Siemens has plenty of plans of its own.
Because the Sparc64-VI processors have two cores--each with two threads--the future APL machines should be able to deliver a little more than twice the performance of single-core "Zeus" Sparc64 V processors running at 1.3 GHz, which first started shipping at the end of 2004 and which used a 130 nanometer process. The Zeus chips had either 1 MB or 2 MB of L2 cache on chip. The current "Olympus-B" Sparc64 V+ processors use a 90 nanometer process and run at between 1.82 and 2.16 GHz, including 3 MB or 4MB of L2 cache. With the "Olympus-C" Sparc64 VI processors, Fujitsu is moving to two cores, and as we now know, is also adding multithreading to each core. According to early benchmark tests, the multithreading support boosts performance by about 20 percent on thread-sensitive workloads like Java. These Olympus-C processors use the same 90 nanometer process as the Sparc64 V+ chips, so moving to two cores nearly doubles electricity usage from 65 watts to 120 watts. The Olympus-C chips will have up to 6 MB of L2 cache on chip and have a target speed of 2.4 GHz. Looking ahead, Fujitsu is planning the "Jupiter" Sparc64 VI+ processors running at greater than 2.7 GHz and having four cores with two threads per core. These Jupiter processors will presumably plug into the Jupiter frames, and because they are based on a 65 nanometer process that Fujitsu put into production for other chips it makes in Japan in late 2005, will run a lot cooler than the 90 nanometer process. Fujitsu has not said how much cache will be on these Sparc64 VI+ processors, but the odds favor it being a lot larger than 6 MB.
The future "Jupiter" server frames, which will use the Olympus-B and Jupiter chips, have a flat SMP architecture, like the current PrimePowers. But the Jupiter frames have been re-architected for dual-core and quad-core processors and have direct connections from the memory and I/O into the processors. With the faster clock speed and multithreading of the Sparc64 VI chips, the new cache architecture, larger on-chip caches, and a better bus design, Fujitsu-Siemens is pretty excited about how well the APL machines it is designing will perform. "The expectations are very high," says McCormack. "And regardless of what anyone says, we are on track for shipping the machines in the second half of 2006."
As both companies have said in the past, the APL product line will include low-end servers based on Sparc T1 processors from Sun as well as midrange and high-end machines based on the Olympus-C Sparc64 VI processor created by Fujitsu. More precisely, says McCormack, the Olympus-C processors will be used in Jupiter frames that span from four to 64 sockets. The current PrimePower machines span from two to 128 sockets, and there has been considerable speculation about how far the Jupiter frames would scale up and down. So, now we know.
Looking out even further, Fujitsu has Sparc64 plans that extend beyond the end of the decade. "We have Sparc64 roadmaps that go out beyond the end of the decade," says McCormack. Fujitsu has announced that it intends to develop a 45 nanometer chip-making process, but has not yet built a factory to support it. Doing that should allow Fujitsu to either stick with four cores and crank up the clocks on the future Sparc64 VII processors or add more cores--whichever makes sense.
If you look at Fujitsu processor roadmaps from 2003, you can see that the Sparc64 VII chips were supposed to be the first quad-core chips and they had a target clock speed of 5 GHz to 6 GHz. But, then again, they were projected to come to market in early 2006. Which is now. As you can see, old processor roadmaps--from Fujitsu or otherwise--are like the early cartography of the New World. Things get distorted. On that same roadmap, the dual-core Sparc64 VI processors were expected in mid-2004 running at 2 GHz and ramping to 3 GHz, and the Sparc64 V processors were never intended to go beyond the 130 nanometer process or 1.62 GHz in clock speed, either. But, chip happens. Making processors and the processes that make them is one of the most complex things humanity has ever done.
As we previously reported, the single-core Sparc64 V+ processors were upgraded to a top speed of 2.16 GHz only a few weeks ago--the last speed bump these chips will see. McCormack doesn't want customers to get nervous, though. "We do not envision the end of life of the existing PrimePower line any time soon, and the PrimePower and APL machines will co-exist," he says.
As for the prospect of using Sun's Rock processors in Fujitsu machines, McCormack is skeptical about the possibilities. "We need to get more clarity on what Rock will be before we can think about it," he explains. "We know exactly what we are doing with Sparc64."
Fujitsu-Siemens Keeps Rolling on Sparc64, Itanium Roadmaps
Sun Reveals Details on APL Partnership with Fujitsu
Sun Allies with Fujitsu for Future Joint Sparc Platform