Sun Reveals Details on APL Partnership with Fujitsu
by Timothy Prickett Morgan
If it seemed like the Sparc partnership between Sun Microsystems and Fujitsu was hastily thrown together when it was announced on June 1, it wasn't. Sun executives said last week that the Advanced Product Line partnership has been in the works for more than a year, and that June 1 was just when the lawyers got done arguing. This partnership makes sense, but the abrupt manner in which Sun told the world about it left a lot of questions unanswered.
Last week, Andy Ingram, vice president of Sun's Scalable Systems Group, gave some insight into how the extended Sun-Fujitsu partnership will work, and also gave a sneak peek at the future Advanced Product Line systems.
Sun has seen its business shrink drastically in the past three years, and is under intense pressure not only to make some profits but also to show that it has staying power in the IT market. Partnering with Fujitsu, which resells its PrimePower Sparc64-based servers against Sun's Sun Fire UltraSparc-based servers, to create a single line of machines that Fujitsu, Siemens, and Sun peddle around the world makes good sense, and is something that many in the IT industry have been calling for since Sun's fortunes started to shrink. (I wrote my first appeal to Sun and Fujitsu to do this in January 2001.) Both companies have executives with pretty big egos, which undoubtedly kept such an alliance from happening. About a year ago, some faction within Sun wanted to shake things up by adopting the Opteron processor for entry machines, by pushing Solaris on X86, and by partnering with Fujitsu for midrange and high-end Sparc systems. Another faction wanted to pursue throughput computing designs, like the future "Niagara" processors, which have just taped out and come back in first silicon, and "Rock" processors, which are due in 2008. (This is the first time that Sun has nailed a date down for Rock.) Yet another wanted to build giant monolithic servers based on the "Millennium" processor, which would have been sold as the UltraSparc-V but was killed off in April. We now know who won and lost those arguments, and as soon as the Sun and Fujitsu lawyers finished their work at the end of May, Sun's chairman and CEO, Scott McNealy, and Ingram put out the word. Sun killed off Millennium because it knew well before that point that it would be using a variant of Fujitsu's Sparc64-VI processors, and indeed whole machines that would have been sold as PrimePowers.
The alliance between the two companies only covers Sparc/Solaris servers. Both Sun and Fujitsu (which means the Fujitsu-Siemens partnership) are pursuing their own agendas in the X86 space. Sun has thrown its weight behind Advanced Micro Devices' Opteron chip, while Fujitsu and Siemens are selling a line of Primergy servers based on Xeon and Itanium processors from Intel.
Ingram reiterated the reasons why Sun opted for the Advanced Product Line partnership: it allows Sun and Fujitsu to demonstrate that the Sparc/Solaris platform has staying power, it allows Sun to focus on the riskier (but potentially more profitable) Niagara and Rock systems, and it helps to stabilize the Sparc/Solaris ecosystem. But the partnership does not mean that Sun and Fujitsu-Siemens will do all Sparc-related things the same. Both companies have their own storage alliances. For instance, Sun resells Hitachi disk arrays, while Fujitsu-Siemens pushes disks from EMC; both companies have their own sets of Solaris drivers, clustering software, service processor and systems management software.
What Sun and Fujitsu have agreed on, in essence, is to resell each other's current UltraSparc and Sparc64 equipment between now and when the Advanced Product Line systems are ready, sometime in 2006. Then Sun gets to design the low-end Advanced Product Line systems and the midrange chassis, while Fujitsu gets to design the high-end Advanced Product Line systems and the high-end chassis. The Advanced Product Line machines will, according to Ingram, be made in Sun's Oregon and Scotland factories and in Fujitsu's Japanese factories, and the idea is to make machines as close to the market they sell into, based on volumes, except the very biggest Advanced Product Line machines, which will probably be made only by Fujitsu in Japan. Each time one of these Advanced Product Line machines is sold, Sun and Fujitsu get money based on how much engineering and manufacturing they did.
While Ingram did not give a lot of details about the Advanced Product Line product line, he did raise the curtain a little. The future Sparc chip being designed primarily by Fujitsu to power the high-end Advanced Product Line machines is code-named "Olympus," and the system interconnection scheme for these SMP systems and their chassis are called "Jupiter." Ingram was cagey about what the Olympus processors might look like, but he said that now that Sun has two cores per processor, there is no way it is going back to a single-core chip. So we know that Olympus has at least two cores.
It is not a coincidence that Fujitsu-Siemens was readying its first dual-core Sparc64, the Sparc64-VI, for late 2005 or early 2006. The Sparc64-VI was to be implemented in a 90 nanometer copper/SOI process and made in Fujitsu's labs in Japan; it was to have 128 KB of on-chip L1 data/instruction cache per core, 6 MB of on-chip L2 cache memory shared by both cores, and an initial target speed of 2.4 GHz, with speeds eventually ranging between 2.1 GHz and 2.6 GHz. This is probably what Olympus will look like. Ingram said that the two companies were planning on systems that spanned to 64-way SMP, which implies that the Advanced Product Line machines will back off considerably from the 128-way SMP of its current single-core Sparc64-V line of PrimePower XA machines. Going beyond 32-way SMP is very inefficient, so this stands to reason, especially considering that the clock speed of the Sparc64 chips will effectively double between the Sparc64-V and Sparc64-VI generations. Very few companies need such large SMP boxes as the 128-way machines Fujitsu-Siemens has designed. There will be midrange Olympus boxes, perhaps with four or eight processors, spanning up to the high-end 64-way machines. The Olympus machines will be aimed at compute-intensive and single-threaded applications.
The other new bit of data from the chat with Ingram was the fact that the future Niagara machines being created by Sun are also being designated as Advanced Product Line boxes. The Niagara processor will cram eight processor cores, each with four threads, and memory subsystems all on a single chip. These Niagara machines are aimed at the infrastructure and modest database and Java workloads that make use of many threads. The Advanced Product Line agreement does not include the second generation of Niagara machines, nor does it include a kicker to the Olympus/Jupiter systems or the future Rock processors, which will have a lot more threads and performance than the Niagara designs. However, according to Ingram, there is nothing to preclude Sun and Fujitsu from extending their agreement to cover these boxes down the road. Since neither party is sharing intellectual property with the other, they can part amicably at the end of the Advanced Product Line agreement and go their separate ways, too.
Alongside the Advanced Product Line systems, Sun will continue to create its Opteron servers and low-end machines based on its UltraSparc-IIIi processors. Texas Instruments is still making the UltraSparc-IV and UltraSparc-IV+ processors today, and will make the Niagara and Rock processors as well, although it would not be surprising to see Fujitsu become a second source for Sun's homegrown chips if Texas Instruments stumbles with the 65 nanometer process that the future Niagara-II and Rock chips require.
Although Ingram didn't say it, Sun, Fujitsu, and Siemens will probably carve up the world into geographies: Sun would do the bulk of Advanced Product Line system sales in North America, Fujitsu in Asia/Pacific, and Siemens in Germany and France. There is no point in competing inside Solaris accounts. These two companies know they need to focus on beating IBM and Hewlett-Packard.