Sun Firms Up Its Sparc Chip Plans
by Timothy Prickett Morgan
Server maker Sun Microsystems has been making a lot of noise about its current and future Opteron-based servers in recent months as it prepares to deliver the "Galaxy" line of Opteron machines, but the company is still working to bring improved Sparc-based machines to market based on its own chips as well as on those provided by partner Fujitsu. Sun's roadmaps from last year for future Sparc products had some pretty wide timeline targets, which is normal in the computer business.
As several Sparc product launches are approaching, Sun is more confident about launch schedules, and Andy Ingram, vice president of marketing for the Scalable Systems Group at Sun, which manages the development, manufacturing, and marketing of Sparc-based servers, is now ready to talk publicly about what is coming when. There appears to have been some slippage with some chips, but one important chip, the eight-core "Niagara" processor, might come out ahead of schedule.
Sun's mantra in the server market for the past several years has been this: different workloads require different processor and server architectures. These workloads differentiate themselves in terms of the degree of threading they have and the degree of the statefulness of the data that the machines process. The first is a measure of how applications are able to take advantage of multiple execution pipelines (whether they are virtual or physical) in a chip architecture, while the latter is a measure of how the application in question is in control (or not) of the access to and modification of the data.
A relational database is extremely stateful, in that databases absolutely control access to the data and usually lock access to it to a single user hitting a single record, while the Web pages being pulled from a server and out to browsers through a Web server is basically stateless data. Data warehouses, databases, and online transaction processing systems have lots of threads and are absolutely stateful; application servers have lots of threads too, but somewhere between 40 and 70 percent of their cycles are dedicated to handling I/O; the nodes in a high performance computing cluster have maybe one or two threads for code and maybe one or two for networking, and their data is more stateful than a Web page, but less so than a database. Ingram says that in theory, Solaris 10 can support up to 1 million concurrent threads, including those used by the operating system and the applications, databases, and middleware that run on top of Solaris. A set of Java applications are inherently multithreaded, with maybe 8 to 10 threads, while a relational database has hundreds of threads.
Sun contends that designing one system that can support all of the possible combinations of statefulness and threading (as well as other parameters, such as SMP scalability, memory and I/O bandwidth, and so forth) that also would hit the appropriate price points all through the server market is not possible. That is why it is creating Opteron-based systems, pushing UltraSparc-III and UltraSparc-IV systems, developing systems that will use its multithreaded Niagara and "Rock" processors, and big SMP boxes based on the "Olympus" Sparc64-VI processor with partner Fujitsu.
In the home-grown Sparc lineup right now, Sun is peddling the "Jaguar" UltraSparc-IV processors in its "Amazon" Sun Fire server line. The Jaguar chips were Sun's first dual-core Sparcs, and they were used in a slightly modified variant of the "Serengeti" server frames that were designed for the original "Cheetah" UltraSparc-III processors from 2001 and 2002. (Sun still sells the UltraSparc-III servers, too) By the end of 2005 and stretching into early 2006, Sun will roll out the next generation "Panther" UltraSparc-IV+ processors, which will be a new dual-core processor, into the "Amazon+" servers. The initial Panther chips will come in at around 1.6 GHz, and Sun will probably deliver parts that run at 1.8 GHz and 2 GHz. (Ingram said that Sun would deliver the UltraSparc-IV+ chips and two kickers, each spaced about nine months apart.) The Panthers will have 2 MB of on-chip L2 cache and an external 32 MB L3 cache, the first such L3 cache in a Sun server. This on-chip cache plus the L3 cache is a substantial factor in the performance boost that the Panthers have over the Jaguars, which have only L1 cache on chip and a 16 MB external L2 cache. The launch date for these improved dual-core Panther chips has gone from "some time in the first half of 2005" when Sun was talking about the Panthers in early 2004 to "mid-2005" when Sun was talking them in late 2004, to the fourth quarter when Sun is talking about them now.
The Panther chips are being created in a 90-nanometer copper/low-k chip process from TI that also adds strained silicon to shrink transistor sizes even further than that copper/low-k process allows. New chips with new processes are always tough to schedule into a server product line. Fadi Azhari, group manager for Sun's Scalable Systems Group, says that the company has not made the decision about what part of the Sun Fire line will get the Panther chips first. It seems that chip yields from fab partner Texas Instruments and marketing needs for specific segments of Sun's entry, midrange, and enterprise server will determine what server gets the Panthers, and how many. In the past, Sun has rolled new chips out in different parts of the Sun Fire product line, and often puts the latest chips out only in very heavy configurations, thereby boosting the amount of money it can get from a system using the new chip and self-selecting for the few customers who need the extra performance. The low number of customers willing to spend big bucks matches the initial low yield on the chip, maximizing profit. It seems reasonable to expect this tactic for the Panthers.
The good news is that the Amazon+ server frames will probably not be substantially different from the current Sun Fire frames, and uniboards based on UltraSparc-III and UltraSparc-IV processors will be able to be used in the Amazon+ servers. Customers wanting the Panther UltraSparc-IV+ chips will have to use the new frames, but they can carry their old uniboards with them if they want, too. Sun is expected to keep the scalability of the Sun Fire line topped out at 72 sockets, or 144 cores. It may boost memory capacity, SMP backplane bandwidth, and memory and I/O bandwidth to compensate for the extra performance in the Panther chips.
On the entry Sparc server front, Ingram said that in the first quarter of 2006, the Scalable Systems Group at Sun would roll out a kicker to the UltraSparc-IIIi, which will be the UltraSparc-IIIi+ chip. This processor, which was expected in the mid-2005 timeframe, is a low-cost, low power variant of the Sparc chip that Sun uses in entry rack machines, and the current "Jalapeno" UltraSparc-IIIi processor only scales to four-way SMP. (Sun is not expected to change this.) The UltraSparc-IIIi+ is a 90 nanometer shrink of the Jalapeno chip that will get a clock speed boost perhaps from the current 1.6 GHz to 2 GHz (that's my estimate, not a Sun statement) and will have its on-chip L2 cache boosted from 1 MB to 4 MB. The Jalapeno has an integrated memory controller as well. Sun will probably provide two kickers to the UltraSparc-IIIi+ as well, and probably at nine month intervals like with the Panther chips.
The Niagara chip taped out a year ago, and it is comprised of a Sparc processor that is based on a cut-down version of the UltraSparc-II core. Niagara crams eight processor cores, each with four of their own execution threads, plus the memory subsystem and the interface for the PCI Express bus, all on a single piece of silicon. The Niagara machines are aimed at the infrastructure and modest database and Java workloads that make use of many threads--in this case, 32 or fewer threads per box, since Niagara is not implemented for SMP clustering. Sun has always been vague about when Niagara would first be available, but Ingram now says that the chip is expected in early 2006--earlier than planned--and that performance might even be higher than the 15X performance factor compared to the Jalapeno UltraSparc-IIIi that Sun promised years ago when it launched the Niagara project. "Performance is coming in at the high-end of our estimates," says Ingram.
In practical terms, what Niagara delivers is a chip that consumes about 56 watts that has about the same performance on infrastructure workloads as a four-way SMP rig of Xeon processors or a two-way setup using dual-core Xeons. These Xeon setups will consume hundreds of watts per processor.
As for the "Olympus" dual-core Sparc64-VI processors from Fujitsu and the "Jupiter" Advanced Product Line servers that are being predominantly designed by Fujitsu but with help from Sun, Ingram says that Sun customers should expect to see them sometime in late 2006 or early 2007. The initial shipments of the Jupiter systems were expected in mid-2006 when Sun and Fujitsu announced their partnership in June 2004. Exactly what is causing the delay with the Jupiter machines is unclear.