Sun Puts UltraSparc-IV+ Chips in Its Big Boxes

by Timothy Prickett Morgan

Server maker Sun Microsystems this week has completed the roll out of its "Panther" dual-core UltraSparc-IV+ processors, the chips that have allowed Sun to get its systems to something approximating parity compared to Unix alternatives from IBM and Hewlett-Packard, by getting them into its high-end E20K and E25K servers. This is something of a triumph for Sun, which was hammered during the UltraSparc-III generation by aggressive Power and Itanium technology and even more aggressive pricing from IBM and HP.

Sun delivered its first dual-core "Jaguar" UltraSparc-IV processors in February 2004, and while these chips offered around 1.8 times the performance of the single-core UltraSparc-III processors they replaced, this was not enough to close the substantial performance gap the 36-socket E20K and 72-socket faced on many workloads compared to IBM's 32-socket Power4 and Power4+ machines and HP's 64-socket Itanium 2 machines. The Sun designs allowed lots of processor threads to be put into a single system image, which was good for certain workloads, but because the dual-core UltraSparc-IV+ processors had relatively low clock speeds, it still took a lot of threads to get high throughput on many workloads. And on some workloads, where the speed of a single thread is important, this was still an issue. The Panther UltraSparc-IV+ chips fix this problem in a number of ways, and Sun's high-end customers and the executives in Sun's Scalable Systems Group, which creates and sells Sun's midrange and enterprise Sparc boxes, must be happy about this. The Sun top brass certainly are happy, and they make no bones about who they want to be compared to. "We are really taking IBM to task," says Fadi Azhari, group manager for the Scalable Systems Group. Sun is doing this in a number of ways.

First, Sun nearly doubled the performance of its homegrown Sparc chips when it moved to the Jaguar dual-core processors, and it has doubled it again in 18 months (right on the Moore's Law curve) by moving to a 90 nanometer copper/low-k/strained silicon chip making process to make the Panther chips. The shrink wasn't just about a clock speed jump--in fact, at 1.5 GHz, the Panthers are not much faster than the Jaguars. Sun used the shrink in die size to bring 2 MB of L2 cache memory onto the chip, which is the first time Sun has had L2 cache on a high-end Sparc chip. (L2 caches have been within the chip package, but not on the Sparc chip itself.) The Panther chips also have 32 MB of L3 cache per chip, the first L3 cache Sun has ever delivered. The integrated L2, the addition of the L3 cache, and a restructuring of the Sparc pipeline to interface with these caches are big factors in the performance gains that come with the Panthers on its midrange and, starting this week, its high-end Sun Fire Sparc boxes.

Improving performance is only half the battle in the server market, and Sun has kept the prices of its Sun Fire Sparc boxes (ranging from the entry E2900 through the high-end E25K) constant as the company is putting Panther chips in them instead of Jaguar boxes. Sun tried to charge a 30 to 40 percent premium for the jump from machines using the single-core "Cheetah" UltraSparc-III processors to machines using the dual-core Jaguars in 2004. This time around, because Sun wants to close the performance and price/performance gap on at least some workloads compared to Unix alternatives, it is just charging the same price for configured Sun Fire servers. This is commendable behavior, considering Sun is now charging a 30 to 40 percent premium for a machine that does about 3.5 times the work of a Sun Fire box from 2003.

Based on my own estimates, I reckon that a high-end E25K with 72 dual-core Panthers running at 1.5 GHz could do about 2 million transactions per minute (TPM) on the TPC-C online transaction processing test. I figure that a fully-loaded machine with 288 GB of main memory and base disk drives would list for around $3.35 million and sell on the street for about $2.3 million after a 30 percent discount, or a little more than a buck per TPM. However, per-core pricing for the Oracle 10g database on such a machine would cost $3.2 million after a 25 percent discount. By comparison, IBM's 64-core p5 595 server with 128 GB of main memory lists for around $3.5 million and would sell for around $2.65 million after a 25 percent discount. This machine can do about 3.2 million TPM, which is about 82 cents per TPM for the hardware. IBM still seems to have the advantage on the hardware pricing. And IBM has a rather large pricing advantage when it comes to Oracle 10g, which would only cost about $1.44 million after a 25 percent discount for this behemoth. Still, at the end of 2003, IBM and HP could field machines that hit 1 million TPM, and Sun could only do about half that, if my estimates are on the mark. I think they are, and I also think Sun has come a long way on performance. Basically, it can match a 64-socket "Montecito" Itanium-based system when HP brings it to market early next year, and it can complain that the TPC-C benchmark is not representative and try to steer customers to other benchmarks that Sun is happy to run.

For instance, Sun is trying to talk up a benchmark on the Manugistics supply chain software suite that shows a Sun Fire E20K with 72 1.5 GHz Panther cores coping with 66.7 million product SKUs/hour compared to a 32-core p5 590 using 1.65 GHz Power5 cores, which could handle 38.4 million SKUs/hour. First, these two tests use slightly different versions of the Manugistics tests--v7.1 for Sun, V7.2.0.1 for IBM--and second, IBM can field a bigger box than this. Just to make it interesting, so can Sun. And the fact that neither of them chose to test their biggest boxes on the Manugistics tests might say a lot about how real applications scale beyond 32 cores or 64 threads. Then again, it may not. Because no server vendor runs the same tests on all of their machines, comparisons are difficult and so is any kind of certainty in a statement like that. And this is absolutely intentional on their part. Having said all that, it has been a long time since Sun could even claim to be in spitting distance of IBM's Unix servers on any benchmark, and that is going to make IBM hopping mad. This is for certain.

Sun has withdrawn the pricing information for the E20K and E25K servers from its Web site, so it is not so easy to get prices for these new boxes. Sun says that an entry E20K with the Panther chips costs around $500,000 and that an entry E25K costs $1 million. These numbers are not very useful--and I told Sun's Azhari as much and asked him to take the high moral ground and put pricing for small, medium, and large configurations back on the Web site. Sun has nothing to hide.

The Panther story will continue to get interesting, particularly if Sun can crank up the clock speed on the chips in the coming year. But IBM will likely still have the performance lead on many tests. Sun has said in the past that it would deliver two kickers spaced about nine months apart after the initial Panthers, and the first kicker will be a 1.8 GHz chip. The final kicker, due around the end of 2006 or in early 2007, will probably run at around 2 GHz. IBM has put its Power5+ chips in the field running at 1.9 GHz in its dual-core module (DCM) packaging running in entry machines and a 1.5 GHz quad-core module (QCM) that puts two Power5+ chips (that's four cores) in a single processor socket. The Power5+ chips are only available in machines with one- and two-socket p5 servers. IBM is expected to roll out its Power5+ chips into its p5 570, 590, and 595 boxes, which scale up to 16, 32, and 64 cores, some time next year. These multichip module (MCM) versions of the chips will probably north of 2 GHz. If history is any guide, the Power5+ MCMs should run at about 2.2 GHz and yield about 15 percent more oomph. If IBM doubles up the chip counts in the MCMs as it did to create the QCMs for the midrange p5 550Q announced a few weeks ago--and there is no indication that it will do this, either--then IBM could put a very, very powerful p5 box in the field. It would have 128 cores and 256 threads (thanks to simultaneous multithreading) and might be able to do 5 million TPM. But don't count on such a box being created. Just ponder the possibilities.