With IBM getting ready to complete its Power5 server rollout, it is not a coincidence that Sun Microsystems has lifted the curtain a little bit higher on its plans for the ‘Panther’ UltraSparc-IV+ processor…
While Sun is talking up its Opteron plans quite a bit, an awful lot of its enterprise customers are on big Sparc boxes, and they want to see big performance and price/performance gains on the midrange and high-end Sun Fire UltraSparc platforms, not just on entry Opteron machines.
Back in February, just two weeks after Sun announced its dual-core ‘Cheetah’ UltraSparc-IV processors, the company revised its Sparc chip roadmap, saying that it would deliver a kicker to the chip called the Panther internally and the UltraSparc-IV+ to the outside world. At the time, it said merely that this chip would offer twice the performance as the top-end 1.2GHz Cheetah chips. Shortly after that, as Sun was working out its joint alliance with Fujitsu on merged Sparc server product line, Sun killed off the future ‘Millennium’ UltraSparc-IV processor and said that the UltraSparc-IV+ had plenty of power to get customers through until the ‘Rock’ massively multithreaded designs arrived several years hence. This was not exactly comforting.
Sun is expected to crank the clock on the Jaguar processors from 1.2GHz to a higher speed (the exact amount is unknown) within about nine months from now, which puts it in the end of May to end of July 2005 timeframe. About nine months after that (meaning the end of the first quarter/beginning of the second quarter next year), Sun is expected to debut the Panther chips. This will very likely be the last of the UltraSparc processors to come from Sun, since the joint Advanced Product Line high-end servers will use the ‘Olympus’ variants of Fujitsu’s future Sparc64-VI, which are clones of Sun’s own UltraSparc chips.
Prior to the deal with Sun, Fujitsu was readying its first dual-core Sparc64, the Sparc64-VI, for market entry in late 2005 or early 2006, and although neither Sun nor Fujitsu want to say so, this chip is almost certainly Olympus with very minor changes. The Sparc64-VI was to be implemented in a 90 nanometer copper/SOI process, with 128KB of on-chip L1 data/instruction cache per core, 6MB of on-chip L2 cache memory shared by both cores, and an initial target speed of 2.4GHz, with speeds eventually ranging between 2.1GHz and 2.6GHz.
For the Panther UltraSparc-IV+ chips to be of any use for Sun, they have to bridge the performance gap between those faster Jaguar chips (which is expected to run at an estimated 1.3GHz to 1.5GHz, higher if Sun’s fab partner, Texas Instruments, can get good yield on the parts) and the future Olympus chips. Sun was previously expected to crank up the clock to at least 2GHz and push as close to 3GHz as it could to boost performance, but Sun has changed the chip memory hierarchy to boost performance instead.
Specifically, Sun is integrating L2 cache on the Sparc chip for the first time, moving from 16MB of external L2 cache to 2MB of on-chip L2 cache. That on-chip L2 cache is enabled because TI is moving from a 130 nanometer process with the Jaguar chips to a 90 nanometer copper/low-k process that also adds strained silicon to shrink transistor sizes even more.
The UltraSparc-III and UltraSparc-IV designs have 100KB of L1 cache integrated on chip as well, and it is likely that this L1 cache size will not change with the Panther chips. Sun says it is also adding an external L3 cache, which will be 32MB in size, to the Panther chips, and that it is expanding buffers, providing better branch prediction, and improving prefetch algorithms to boost performance. Presumably these factors are how Sun can claim that it can double the performance from Jaguars running at 1.2GHz to Panthers running at 1.8GHz, which Sun has now said is the initial target clock speed for the dual-core chips that will take the company up to the APL products.
While Sun is adamant, as always, that the Panther chips will offer binary compatibility for Solaris applications compiled on older Sparc chips, it is unclear if the Panther chips will plug into the new Sun Fire Enterprise line or if they will require a slight modification of the underlying hardware, like the jump from the UltraSparc-III to the UltraSparc-IV processors did. The odds favor Panthers plugging into any server slot that supported Jaguar chips.