December 07, 2010, 9:45 PM — Oracle will halve the number of cores in its next Sparc processor and instead improve its single-thread performance, a weak area for the chip but one that's important for running large databases and back-end applications.
The next Sparc chip on Oracle's road map, the T4, will have eight cores on each chip, down from 16 in the current Sparc T3, according to a slide shown at Oracle's systems launch last week. Both will run the same eight compute threads per core.
Chip makers have been adding cores to boost performance as an alternative to increasing clock speeds, and the Sparc 3 became the first 16-core server processor on the market when it began shipping in September.
But chips with high core counts tend to be better at some workloads than others. They are good at jobs that can be broken into many smaller parts, such as processing high volumes of Web requests and online transactions, but fare worse at big databases and ERP applications, where single-thread performance is important .
That's been a weak area for the Sparc chip historically, something Oracle CEO Larry Ellison acknowledged at last week's event. The Sparc T4 is "crucial" for Oracle, in part because it "addresses the single-thread issue," he said.
The goal is to boost the T4's single-thread performance three-fold, according to a roadmap presented by Oracle partner Fujitsu at the recent OpenWorld conference. Oracle already has T4 chips running in its labs and the processor will ship sometime next year, Ellison said.
Sun always positioned the 16-core Sparc chip -- which it used to call Ultrasparc -- as being for "network-facing tasks," or those that can be highly parallelized. But Oracle is positioning Sparc equally as a general purpose processor. The Sparc Supercluster Oracle introduced last week is a "general purpose computer," Ellison said.
It appears Oracle wants the T4 to be a more balanced processor, offering strong parallel performance but also capable of handling big single-thread workloads.
"They may want to have larger caches to support that, and if you want to devote more of the chip area to caches then that leaves less room for execution cores," said analyst Nathan Brookwood of Insight64.
Oracle also sells high-end M series servers based on the Sparc64 VII+ processor designed by Fujitsu. It's a different implementation of the same underlying Sparc v9 architecture, and one that does offer strong single-thread performance.
A Fujitsu executive appeared at the Oracle event last week to show the companies are committed to working together. He announced a new Sparc64 VII+ that doubles the cache size over the current chip and increases the clock speed to 3.0GHz.
But some analysts think that, longer term, Oracle would prefer to standardize on a single Sparc implementation, instead of supporting both the Sparc64 and Sparc T series. It could make life easier for developers who today need to optimize code for both implementations, Brookwood said.
"I think the handwriting is on the wall that, ultimately, Oracle will want to have one core that they can deploy in both these highly parallel throughput type environments and also in the higher-end environments," he said.
Indeed, the roadmap shown at OpenWorld had Sparc-based systems scaling from four sockets per server today, to eight sockets in 2013 and up to 64 sockets in 2014 -- in line with the Sparc64 -- suggesting the Sparc T chips will power some very large, scale-up systems.
"If I had to guess, I'd say that somewhere in the distant future Sun will take the cores in the T4 and use them for high-end chips, as well as having variations they use at the low-end," Brookwood said, noting that Intel uses a similar strategy for its Xeon processors.
Building a Sparc T4 with fewer cores but faster single-thread performance could be a first step in that direction.
