SAN JOSE, Calif.--Sun Microsystems and Fujitsu engineers on Tuesday promised significant performance increases with two next-generation chips, models that they hope will help keep the Sparc line relevant.
Both processors, Sun's eight-core Niagara 2 and Fujitsu's dual-core Sparc64 VI, run Sun's Solaris operating system. However, they are aimed at different markets. Niagara 2 is for lower-end servers running Java or Web server software, while the Sparc64 VI works in higher-end servers with numerous processors that can handle tasks such as working with massive databases.
On integer processing--which represents the bulk of server tasks--the dual-core design of Sparc64 VI gives twice the performance per chip socket as Sparc64 V, said Aiichiro Inoue, chief scientist in Fujitsu's server group. And for floating-point processing, which is used in many mathematical calculations, its performance is 2.5 times that of Sparc64 V.
The chip measures 420 square millimeters, has 540 million transistors, will run at 2.4GHz and will consume a maximum of 120 watts and more typically 80 watts, Inoue said. It also has 6MB of high-speed cache memory shared between the two cores.
The follow-on Sparc64 VII also has a shared 6MB cache, but that chip will have four cores. As with the Sparc64 VI, each core can handle two simultaneous instruction sequences, known as "threads." That chip will measure 460 square millimeters, Inoue added.
Under Niagara's hood The Niagara processor takes multiple cores and threads to an extreme among mainstream central processors. The first generation has eight cores, each able to process four threads simultaneously. Niagara 2 has eight cores that each can run eight threads.
Sun likes to measure Niagara performance in terms of "throughput"--the amount of work performed in aggregate. The chip family is designed to run many tasks simultaneously, but a single thread won't run as fast as on conventional processor designs.
For integer tasks, the throughput is doubled on Niagara 2 compared with its predecessor, said Robert Golla, a principal architect at Sun. Throughput per watt of power consumed also is doubled, he said.
Single-thread performance improves, too, by a factor of 1.4 on integer processing and by a more dramatic factor of 5 on floating-point processing. Floating-point throughput has been boosted by a factor of 10.
The floating-point improvements are expected: Each Niagara 2 core has its own floating-point execution unit, while Niagara 1 had a single unit shared among all eight cores.
The Sparc64 V and UltraSparc TI are built with manufacturing processes that can create electronics elements measuring 90 nanometers. Their successors use a 65-nanometer process, letting more circuitry be squeezed onto the same chip. A nanometer is a billionth of a meter.
Sun considered making Niagara 2 a 16-core model, but determined that adding more threads was a more economical use of silicon chip real estate. "Doubling threads is more area-efficient than doubling cores," Golla said.
I'd love to see some head-to-head comparisons of the Sparc64 VI versus IBM's Power 6. As a Sun investor, I'm keen to see the Sparc win the benchmark battles, and come to market earlier than IBM.
There will doubtless be several benchmarks to pick from, but there's an unpleasant void when it comes to higher-end server tests. IBM has optimized its processors for TPC-C, an expensive test of high-end database performance. Sun, though, refuses to run TPC-C on the argument that it's obsolete. (It was created in 1993.) It's possible we'll see a TPC-C replacement (<a class="jive-link-external" href="http://news.com.com/Repairs+under+way+for+server+speed+tests/2100-1013_3-5708963.html" target="_newWindow">http://news.com.com/Repairs+under+way+for+server+speed+tests/2100-1013_3-5708963.html</a>), but don't hold your breath. There are other options, such as SAP's SD (see <a class="jive-link-external" href="http://www.sap.com/solutions/benchmark" target="_newWindow">http://www.sap.com/solutions/benchmark</a>), but that doesn't explicitly factor in price, alas.
SPARC64-VI will ship in servers in early 2007, about six months ahead of POWER6. The real comparison will be SPARC64-VI to POWER5+ and Intel's Montecito (all three are 90nm designs). Then the next big comparison will be POWER6 and Niagara 2 (both are 65nm designs), and Intel's Montvale (which remains at 90nm).
As the three vendors are on different tempos, it is hard to make head to head comparisons. Instead, leapfrogging is inevitable, and direct comparisons are somewhat unfair.
In 2008, POWER6+ (remaining at 65nm, number of cores unknown), SPARC64-VII (65nm, quad-core), and Intel Tukwila (65nm, quad-core) will be an interesting comparison.
Prominent corporate governance organization says Facebook's dual-class stock structure gives CEO Mark Zuckerberg too much control over the company's future.
Google creates an animated doodle that features a boy, a girl, Google's search engine, and a jump rope. But might there be darker, more analytical, more troubling interpretations to this tale?
When the sun goes down, that's when the iPad gets busy for folks with news readers. The iPhone? It's more of a daytime habit. If you're building an app for both devices, heed the lesson.
Is the public ready for Samsung's new Galaxy Note device, which melds tablet and phone into one unique mobile device? We hit New York streets and received some surprising results.
EnerG2 opens a plant to make an engineered carbon that will improve performance of energy storage devices and make storage for start-stop hybrid cars less expensive.
As the three vendors are on different tempos, it is hard to make head to head comparisons. Instead, leapfrogging is inevitable, and direct comparisons are somewhat unfair.
In 2008, POWER6+ (remaining at 65nm, number of cores unknown), SPARC64-VII (65nm, quad-core), and Intel Tukwila (65nm, quad-core) will be an interesting comparison.
to be an Intel-only world; at least not anytime soon.
I'd be curious to see those comparisons myself. I'll say that I've
been very impressed (overall) with IBM's technology.
Charles R. Whealton
Charles Whealton @ pleasedontspam.com