Sun SPARC catches fresh fire with new Fujitsu UNIX servers

A quad-core, dual-threaded processor that until now had been tested by the likes of the Japanese space program, is being unveiled this morning for the consumer space by Sun and Fujitsu.

Last May, the Japan Aerospace Exploration Agency (JAXA) announced it had built a supercomputer assembled from 424 new Sun processors delivering 3,392 computing cores, with a design that even some of Sun's employees weren't expecting. With eight threads per processor, they couldn't have been "Olympus" processors, which were only dual-core, dual-threaded. JAXA placed the order from Fujitsu back in February.

Indeed, JAXA had an early hand in testing Sun's quad-core, dual-threaded 65 nm "Jupiter" processors, which today found their way into commercial-grade servers for the first time: Fujitsu's M4000, M5000, M8000, and M9000, designed for Solaris 10. They're called SPARC64-VII processors now, which you'd think would be no surprise. As it turns out, some of Sun's own developers were tripped up to the development when they saw references to "Olympus" paired with "Jupiter" inside the comments of newly minted source code they'd received from the company.

What's particularly important from a developer's standpoint is Sun's approach to parallelism. If you're a veteran of low-level programming for Intel processors, you know that company simultaneously implements two approaches to running multiple threads simultaneously: Its Itanium processors require programmers to explicitly build thread-forking instructions into their code. Meanwhile, x86 programmers can rely on Core 2 Duo and multicore Xeon processors to find the best way to fork threads and distribute them among cores on their own, although they must often invoke certain features in their code compilers for optimizing their code to enable CPUs to do this.

Fujitsu's FX1 server rack, comprised of its new M4000 servers with SPARC64-VII quad-core, dual-threaded processors

Sun's UltraSPARC architecture tries a more radically simplified approach: enabling purely automatic multithreading without the programmer's intervention. Among the highly optimized procedures in Sun's new microcode is a single low-level instruction that enables multiplication of two operands and the addition of a third: a kind of x * y + z. Why is that important? It dramatically increases the speed of encryption/decryption operations, driving forward a scenario where 100% encrypted data storage and throughput can take place with negligible performance degradation.

JAXA claims its SPARC64-VII-endowed supercomputer is capable of 135 teraflops (trillions of instructions per second, or Tflops). That would put its cluster at #11 on the University of Mannheim's Top 500 list, though as of June, not a single Sun-based cluster was counted on that list. Sun claims each SPARC64-VII CPU itself is capable of 40 gigaflops (Gflops, or billions of instructions per second).

Fujitsu's server packages enable symmetric multiprocessor (SMP) arrays of up to 256 cores each, though using fewer processors thanks to core consolidation. That feature alone contributes to what Fujitsu is calling a 44% power reduction.