Sun chip engineers talk multicore, SPARC delays

Production slipping

DVD it in many colours

It was a bizarre afternoon Monday at the Sun Microsystems office here, as one chip engineer after another came bounding out from behind a thin client to tell us more about multicore processors.

With all the Sun Fellows in the room gabbing to the press, we couldn't help but wonder if Sun's chip production would be slowed by a day or two. It turns out that the meeting had a bigger impact than expected, as Sun has pushed back the UltraSPARC IV release until the first quarter of next year. UltraSPARC V is not likely to appear until the start of 2006.

The Sun brain-trust was quick to move past these two, little slip-ups and focus instead on the new H-Series line of processors that will arrive in 2005 and 2006.

Anyone who keeps an eye on these things knows that Sun likes to talk up technology it's calling CMT (chip multithreading) and/or throughput computing. Instead of building a hulking mass like Itanium, Sun wants to put a number of simplified processor cores on a single chip and surround the sucker with copious amounts of memory.

By putting lots of processor cores on one chip, Sun will create a type of mini-SMP that can outperform faster processors. Sun argues that chips such as Itanium are designed to handle one software thread well but cannot deal with the multithreaded software common in some modern workloads. The big chips spend all their time waiting for memory, while the multicore chips will always have at least one core processing data at all times.

The trend toward multicore processors is nothing new. IBM already ships the dual-core Power4 chip, HP will go dual-core with PA-RISC later this year and Sun will follow with UltraSPARC IV. Itanium is also limping along on its way to the dual-core Montecito.

The CMT chips, however, are meant to handle certain types of software and tasks better than these high-end 64bit chips. Sun sees any applications that create a lot of network traffic such as Web services software, Java, Web servers and application servers as ones that will benefit from spreading lots of software threads across low powered processor cores.

Sun acquired the technology from Afara and was able to bring UltraSPARC I designer Les Kohn, a Sun Fellow, back into the fold with the deal.

Sun claims there is some "secret sauce" in its H-series designs and won't provide much detailed information about the technology. Marc Tremblay, also a jolly good Sun Fellow, did say that Sun is working on putting a high performance networking subsystem on the chip that will run at wire speed and handle functions such as encryption and IP acceleration. There will be three interfaces into the chip - one for networking, one for memory and one for doing I/O to disk.

Tanglewood

Sun is so bullish on its CMT processors that it claims to have made competitors quake in their fabs.

"We strongly believe that in the last three months (Intel and IBM) have revised their processor roadmaps," said David Yen, head of Sun's processor group.

Yen pointed to Intel's public disclosure of the multicore Tanglewood chip as one bit of evidence that Sun's technology made rivals nervous. The only problem there is that Intel has not disclosed Tanglewood. The Register has and no one else.

Yen cited Tanglewood's predicted performance, specs and delivery date.

While Sun suggests Intel made a recent about face, it seems more likely that the company altered the course of the Itanic back in January. Intel felt pressure from rivals and decided to give Montecito two cores instead of one and deliver the chip in 2005. Next, it killed the Shavano processor, pinning its multicore hopes onto Tanglewood.

All of this shifting does seem to show that Intel miscalculated a bit on the importance of multicore chips. This is a company that likes to make one, screaming processor and do it better than anyone else.

Sun, however, makes a strong case that different chips will evolve to handle different tasks. The old Alpha engineers that have made their way from DEC to Compaq and now Intel realized this back in 2000 (PDF) . IBM has known this for some time as well.

It seems hard to believe that Intel would miss out on a major movement in processing.

Any way you slice it, Intel, HP and IBM appear prepared to let Sun run off and talk up this CMT thing for some time. Sun needs a positive chip story to tell its users. The Sun Fellows admit the competition can pull something similar out of the hat, but argue that the designs Afara was able to pull off as a start-up will blow away anything a large company would dare to attempt.

Lots of trade offs have been made, but Sun says its happy with the results. ®

Sponsored: Network DDoS protection