Schwartz Blogs a Bit About the Dud Rock Chip on His Desk

Published: April 12, 2007

by Timothy Prickett Morgan

Increasingly, companies are relying on their executive blogs as a way of communicating with the outside world rather than the more traditional means of using public relations channels and relying on the interested media outlets to provide coverage. And so it is that we learn from the blog of Jonathan Schwartz, Sun Microsystems' chief executive officer, that the company has received first silicon on its future "Rock" multicore Sparc processor.

The chip that Schwartz has in a baggie on his desk, which is shown below, is actually named the UltraSparc RK. This particular chip is a dud that doesn't work--which is what happens when you first implement a chip in silicon; you get some duds that you can use for PR purposes as the few good ones you get end up in alpha systems to test software.

The Rock chip has a total of 2,395 pins, and according to Schwartz, there are 812 pins that feed into the Rock cores and another 1514 pins are used for power and grounding. That leaves another 69 pins that are not yet connected to anything, and it is interesting to ponder what other things Sun will eventually cram onto the Rock chip as the chip evolves and chip fabrication processes allow the electronic components in the chip to be shrunk. As we previously reported, the Rock chip will have 16 Sparc cores, and Schwartz added another piece to the puzzle in his blog by saying that the Rock chip will be able to support 256 TB of memory in a coherent manner. Top-end RISC/Unix and X64 servers are stretching to hit 2 TB these days.

"That is an awful lot of RAM in a single system (and given the cost of memory nowadays, you'd want to post an armed guard next to that machine)," Schwartz quipped.

Sun has not said how many Rock processors it will gang up in its future "Supernova" servers when these boxes start shipping next year. Sun has not said how many threads each Rock core will have or what their clock speeds will be, but the company did say back in early February that a top-end Rock machine would have 16 times the performance of a Sun Fire E25K server using the dual-core UltraSparc-IV+ processors. That E25K machine has 144 cores in total spread across 18 four-socket system boards. Each system board can support up to 64 GB main memory using 2 GB DDR2 DIMMs, for a maximum main memory of 1.15 TB. That works out to 8 DIMM slots per CPU socket, or 4 DIMM slots per core in the E25K. The UltraSparc-IV+ chips are interconnected with the 150 MHz Sunplane interconnect, which provides NUMA/SMP clustering to give Solaris a single system image to play with.

Back in February, Sun reconfirmed that it was on track to deliver the Rock processors in the second half of 2008. It has not said much about the Supernova systems, but it seems likely that Sun will use a uniboard design as it did with the past Sun Fire product lines. But Rock machines may look a lot more like the high-end "Galaxy" X4600 servers, which are based on AMD's Opteron Rev E and Rev F processors. In the X4600, a single processor, its main memory, and its interconnect are all put on a single board rather than on a board with four sockets. (This is a hunch, not any inside dope.) But, Sun could also deliver uniboards with two, four, and eight sockets, too. That would put from two to eight Rock chips in a box, which works out to 32 to 128 cores using 16-core chips; Sun will almost certainly have Rock chips with four, eight, and 12 cores activated, thereby recycling chips that have some dud cores that it would otherwise throw out. Whether Sun uses single-socket uniboards or more standard uniboards with multiple sockets is unclear. What does seem clear is that Sun will probably not go beyond 256 threads in the biggest system, which means Rock is a 16-core, 32 thread processor. Future Rock designs and future Solaris releases will almost certainly support more threads. And that is probably what those extra pins are all about.

So how is Sun going to get 16 times the performance of the current crop of generally available Sun Fire 25K systems? Well, just moving to 65 nanometer processes should have allowed the Rock chips to hit 3 GHz. That's 67 percent more oomph just from clock speed alone, and it gets Sparc chips humming at the same clock speeds as Opteron and Xeon processors. But this theoretical clock speed improvement is mitigated by the fact that Sun appears to be moving from 144 cores with the E25K down to 128 cores with the top-end Rock machines. When you wash it all out, it should be a 1.5X performance increase. How, add in the effect of simultaneous multithreading and the "scout" threads that the Rock will have. SMT should be a 40 percent boost right there, which puts you at 2.1X. The scout threads are not traditional instruction streams, but a means of running instruction streams way ahead to make sure local cache memories have the data they might need as the real instruction streams progress through the CPU cores. This is a very sophisticated kind of prefetching that instead of predicting what data is needed actually runs ahead real fast in the instruction stream and gets the data. (Maybe this is unsophisticated and therefore very smart.) Sun has been vague about the scout threads and their effect on performance, but it is safe to assume that it is substantial.

Moving to fully buffered DIMM main memory, and having a lot of it--128 DIMMs on the two-chip Rock machines and 512 DIMMs on the eight-Rock versions, reportedly--is also going to significantly boost performance, too, for the Rock chips. And if Sun moves its main memory controller as well as cache controllers and tables onto the Rock, chip, too, this will provide a big performance boost.

Here's what I want to know: What does "RK" really stand for? Real Killa? Rice Krispies? Radial Keratotomy? R.K. Narayan?

RELATED STORIES

Yen Explains Sun's Chip Strategy

Sun Boosts Performance of UltraSparc-IV+ Chips

Sun Details Server Chip Roadmaps at Analyst Summit

Sun Tapes Out Rock Sparc Chip, Gooses Clocks on Niagara Sparc T1

Sun Talks Up Niagara-2 Sparcs, Begins Work on Niagara-3

The X Factor: High-End Chips Draw Even, Vendors Prepare to Differentiate

Sun Merges Server Units, Taps Key Exec for Storage

Sun Finally Announces Niagara-Based Sparc Servers

Sun Firms Up Its Sparc Chip Plans

Sun Allies with Fujitsu for Future Joint Sparc Platform

Sun Pulls the Plug on Future Millennium, Gemini Chips

                     Post this story to del.icio.us
               Post this story to Digg
    Post this story to Slashdot