|
Sun Details Server Chip Roadmaps at Analyst Summit
Published: Febraury 7, 2007
by Timothy Prickett Morgan
Sun Microsystems hosted its day-long analyst summit meeting in San Francisco yesterday, and the event was a lot less glum than it has been in prior years now that the company's new management is seeing some market traction and financial results on its hardware products and open source software strategies. There was not a lot of news at the meeting, but Sun did talk a little bit about roadmaps, and divulged the details of a future Sparc T1 processor that many suspected was in the works.
That particular processor, code-named "Victoria Falls," is a kicker to the "Niagara" line of multicored, low-powered Sparc T1 chips that have been part of Sun's resurgence in the server space. The Niagara chips are very efficient processors that have eight four-threaded Sparc cores (based roughly on the UltraSparc-II core) that have been rejiggered to be very efficient at running multithreaded applications. The Niagara design allows the 32 threads in the chip to run at an aggregate of about 75 percent efficiency, which is a lot higher than chips with big instruction pipes and higher clock speeds (such as Sun's own UltraSparc-III and UltraSparc-IV processors). Because of this, a single Niagara chip can do the work of a dual-core processor running at twice the clock speed or more, and do so within a 75-watt thermal design point (TDP).
The Victoria Falls chip, which is so named because Victoria Falls in Africa (between Zimbabwe and Zambia) is much larger than Niagara Falls (which separates the United States and Canada), will be available in future Sun Fire servers sometime in the first half of 2008, according to a presentation given by John Fowler, executive vice president of systems at Sun. The Victoria Falls processor probably be called the Sparc T2 and has been referred to in the rumor mill as the Niagara-3 chip up to now. The "two" in T2 would stand for the number of chips that can be glommed together, given that Victoria Falls, unlike Niagara, has the clustering electronics on the chip to support shared main memory and coherent cache memories across two processors. Sun could use either SMP or NUMA clustering to accomplish this, or some hybrid variant based on its own interconnect. The latter is the expectation, but Sun could graft HyperTransport NUMA links onto the chips, too.
The Victoria Falls chip taped out in October 2006, and Sun's foundry partner, Texas Instruments, has returned samples to Sun, which has put them into alpha systems for design and testing.
As we reported last October, when Sun had taped out the Niagara-2 kicker to the current Sparc T1 chip, the Niagara-2 chip would move from the eight cores and 32 threads of the Niagara-1 chip to eight cores and 64 threads, with a floating point unit added to each core as well with the Niagara-2 design. Measured against an unspecific metric that Sun calls "throughput," the Niagara-2 chip is expected to deliver about 2.5 times the performance of the initial Niagara-1 chips, and 35 times the performance of the stripped-down UltraSparc-IIIi single-core processors that were used in Sun's entry servers and workstations a few years ago (and which are still sold in systems today). By putting two Niagara-2 cores in a single system with the Victoria Falls design, Sun will be able to offer entry Sparc servers that have about 4.6 times the performance (again, that is throughput, not a particular industry-standard benchmark). The Victoria Falls systems are expected to have about 65 percent more oomph than the Niagara-2 servers. Victoria Falls chips will appear in rack-mounted servers and blade servers.
The Niagara-2 chips taped out in April 2006, and are expected to appear in Sun Fire servers in the second half of 2007.
Also in the second half of 2007, Fowler said, Sun will launch its next-generation of "Galaxy" X64 servers. He also explained that the Sun and partner Fujitsu would get the so-called Advanced Product Line, or APL, servers to market in the first half of this year. The APL machines are based on the "Olympus" Sparc64-VI processors and their related "Jupiter" server frames, and are essentially the servers that Fujitsu had planned to bring to market to compete against Sun in 2006. But in June 2004, when Sun was on the financial ropes and its own UltraSparc processor development was shifting gears toward the Niagara and Rock processors and its dual-core monolithic UltraSparc-IV chips were delayed, the two companies decided, in essence, that Sun would use the Olympus chips and the Jupiter servers as a stop-gap until the Rock processors and their Supernova servers.
Fowler said that the APL systems will offer about 1.5 times the performance of the current Sun Fire Sparc machines, which use the "Panther" UltraSparc-IV+ processors. This is a considerable performance boost--and one that would have been more dramatic if the APL machines had shipped in late 2005 or early 2006, as Fujitsu had originally planned, or in mid-2006, when Sun and Fujitsu were expecting to get the machines out after they announced their partnership in mid-2004.
The Sparc64 VI processors have two cores, each with two threads, for a total of four threads per socket, and is expected to run at around 2.4 GHz; the Jupiter server is expected to top out at 64 sockets, or 256 threads. The biggest Sun Fire box today using UltraSparc-IV+ processors has 72 processors, 144 threads, and runs at 1.8 GHz. If you multiply threads by clock speed and divide the Olympus numbers by the UltraSparc-IV+ numbers, you would expect to see a performance boost of around 2.4 times moving from Sun Fire to APL machines. This might mean Fujitsu has had to cut back on the clock speed in the Sparc64 VI processors to get them out the door. Fujitsu and Sun have said very little about the APL products--even when asked to.
On workloads where applications are highly threaded, top-end Supernova machines based on the Rock Sparc variant are expected to deliver more than 10 times the performance of the APL machine and about 16 times of that of the current UltraSparc-IV+ machines using the 1.8 GHz processors. That gives the APL boxes a very narrow appeal in terms of time--four, five, or six quarters, depending on when APL and Supernova machines actually ship--and performance. Fowler reiterated that servers using the Rock processors would debut in the second half of 2008. The 16-core Rock processors will probably gang up into machines using uniboard designs like past Sun Fire machines, with two through eight chips in a box, which works out to 32 to 128 cores. If the Rock has at least two threads, that works out to the magic top-end of 256 threads in the largest machine.
RELATED STORIES
Sun Tapes Out Rock Sparc Chip, Gooses Clocks on Niagara Sparc T1
Sun Talks Up Niagara-2 Sparcs, Begins Work on Niagara-3
Sun Delivers Sparc T1 in Netra and ACTA Blade Servers
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
|