Say no to commodity
Phillips stole some of Fowler's server thunder - he is the boss, so he can do that - when he said that Oracle was not putting the future of Sun's hardware business on the commodity x64 server racket, as Sun's own chiefs had done back in 2004 as a means of bolstering the company. "We're not too interested in being in the commodity x86 Windows market," Phillips said. "Other companies - Dell and so forth - are good at that. Let them do it."
Fowler didn't say much more about Oracle's x64 server plans, other than to add that the company would be focusing development in x64 machinery on enterprise-class products used in clusters and was more interested in engineering systems "from application to disk."
As El Reg has been guessing for some time, Oracle is going to hunker down with the Sparc T series of multicore chips for entry, midrange, and clustered systems. In a call after the five-hour event with the press, Fowler could not bring himself to say the Rock chip was dead, but merely said that Oracle had "streamlined down" to the CMT processors, by which he meant the T Series. (Of course, Rock was supposed to be a 16-core chip multithreaded processor as well, but that is beside the point).
In his presentation, Fowler said that there were upgrades on the way for systems based on both the Sparc T series (now made by Oracle) and the Sparc64 chips made by Fujitsu and put into the Sparc Enterprise M midrange and high-end SMP servers that are sold by both Sun and Fujitsu. This seemed to imply that Sun and/or Oracle had extended a co-development agreement with Fujitsu, but maybe not. An upgrade of the quad-core Sparc64-VII+ chip is already in the works, and Fowler did not even mention Fujitsu's eight-core "Venus" Sparc64-VIII processor.
What Fowler did say is that Oracle could field Sparc machines that had one to four processors and up to hundreds of threads and terabytes of memory and that it was committed to delivering machines with thousands of threads and double-digit terabytes of main memory. He added that Oracle's engineers, both on the hardware and software side, would be focused on application performance because this is what the customers care about.
Fowler brought Splain up on stage after a while, and he presented a very skinny Sparc roadmap, one with far less detail than the one Sun was showing customers back in June 2009, as El Reg previously reported.
The new Oracle Sparc roadmap shows the UltraSparc T3, also known by its code-name "Rainbow Falls," coming out this year based on 40 nanometer processes with double the number of cores of the T2+ chip, an improved floating point unit, improved security, expanded cache, and faster memory. Sun talked a bit about Rainbow Falls at the Hot Chips conference last summer.
The next generation of Sparc T chips will have a new core, run at higher frequencies, be implemented in a 40 nanometer process, and be compatible with Sparc T3 systems, and the focus here will be on improving application response time, according to Splain. If I had to guess, this refers to a chip called "Yosemite Falls," which was due in late 2011 and which hopefully will be moved up. Let's call it the Sparc T4 so we don't get lost.
After that comes a future Sparc T chip using the new core, but packing more of them on a chip and running them at higher frequencies, and if I had to guess, that looks like "Cascade Falls," which on the old Sun roadmap was die in late 2012 and had 16 cores and eight threads per core.
Today, Oracle said that this next-generation chip - let's call it the T4+ - would have more cores, higher clock speeds, larger cache memories, a next-generation memory controller, new I/O, and more sophisticated power management; the aim here is to have increased system throughput compared to earlier Sparc T machines. This chip will use 28 nanometer processes, too.
Further out, there is yet another Sparc chip, also in 28 nanometer processes, with higher clock frequencies, much larger caches, plus the next generation memory and I/O support and the ability to expand beyond more sockets. An "improved everything" chip, as the Oracle Sparc roadmap put it.
This chip does not appear to be the Cascade Falls or Yellowstone Falls chip, the latter of which was due in 2012 originally with four cores, eight threads, and supporting from 4 to 192 sockets in a single system image. I just don't believe that Oracle is going to create such a huge box, unless it is using a 120 Gb/sec InfiniBand switch as an SMP backplane and it is really a glorified cluster.
Not that there is anything wrong with a glorified cluster, so long as it looks and smells like an SMP or NUMA server to the operating system and applications.
Beyond this, Splain was all excited about Oracle having four chips in development at the same time, as if this had never happened before. And he added that within the next 12 to 15 months, a faster Sparc64-VII+ chip was due. At 3 GHz compared to the current 2.88 GHz, as the original Sun roadmap from June 2009 showed, don't get too excited. And Oracle did nothing to clear up what happens to the Sparc64-based product line beyond this point.
Cindy Reese, the senior vice president at Oracle in charge of the Sun supply chain, said that Oracle was shifting its new hardware manufacturing operation from a build-to-stock manufacturing operation, typical of a channel-focused company, to a build-to-order operation, typical of one that has direct relations with its customers.
Oracle has already cut back on the Sun product line by 50 per cent, according to Reese, in preparation for the acquisition, and by simplifying its supply chain it can cut its manufacturing partners in half and save money while negotiating better prices. Sun will also be able to build and distribute in fewer locations, and be able to carry much smaller inventories.
All of these things - as well as getting out in front of Oracle and Sun customers who now know that Sun is not going to go bankrupt or be eaten by IBM or Hewlett-Packard and have its products killed - are key to making the Sun division profitable for Oracle. ®