Talking to ComputerWire, John Fowler, general manager of the Network Systems Group at Sun, laughed at the suggestion that he would say very much about these future machines, the development of which is being spear-headed by none other than Andy Bechtolsheim, a founder of Sun. Bechtolsheim left Sun, created a networking company, which he sold to Cisco Systems, and then created a super-secret server startup called Kealia, which Sun snapped up in February, when it first launched Opteron machines and redid its Sparc roadmap, laying the ground work for its partnership with Fujitsu for Sparc processors and servers.
Bechtolsheim was chief technology officer for Sun between 1984 and 1995, and his departure led to Sun buying the Sparc SMP server business, which is at the heart of the high-end UltraSparc-II and the midrange and high-end Sun Fire UltraSparc-III and UltraSparc-IV server lines. After selling Gigabit Ethernet switch maker Granite Systems to Cisco, in 1998, Bechtolsheim got the urge to create a new company again and founded Kealia, which was rumored to be working on network devices and servers based on the Opteron, and possibly even a video server, in February 2001.
Kealia, the name of a coastal resort town on the main island of Hawaii near the Kona coffee region, was working on a broad line of servers and storage devices, and having bought that company, in February 2004, Sun intends to bring these products to market. In the interim, Sun is shipping two- and four-way Opterons based on engineering work done by Newisys, a unit of contract hardware manufacturer Sanmina-SCI; we know them as the Sun Fire V20z and V40z. Fowler says that these machines are plug-compatible with the dual-core Opteron processors that AMD will launch them around the middle of next year.
Furthermore, when Sun says that it will deliver an eight-way Opteron box, which it will do in 2005, it apparently does not simply mean putting dual-core Opterons in the V40z, which it will do. Sun will deliver a real eight-way Sun Fire, presumably to be called the V80z. With the then-current Opteron line, Sun will have machines with two, four, or eight cores using single-core Opterons or four, eight, or 16 cores using dual-core Opterons. That in itself is a pretty broad product line.
But such a simplified line is not what Sun aspires to. When asked about the future Opteron machines, Fowler would not say much about what they might look like. The two-way and four-way rack-mounted servers will be the most boring things we do, he said with a laugh. Remember, our business model is not just about producing low-cost boxes, but creating scalable computing systems. The expectation is that these future Kealia Opteron machines will blend server and storage functions in a slightly different way than we are used to in rack-mounted servers. The odds favor a sophisticated blade server design with integrated storage arrays, with lots of server and storage virtualization features.
While Sun will say nothing about it, it also has other options beyond the entry and midrange with Opterons, particularly now that Newisys is working on the Horus chipset for Opterons. For all we know, Kealia and Newisys were already in cahoots on building systems from the Horus chipset, which will scale from four to 32 Opteron processors in a single system image.
It is ironic – perhaps intentionally so – that Horus is the Egyptian god of the rising sun. And even if Sun doesn’t adopt the Horus chipset, Solaris will almost certainly be adopted for it by third parties and will scale reasonably well on it. Some other intrepid company will step into the gap if Sun leaves one. Newisys says that OEM customers will get beta systems based on the Horus chipsets around the middle of next year – quite possibly when Sun is first getting ready to show off its own Opteron wares. If they are not cooperating, they will certainly be competing in the Opteron server space.
As for the Network Systems Group, which contains all x86 server development, Fowler says that he is not against using any Intel x86 processors in future products. We do not have an exclusive deal with AMD, he said, and should Intel come up with something competitive, we will consider it. However, so far, ugly, real-world code runs better on Opteron than it does on Xeons with 64-bit memory, even better than any benchmarks might suggest.
