By Timothy Prickett Morgan
The latest word from one of the big Wall Street brokerage houses is that Sun Microsystems Inc’s Excalibur Ultra 100 workstations, which were expected in March 2000, have been pushed back to some time in the second quarter. The Excaliburs, which will use the long-awaited Cheetah UltraSparc-III processors, are the follow-ups to the current Ultra 80 workstations, which can be configured with up to four 450MHz UltraSparc-II processors.
The Excaliburs will have integrated Firewire and USB ports on their motherboards, and it’s these that appear to be giving Sun the biggest headache. That and the enhanced graphics applications, the Zulu frame buffer and Shaka Zulu multiplexed frame buffer. It wouldn’t be the first time that Sun has had trouble delivering its high-end workstation graphics subsystems. Sun gets it right in the end, but building what amounts to a mini desktop supercomputer is a complex problem. Its perseverance is why Sun owns the Unix workstation market and why HP and IBM have to a large extent refocused on NT workstations.
We’re also hearing about a new low-end UltraSparc-III-based workstation dubbed Grover, but the exact configuration of this machine is not yet known. Grover would be a follow-on to the current Ultra 10 workstations. Exactly when it might ship is unclear, but more than one person has told us that April is now likely. As far as we know, Sun has 400MHz and 416MHz UltraSparc- IIIs up and running with both 4Mb and 8Mb L2 caches in actual machines.
On the server side, Goldman Sachs is telling investors that they should not expect to see UltraSparc-III-based server announcements from Sun until March 2000. Some big Sun customers have been talking about the so-called E12K, a follow-on to the current 64-way Starfire server that has 128 processors rather than 64. Presumably this machine would use faster 450MHz or 480MHz UltraSparc-II processors. By our own estimates, a 64-way Starfire with 450MHz UltraSparc-IIs should be able to top 130,000 TPM, about 10% more power than the current 400MHz models and with 480MHz processors the Starfire should be able to hit almost 140,000 TPM. If Sun could get similar scalability efficiencies on the 128-way Starfire machine as it does on the 64-way box, it might get performance as high as 175,000 TPM with the existing 400MHz chips or as much as 200,000 TPM with 480MHz processors. That would put Sun back in the driver’s seat in terms of performance, probably even ahead of Hewlett-Packard’s 550MHz PA- 8600 chips running in its 32-way V2500 Unix servers, which are due in January or February 2000.
The follow-on to the Starfire servers is Serengeti, which reportedly will have up to 1,000 UltraSparc-III chips in a single frame. Sun has been telling customers that they shouldn’t expect to see Serengetis until September 2000, but that could also change depending on the reliability of these boxes, which include new ASICs and I/O subsystems and which appear to be having plenty of software glitches in their beta testing.
Given the delays in the Serengeti servers, which will be more than a year late when they ship next year, there seems to be renewed interest in the Starcat clustering. Starcat is a set of fiber optic clustering hardware that presumably works in conjunction with the eight-node clustering in Solaris 8 that will ship after the February 2000 ship date of Solaris 8. The exact specs for Starcat are as yet unknown, but some techies are saying that the setup in Solaris for Starcat implies a maximum of 554 processors in the cluster. It is unclear if Starcat will use NUMA clustering, but Sun didn’t buy the engineering teams from Thinking Machines and Kendall Square just to keep them from falling into the hands of its competitors. The latest word is that Starcat won’t be available until November or December, which puts it a few months after Serengeti.
Finally, no one is talking about the fact that the Cheetah chips are supposed to come out at 600MHz but appear to be only running at 400MHz or 416MHz. At those clock speeds, it will barely outpace the 450MHz UltraSparc-II. The 450MHz UltraSparc-II with 4Mb of L2 cache has a SPECint95 rating of 19.6 and a SPECfp95 rating of 27.1. At 600MHz, the UltraSparc-III is expected to have a SPECint95 of about 35 and a SPECfp95 of about 60. Choke the clock speed down to 400MHz, and those numbers will drop to a SPECint95 of about 24 and a SPECfp95 of about 40.
