By Timothy Prickett Morgan
IBM Corp will attempt to revitalize its RS/6000 Unix server business today (Monday) with a new line-up of systems that includes the Condor RS/6000 S80 server as star of the show (CI No 3,735). The new systems represent IBM’s effort to become the performance leader in the Unix market and mount a serious challenge for Sun Microsystems Inc, which shows no sign of being ready to launch its own new range of faster UltraSparc III Cheetah boxes, which were originally due onto the market by now.
As reported, the PowerPC Pulsar chips used in the S80, the follow-on to last year’s PowerPC Northstar Blackbird S7A server, are running at 450MHz. The S80 is based on a new six-way motherboard, and the server has four of these cards lashed together to give it SMP scalability of up to 24-way processing. That’s twice as many processors in the current Northstar line as well as in the S/390 G6 mainframe line, both of which top out at a dozen processors. It will also be available in 12-way and 18- way configurations. In effect, the Condor replaces the S/390 mainframe as IBM’s most scalable SMP platform – at least until AS/400s based on the follow-in I-Star chips arrive next year. Based on various preliminary benchmark tests, it looks like a twelve-way Condor will have as much as 20% more power than a twelve-way 9672-ZZ7 mainframes, and that because of IBM’s very good SMP ratios, the 24-way Condor will have about twice the power of that twelve-way G6 Turbo mainframe.
The S80 Condors will come with up to 64Gb of main memory, twice as much as on the S7A, and nearly four times the internal bandwidth. That, along with tweaks to the operating system, AIX 4.3.3, means that Pulsar-based S80s can do anything from two to four times the work of an S7A, depending on the workload. IBM has not yet divulged the amount of L1 or L2 cache on the chip. Systems will ship on September 24. Preliminary pricing set last week had the base price of an S80 with six processors at $290,000, but this could change on announcement day morning, with IBM locking in prices at 5am. Additional six-way Pulsar processor cards are expected to cost about $200,000 each. There are generous upgrade offers for S7A users, and the upgrade process is apparently very easy. In addition, IBM says that it will be cutting S7A processor prices by 20%. No other price cuts are expected in the RS/6000 line.
As far as performance goes, IBM says that when it finishes putting together its final TPC-C online transaction processing benchmark test results on the Condors, it will be able to process over 120,000 TPC C transactions per minute (TPM) at a cost of under $80 per TPM. That’s about 3.5 times the scalability of the Blackbird S7A, although it is only about 10% cheaper. That Condor performance will be marginally better than Sun’s best and only TPC-C results on a 64-way Starfire Ultra Enterprise 10000 server, which could handle 115,396 TPM, but which cost $106 per TPM. The 32-way Hewlett-Packard V2500 similarly topped out at 92,834 TPM on the TPC-C test, and did so at a relatively impressive (at the time) $88 per TPM. IBM also says that the Condor’s 16,640 concurrent users on the SAP R/3 Sales & Distribution benchmark test significantly beats the 14,400 users that Sun’s Starfires can support running the older, less dense version of R/3 – Starfires running the current version would probably support around 6,600 concurrent users, estimates IBM.
IBM says that the 24-way Condors will show 62% better performance than a 64-way Starfire on three tier Baan implementations, and 32% better performance compared to a 32-way Starfire running Baan database and application servers on the same box. With these kinds of performance numbers, Sun is under tremendous pressure to get the 600MHz UltraSparc-III Cheetah follow-ons to the current 400MHz UltraSparc-IIs used in the Starfires out the door. These were supposed to be out by mid-2000, but now it looks like it will be September next year before Sun gets them out the door. That’s a very long time in the high-end Unix market, as IBM knows full well from being an SMP laggard for two years.
The performance gap could be even greater if IBM uses its silicon-on-insulator I-Star chip in future RS/6000s as well as AS/400s. The question now is if IBM has learned any lessons. Will it make the Pulsar chips used in the S80 obsolete as soon as I- Star is ready in mid-2000 or so? It had better do that to keep sales running. When the 560MHz I-Star is ready, today’s S80 buyers will want it and should get it, and later in 2000 when the 800MHz I-Star is ready, they should get it, too. IBM should similarly make obsolete the new 222MHz Power3 chip used in the new SP nodes and T70 technical servers (see separate story) as soon as Power3-II is ready, perhaps by the end of this year, perhaps early next year, perhaps running at 300MHz, perhaps even faster. Big Blue has got to run full out and stop trying to control the release of technology. The company should let customers pull products off its shelves, not try to push them, and customers always want faster components. It’s success with the S70, S7A and H70 servers shows that this strategy will work.
