IBM, NCR ready Unix systems

Electronic News, April 27, 1992 by Gerry Khermouch

NEW YORK -- IBM and NCR are getting ready to ship high-performance, mid-range Unix systems targeted at commercial users who might otherwise rely on mainframe or proprietary mid-range solutions.

IBM, as anticipated (EN, April 20), last week unveiled a new high end to its RS/6000 line of RISC servers, the PowerServer 970, with a claimed performance rating of 100.3Specmarks, making it the first single-processor RISC machine to break the 100Specmark barrier.

More pertinent to potential commercial users, IBM officials said the new model also rated 100.9 transactions per second on the TPC-A benchmark monitored by the Transaction Processing Performance Council. That vaults it ahead of such machines as Sun's $800,000 SparcServer 690MP multiprocessing system, at 95tps, and Hewlett-Packard's $1.2 million HP 9000 Series 877S, at 75tps its highest-rated uniprocessor. The new IBM model, based on a 50MHz processor, will ship late in June at a price of $97,822 for an entry-level configuration. Price/performance works out to $12,500 per tps.

At the same time, officials of AT&T's NCR unit confirmed that their Intel-based 3450 and 3550 multiprocessors, which had been delayed for three quarters because of ASIC problems (EN, Nov. 4, 1991), have now reached general availability, at newly available benchmark levels that place those multiprocessor machines close to the head of the pack in raw performance and out front in price/performance. A four-processor 3450 clocked in at 100.3tps, or $8,422 per tps.

With HP said to be preparing enhancements for the high end of the 800 line as part of its own aggressive foray into commercial accounts, observers expect the Unix-based systems to have taken another price/performance leap in challenging mainframe and proprietary systems for some of their core data-center accounts.

Details were scarce last week on the planned HP announcement, expected by the end of May, but HP is likely to offer improvements to the systems' backplane and memory management, along with some increase in sheer processing speed. Its two-way and four-way versions of the Model 870S already are well ahead of 100tps, but at prices of $2 million and $2.7 million, respectively.

Jeffry A. Erramouspe, manager for marketing programs at NCR's Multiprocessor Systems Business unit, said that while the 3450 and 3550 are only now reaching general availability, systems numbering in the "triple digits" already have gone out to users. He insisted the delayed production ramp hadn't hurt NCR, in part because the economic slowdown stretched out purchasing decisions.

Mr. Erramouspe said that while the machines still lag some specialized providers like Sequent in raw performance, they are well ahead in tps per dollar. Anticipated future improvements include a shift from 5MBps SCSI today to 20MBps by year-end and, by next-year's first half, a move to the new P5 microprocessor from Intel, which should boost raw performance by 50 to 75 percent with no change in software. Secondary cache, now 256KB, may be doubled, and third-level cache of 2-to 4MB may be added to each of the four system processors at the high end.

At IBM's Advanced Workstation division, Phil Hester, vice president of systems and technology, said the Model 970 is aimed at both commercial and technical accounts, but acknowledged that it was offered in part to accommodate commercial customers who had been requesting additional disk-storage and transaction-processing capabilities. The Model 560, introduced in January, already has received a strong reception from technical customers, so the Model 970 currently fills a greater need on the commercial side than on the technical side, he indicated.

Part of the performance improvement was attributed to improvements in the 0.5-micron CMOS process used in the Model 560, with higher back-end densities in the I/O wiring now supporting a quadrupling of on-chip instruction cache to 32KB. A two-way, set-associative design requires a heavier emphasis on ASICs in the system design, but improves the hit rate over that of more conventional direct-mapped caches, particularly for such larger-footprint workloads as database management and on-line transaction processing, applications often demanded by commercial users.