IBM Corp’s new RS/6000s models include 100MHz, 120MHz and 133MHz PowerPC 604, PCI/AT-based 43P workstations. They are almost identical to the company’s Power Series desktops but have the more robust hard disk drive controller interface, SCSI, instead of IDE and two- and three-dimensional graphics support. They come with from 16Mb to 192Mb RAM, 540Mb to 5Gb disk, three AT and two PCI/AT slots, E15 graphics adaptor and 512Kb cache (256Kb standard on the 43P-100). They are rated at 128.1, 157.9 and 176.4 SPECint92 respectively and 120.2, 139.2 and 156.5 SPECfp92 respectively with their caches. With 13.6 colour screens, CD-ROM and floppy drive the units start at $6,200, the 43P-100, $7,315, 43P-120, and $7,620, 43P-133, and ship on July 7. They come in above the existing 40P and 25T workstations. The 40P to 43P upgrade is $2,000. The 120MHz 604 Micro Channel Architecture bus-based model 42T upgrades the existing 41T. It comes with from 16Mb to 256Mb RAM, 1.1Gb to 4.4Gb disk, four Micro Channel and one graphics slot and 512Kb optional Level 2 cache. It’s rated at 118.2 SPECint92 and 116.5 SPECfp92 (150.2 and 146.5 respectively with L2 cache). With a 16 colour screen, floppy drive, Power GXT150L two-dimensional graphics and Ultimedia audio adaptor it’s from $10,950 from July 7. Models without the adaptor or screen – the 42W – costs from $9,550. New GXT500 and GXT500D 1,280 by 1,024 graphics boards are available for the 42W (no monitor or audio adaptor), 42T (with adaptor and screen) and 41W and 41T. The GXT500 is a double-buffered 12-bit add-on, the GXT500D is a 24-bit double-buffered system. Both support OpenGL, PHIGS application programming interfaces and cost from $5,800 and $7,800 respectively. The 80MHz C10 uniprocessor server is succeeded by the 120MHz 604-based C20 with 16Mb to 256Mb RAM, 1.1Gb to 6.6Gb disk, four Micro Channel slots and up to 1Mb L2 cache. It’s rated at 118.2 SPECint92 and 116.5 SPECfp92 (155.0 and150.2 with L2 cache) and costs $11,500, also from July 7. Dual-processor 75MHz 601 upgrade boards that turn the J30, G30 and R30 symmetric multiprocessing servers into six- and eight-ways from the current four-ways, start at $15,000 from the end of August. IBM reports an eight-way J30’s throughput at 3,119 tpmC against the similarly-configured Hewlett-Packard Co’s HP 9000 T500 at 3,118 tpmC and Sun Microsystems Inc’s SparcServer 1000E at 1,204 tpmC. The Transaction Processing Council said the comparison is not vaild because IBM is using version 3.0 of the TPC-C benchmark and the Hewlett-Packard and Sun figures were cut with version 2.0. Although vendors and other observers say throughput measurements using TPC-C 2.0 and 3.0 test suites are comparable, price-performance comparisons between results from the different tests are useless because 3.0 has eliminated the use of terminals. The Council, meanwhile, is struggling to eliminate confusion in TPC-C reporting and wishes to invalidate altogether the mixing of results from the different suites. Two TPC-C suites became necessary when vendors began configuring cheaper, commodity terminals in place of their own more expensive ones to lower overall system costs and produce better TPC-C 2.0 cost-of-ownership numbers.

Processors falling short of expectations

The 604-based symmetric multiprocessing RS/6000s are not now expected until later in the year – IBM said six-way 601s will perform at the same level as quad-604s. Although the company’s new units use the latest 120MHz and 133MHz PowerPC 604s, the systems do not perform as well as Somerset’s processor performance marks. Originally the 100MHz part was estimated to reach a SPECint92 of 160 and a SPECfp92 of 165, the 120MHz to go to 180 SPECint92 and SPECfp92, and the 133MHz to 200 SPECint92 and SPECfp92. The new RS/6000 models show the processors falling short of expectations. They also show the extent to which it is unwise to assign SPECmarks to a chip, rather than an entire machine. The 43P shows a 100MHz 604 with 256Kb of asynchronous level 2 cache doing SPECint92 128.1 and SPECfp92 120.2. D

oubling the size of the cache and making it synchronous improves the figures to SPECint92 140.8, SPECfp92 129.1. At 120MHz, and with the 512Kb cache, the machines manage 157.9 SPECint92 and 139.2 SPECfp92. The speediest, 1 33MHz, part is rated at SPECint92 176.4 and SPECfp92 156.5. In summary, a 133MHz 604-based machine is carrying out floating point operations slower than the alliance suggested a 100MHz chip would, while integer figures at 133MHz are worse than we were led to expect last week from the 120MHz part. IBM blames early versions of the memory controllers used in the machines. As for the next stop, it said 150MHz and 180MHz versions of the 604 are planned, but no timescales were given. Motorola Inc is the one member of the alliance yet to announce machines using the 120MHz or 133MHz processors.