and succeeded in producing a line of IBM-compatibles that still compete today. Amdahl has about 7 percent of the IBM mainframe market.

In 1979, Amdahl left to form Trilogy, a quarter-billion dollar startup that would employ wafer-scale integration to create a 2.5" superchip. Far ahead of its time, the superchip never came about. Amdahl later founded Andor, another IBM-compatible venture, but suffered from the "mainframes were out" syndrome in the early 1990s and Andor didn't make it either. Not to be undone, Amdahl is still at it with Commercial Data Servers, which recently introduced entry-level mainframes to replace the several thousand 43xx and 9370s out there at half the cost of IBM Multiprise systems. Amdahl hopes once again to have the premier mainframe, beating out everybody by the end of 1998 with a cryogenic-based 270-MIPS machine.

In addition to Amdahl 1 (Amdahl Corp.) and Amdahl 2 (Commercial Data Servers), Hitachi is the other major IBM-compatible mainframe vendor. Hitachi claims 24 percent market share overall and 50 percent share in the top 4 percent of customers. Its Skyline series indeed is the fastest single processor machine on the market today at 124 MIPS. A new 152-MIPS machine is expected shortly.


Is a mainframe a license to steal? How can Hitachi sell a 10-CPU system for $10 million that delivers a total of 1,000 MIPS (million instructions per second), when five Pentiums will deliver the same 1,000 MIPS? There are reasons, but comparing MIPS ratings and raw megahertz is about as misleading as it gets. This is the kind of bean-counter thinking that made client/server so appealing. In actuality, many mainframe instructions will do more actual processing than the equivalent Intel instruction, but that's a fraction of the issue. IBM and the compatible vendors have had 34 years to perfect the System/360 architecture, and those 34 years have produced a formidable machine. There is a difference!

First of all, mainframes offload input and output to separate computers called channels, so data transfers take place simultaneously with CPU processing and other data transfers. Additional processors may act as I/O traffic cops between the CPU and the channels to handle the processing of exceptions (What happens if the channel is busy? If it fails? etc.). IBM offers 256 channels per system and Amdahl and Hitachi support up to 512. All these subsystems handle the transaction overhead, freeing the CPU to do real "data processing" such as adding up columns and updating account balances, the purpose of the computer in the first place.

Secondly, instead of one pathway into memory as in a PC, there are multiple memory banks providing multiple ports into memory. For example, Hitachi's Skylines have 16 ports into main memory and 32 ports into cache memory, and that cache memory is 10 times faster than regular, main memory.

Thirdly, while a 200MHz Pentium has an internal bus of 66MHz, a 200MHz mainframe may have a data bus that also runs at 200MHz, three times as fast. The multipliers add up: three times the bus speed, 10 times the cache speed, perhaps 32 or 64 overlapped data transfers. Multiply one times the other, and the combination makes a processing machine unlike anything else.

Mainframes are very scalable. While lesser machines can poop out, mainframes keep on going, and going and going (shades of the pink bunny!). Using symmetric multiprocessing (SMP), they offer systems with up to a dozen processors that share a single memory. To increase capability beyond a single system, IBM uses its Parallel Sysplex technology to cluster up to 32 systems, making them all appear as a single system image. In practice, most users don't scale beyond six or seven systems, because it gets very complex (that's six or seven systems, each having multiple processors). Amdahl recommends no more than four or five.

Both Amdahl and Hitachi rely on IBM's OS/390 (formerly MVS) software, which after 34 years