The proliferation of iAPX-86-compatible processors is causing headaches for those that have to measure the speeds of the things. The problem is that the new generation of Pentium class processors including the Pentium itself, the M1, Nx586 and the K5 mean that building synthetic benchmarks will become more difficult. In the good old days of the 80486, if a benchmarker knew that, say, on average 10% of instructions in Windows applications were branch instructions, all that was needed was to shove the requisite number of branch instructions into the synthetic benchmark and sit back. However, now that different manufacturers’ chips have different sized caches and different levels of branch prediction it is not just the right number of instructions that is important – getting them in the right order also counts. There are many similar new potential pitfalls for the iAPX-86 bench test designer. Ziff plans a new 32-bit CPUMark32 bench test next year to cope with the new clutch of processors.It is part of a new WinBench32 suite designed to benchmark entire machines. The labs are asking chip manufacturers for statistics and technical details.
