Well I've downloaded the oldest version of passmark, I could obtain on the official site which is version 5 and run the benchmark with and without L2 cache. Both CPU and memory benchmark. I thought that the differences would be bigger, it seems that passmark's tests fit into L1 cache, so disabling L2 had no big effect. Only memory benchmarks showed some differences and that was with EDO CL2 memory, SDR memory and higher fsb would help.
I also run Quake (1.06) in software mode in 320x200 and 400x300 resolution with sound, but to my surprise it didn't drop dramatically.
P233 MMX with 512kB L2 cache, FIC VA-503+ (MVP3) board, 64MB EDO RAM CL2, S3 Virge/DX 2MB, Win98SE, AD1816 sound card:
320x200 44.1 fps
400x300 29.2 fps
P233 MMX without L2 cache, EDO RAM:
320x200 38.1 fps
400x300 26.4 fps
P233 MMX with 512kB L2 cache, FIC VA-503+ (MVP3) board, 128 SDRAM 100MHz CL2:
320x200 43.7 fps
400x300 29.8 fps
The attachment PerfRes.gif is no longer available
EDIT: updated results with SDR memory running at 100 MHz with CL2 timings and L2 enabled. There is not much difference, actually passmark show performance decrease which also happened in quake in low res. Only sisoft sandra showed some interesting memory results.
It seems like 128MB is too much and is not cached, but i thought i was in cacheable area 😕