PCI controller bottleneck
Sender throughput using Intel Pentium (100 MHz and 166 MHz) Triton motherboard (50 MHz and 66 MHz) and comparing performance of 82437 FX (25 MHz) and 82437 VX (33 MHz) PCI controller chipset using DMA transfer over AMCC S5933 MatchMaker bridge.
32b x 25 MHz (33 MHz) = 0.80 Gb/s (1.06 Gb/s) maximum possible burst rate.
32 kB packets using machine code (bypassing NT OS) observe 300 Mb/s sustained throughput for VX and 340 Mb/s sustained throughput for FX implementation.
Sustained DMA throughput using Windows NT 4.0 for packets larger than 64 kB is < 140 Mb/s.
Windows NT performs virtual address translation for packets greater than 64 kB. NT device driver improvements include common buffer aligned to 4kB page large enough to accommodate the complete message in a single contiguous region of physical memory.