On-board RAID woes: Smarter RAID 10 and RAID 5?

MANY HIGH end PC mainboards had the RAID option available for quite a while already, usually a dedicated Promise chip offering a dual-channel RAID 0 (striping data across two disks for parallel access but with lower reliability) or RAID 1 (mirroring two disks for greater reliability and parallel reads). Sitting on the 32-bit 33 MHz PCI bus, however, reduced the performance somewhat since two really fast disks could saturate it.

Now the RAID-enabled Serial ATA in Intel ICH5R south bridge, present in their newest 800 MHz FSB chipsets, provides RAID 0 (soon RAID 1 too via BIOS update) for two SATA drives, connected via a faster, lower-latency hub interface instead of PCI. However, there is quite a bit of extra CPU utilisation spent on using two disks in striped fashion.

The next batch of SATA-enabled south bridges from Intel and others will have four SATA channels, enough for RAID 10 (mirrored striping, the best of speed and reliability in one, but net cap only 50% of gross) or RAID 5 (better net capacity at 75% gross, but complicated XOR-based ECC).

Now, some quarters may have the horrendous idea of software-based RAID 5 to essentially drain any spare CPU power even in your Prescott CPU. In the best case, there may be a hardware XOR engine in the chipset to assist on the parity/ECC calculation and that's about it.

However, with CPU power always never enough when the next great game, 3-D modeler or movie editing app comes, why not just add one of those cheap 32-bit embedded controller cells into the south bridge, to handle not just SATA RAID, but also Ethernet, sound, USB etc, so that the CPU can get some rest of all the unnecessary interrupts?

Such a move, eliminating the software-based sound, Net, RAID etc, would probably move the real system speed up much more than, say, a frequency jump from 2.8 to 3.2 GHz. It would be interesting to see how will south bridges from Intel, VIA, Nvidia and SiS handle this I/O processing overhead in the future - after all, intelligent I/O is common on servers and workstations for ages, why not move it to the desktop too? µ