On Sunday night, after returning from a weekend trip to Hamburg, I sat down and looked at the NAND and S3C2442B data sheet to figure out the actual timing performance. Interestingly, the NAND timings were much more verbose and detailed (and had different names) than the timings described in the NAND controller section of the S3C24xx manual - and both are from Samsung ;)
Anyway, it seems like the current timing settings for the various stages (reading u-boot by the stepping stone mechanism, reading the kernel by u-boot as well as actual mtd-based access inside the Linux kernels) were extremely suboptimal. They're basically standard timings designed to work with most NAND flashes out there, ignoring the fact that GTA02 uses one specific flash with very good (fast) timings, at least according to the data sheet. There should also be no PCB / routing related issues such as capacitive overload preventing higher speeds, since the NAND flash die is stacked onto the CPU die in one package, and the NAND controller signals are not routed on the PCB anywhere.
Some initial experiments based on the calculations show that the performance can be easily improved by 41% over the stock GTA02 NAND performance. However, the actual speed (6.59MBytes/sec) is still much lower than the theoretical maximum read performance of 15.64MBytes/sec. It seems there is more room for improvement inside the MTD layer of the Linux kernel.
It's again quite amazing how much room there is for improvement in GTA02 performance, both power consumption wise (see my recent post about framebuffer blanking), as well as actual data throughput. Those are really low-hanging fruits, and it's very surprising that nobody working for Openmoko or in the Openmoko community has been able to spend some time to look into those...