Single-cycle ARM processor (flags)

I vote for DMA. It's really flexible in Cortex-M3 and up - and you can do all kind of crazy things like automatically getting data from one place and outputing into another with specified rate or at some events without spending ANY CPU cycles. DMA is much more reliable.

But it might be quite hard to understand in details.

Another option is soft-cores on FPGA with hardware implementation of these tight things.

I'm not familiar with the ATSAM's, but I know the STM32F4's (as a comparative ARM Cortex-M) wouldn't work with an external 96MHz source - their working range is 4 to 26MHz.

I'd try a slower crystal or oscillator and use the PLL (assuming the ATSAM has one) to increase the system clock to 96MHz.

Re your problems with odd behaviour following an erase/program cycle, maybe try a different toolset. The Coocox IDE (at v.1.7.7) appears to support your MCU (AT91SAM3U4C?), although I'm less sure that it supports the SAM-ICE. If it doesn't then you can buy a CoLinkEx debugger that is definitely supported and should work with almost any Cortex-M for around US$20.

Disclaimer: I have no relationship with Coocox, I just think they have an awesome free product.

I suggest this because I've encountered a situation in the past where if I used vendor A's tools then I had to power cycle between erase/program cycles; but if I used vendor B's tools then I didn't, despite flashing exactly the same binary.