I have a firewire based system based on a MOTU 896HD and a second MOTU 896HD via ADAT bridge.
 
Back when I ran Win XP Pr 32 bit I was able to use a MOTU driver that let me run a 64 sample buffer with a round trip latency of 178 samples or 4.03ms.
 
Ever since I bought my Win 7 Pro 64 bit system I have had to use a more recent MOTU driver, one that includes extra buffers for MOTU mkIII Cue Mix efx dsp functions. My gear doesn't have any dsp functions but I have to put up with the extra latency.
 
My current round trip latency for a 64 sample buffer is 308 samples or 6.98ms. Nearly twice as long as I enjoyed back when I ran a 32bit Win XP machine.
 
I'm wondering if MOTU's PCI-E hardware comes with 64bit driver options that have the extra low round trips I used to enjoy.
 
It seems like it might be worth it to add a PCI-E card and a run both MOTU896HDs via ADAT to an interim device like a 2408. That would give me 24 channels at 24/44.1.
 
Another option might be to pick up two used MOTU HD192s and a fresh new PCI-E card which would give me the option to run 24 channels at 24/96.
 
 
If that isn't worthwhile I may start saving up for a RME UFX.

Hi Mike,
 
I believe the PCIe 424 controller card can go down to a 16-sample ASIO buffer size.
That's extremely low (sub 2ms) round-trip latency.

The PCIe 424 controller card is one of the best when it comes to round-trip latency.
 

Thanks for the info Jim.
 
I was looking at the Antelope Orion 32 and thinking that it was just what I wanted but then I began to understand that it uses a 128 sample buffer and that Antelope suggests that you record at 192kHz if you want really low latency. It seems like they could have a 32 sample buffer for 44.1 and it would be about equal in performance... but they don't. 
 
That got me thinking about the PCI-E again. :-)

PCIe is for sure the answer for lowest latency. Now, that said different vendors can have different quality of hardware and you can find USB/FW from one vendor that is better than PCIe from another. However all other things being equal, PCIe will win. It is a bus that is very close to the processor and has very little overhead, hence the ability to have low latency.