vs3, quick question: why do the bitfurys in the nf6 design not run at their top frequency (6x3.5 = 21GH vs ~12.5GH)?
The main reason is voltage.
First, not all chips will get 3.5GH. Each chip is slightly different and will get the best results at a specific voltage and speed (osc6 bits).
The voltage for that is usually in the vicinity of 0.95-1.1V.
With NF6 specifically you can't easily adjust the voltage for each individual chip - they split somewhat even the input voltage which on a theory is supposed to be 5.0V.
In practice I've seen it vary all the way down from 3.8V to 5.3V. Some crappy USB hub power supplies will start dropping the voltage with higher amperage usage.
For example - I just got a very nice SIIG 10-port USB3 hub with 12V/5A power supply - and the results speak for themselves:
bfgminer version 4.5.0 - Started: [2014-08-01 23:46:09] - [ 0 days 00:16:48]
Manage devices Pool management Settings Display options
Pool 0: stratum.bitcoin.cz Diff:47 +Strtm LU:[00:02:50] User:vs3.nanofury1
Block: ...3f31a753 #313602 Diff:18.7G (134.1Ph?) Started: [23:59:20]
ST:25 F:0 NB:2 AS:1 BW:[ 53/ 42 B/s] E:272.54 I:61.47uBTC/hr BS:3.49k
6/23 | 57.15/56.82/55.90Gh/s | A:351 R:0+0(none) HW:331/2.5%
--------------------------------------------------------------------------------------
NFY 0: | 2.20/ 2.34/ 2.30Gh/s | A: 21 R:0+0(none) HW: 27/4.8%
NFY 1: | 4.85/ 4.80/ 4.91Gh/s | A: 26 R:0+0(none) HW: 19/1.6%
NFY 2: | 4.92/ 4.93/ 4.86Gh/s | A: 27 R:0+0(none) HW: 10/.87%
NFY 3: | 16.56/16.79/16.11Gh/s | A: 85 R:0+0(none) HW:156/4.0%
NFY 4: | 15.85/15.88/15.63Gh/s | A:106 R:0+0(none) HW: 89/2.4%
NFY 5: | 12.74/12.69/12.69Gh/s | A: 87 R:0+0(none) HW: 35/1.2%
--------------------------------------------------------------------------------------
#0 is a NF1, 1 and 2 are NF2s.
#3 (56 bits) and #4 (55 bits) are NF6 both with Gen2 chips, #5 (56 bits) is a NF6 with Gen1 chips.
So, if you want to get a few extra hashes you need to increase the 5V voltage. With 5V there are about 4.9 to be split among the chips, or that gives you 0.817V per chip.
If you increase that to 5.3V that will give you 0.87V per chip and you should easily be able to get over 3-3.2GH/chip.
You can try increasing it even further, but there is a risk that you may damage the USB hub or your computer. There is also the risk of overheating, so you have to be careful. Increasing it to 5.5V will give you 0.917V/chip and with that you should be able to get even more hashes.
Aside from that - what the others wrote is also valid to some extent - standard USB2 port provides 0.5A and USB3 is 0.9A. So in either case you will need a powered USB hub and you need to make sure that you have at least 1.2-1.3A per NF6 device (and 1.5A or more if you're going to be pushing it).
The MCP2210 also contributes to the limitations - it adds a gap between the transmissions which limits you how quickly you can scan all chips. If you search in the previous posts I think there was someone who experimented with longer chains and basically started hitting the limits after the 8th chip. In my guesstimate the limitation will be at around 20-25GH.