Hi,

I'm trying to implement a highly linear current mirror as a transconductance stage of a passive current-mixer. The mixer is driven by a current source that drives current into the mirror input. The output current (from the mirror) is then driven into a load which the passive mixer then (voltage) switches into the load of the mixer (resistive, small). This is for a transmit chain.

I've tried some basic topologies (e.g. source degeneration) which give me decent results for linearity (> +30dBm), but it burns up a lot of current and is noisy. Could anyone suggest a very linear (>30dBm) transconductance or a means to further linearise a current mirror and yet achieve <2nV/√Hz (<-160dBm/Hz) output noise floor with some small gain ? Or even any seminal papers or interesting designs/topologies ?

I know this isn't all of the information about my design, but I would be happy to answer any questions and would welcome any suggestions !

cheers

A current source, by design is high impedance.
High impedance = high R equivalent for noise.
Current source will always be noisy, due to being a current source!

sandman wrote on Dec 24^th, 2011, 4:12pm:


For noise, I suggest starting out with the equations and substituting 2*I/(Vgs-Vt) for gm in your final equations. This will give you a physical feel for the trade-offs that you have. Also, you should be able to look at the simulation results and determine the contribution of each device in the circuit after doing your noise analysis.

Current mirrors are pretty basic: the lower the gm, the lower the noise they produce. If you lower the gm by decreasing the width, the mirror will be faster. If you lower gm by increasing the length, the mirror will be slower. Physically, lowering gm means making Vgs-Vt as large as possible for a given current so a low noise current mirror will need lots of headroom.

Alternatively, resistive degeneration accomplishes the same thing, and is even better for 1/f noise. Drop as much voltage as you can stand across the resistor. I do not know why this didn't work for you, but you must be misinterpreting the results or describing the circuit incorrectly to us. Furthermore, your current consumption should stay the same - I suspect you are doing something wrong as the current consumption increase should reduce the noise, assuming you kept the gain the same.

The noise could also be generated before the mirror: garbage in, garbage out.

I assume finite output impedance of the mirror is causing your linearity problem. Resistive degeneration will help as you found out, as will a cascode. The former will reduce noise,; the latter will add negligible noise for a given operating point, but the bigger impact of the cascode will be loss in headroom which will limit how much vgs-vt you can have. In other words, you will lose headroom by adding the cascode, forcing you to reduce vgs-vt of the mirror (or the IR drop if degenerated). The loss in headroom will force a design with higher noise.

rg

Hi sandman,


just wondering why do you have a current input in the first place? Is the entire transmit chain current-mode?

regarding your question, you have two figures that need optimization: noise and linearity. So I guess you need to do some kind of study to see which type of current mirror offers the best trade-off.

Alternatively, convert your current into a voltage using an op-amp in feedback. It can be a simple op-amp design to minimize noise.


Aaron

Hi Raj,


Quote:


I suppose that different current mirrors offer different noise and linearity tradeoffs. For example, a cascode current mirror may be more linear since under large signal conditions, it is less affected by the output voltage (total Vds). However, because more devices are involved, it is likely to be noisier than a simple current mirror. So the question becomes, for the same noise budget, which current mirror offers the highest linearity? There are of course other tradeoffs.

My comment about possibly using an op-amp refers to the fact that a transconductance stage is needed at some point anyway. Unless you have a current-output DAC and a current-mode transmit chain. Not sure exactly how that would be implemented.


regards,
Aaron

loose-electron wrote on Dec 25^th, 2011, 4:03pm:


Hi loose-electron,

I have a current-source (sink rather!) DAC at the input of my current-mirror (CM) amplifier. The input current noise of the CM amplifier is amplified by the Gm (including degeneration,) of the diode connected device and then the mirroring ratio (>1 assumed).

The output current noise of this combination sees the parallel combination r_out (small signal) of the CM and the R_load, which is typically smaller than r_out, resulting in a smaller output voltage noise (than across r_out alone). I wonder then if what you say about the r_out still holds, taking the load into picture.

Hi Aaron,

aaron_do wrote on Dec 27^th, 2011, 7:20am:


I have  current-'sink' DAC that drives modulated signal current into the CM. The CM amplifier drops the current across a load. The voltage across this is switched by a passive mixer. This is for a Tx. chain, so the mixers draw no current (theoretically).

I set out to explore the noise equations for a CM amplifier by hand but this has be expected-ly slow. A couple of papers treat noise in CM's but only through simulation results.

Interestingly, I haven't come across any papers or articles on the linearity of the various current-mirrors. Any suggestions ?

I did consider using op-amps at the input of the mixer. But, I needed some gain control and an output noise lower than -160dBm/Hz. Some basic analysis ruled out an Op-Amp configuration where I tried to enhance linearity.. What kind of structure did you have in mind ?

sandman wrote on Dec 30^th, 2011, 9:33am:


Where is the hangup? The noise source should just be current sources in parallel with the devices in the mirror. The analysis is pretty straight forward from there. As I mentioned earlier, about all you can do to reduce thermal noise is increase the overdrive (i.e. Vgs-Vt).

Raj - I don't think you want to use an opamp; it would be overkill, limit bandwidth, use extra power, add noise, etc. Apparently he can get the linearity with resistive degeneration so in the worst case he should be able to do far better with a cascode device.