How to optimize your DSP power budget


Multicore DSP implementations require intelligent power management given the plethora of voltage and current requirements for core, memory, I/O, and other rails. A key performance benchmark for the DSP’s core voltage supply in particular is the ability to adjust VCORE on-the-fly according to the DSP’s usage case scenario and environmental condition. The VCORE command is typically provided in digital format and should be readily interpreted by the power supply. The VCORE rail normally has the highest current specification, and a small form-factor power solution balancing efficiency and size is vital. The key is to use a low-cost interface between the DSP and an analog PWM stage to accomplish this voltage-identification (VID) function.

With that in mind, illustrated below is a multicore DSP with core rail denoted as CVDD. As well, I wrote an article on the topic in EDN magazine “Optimizing DSP power budget by adjusting regulator,” that goes into this more in depth.

A 500-kHz buck converter rated at 15A powers CVDD. This design exploits VID control using a 4-wire digital interface to a VID programmer that interfaces directly to any analog power stage or controller. Watch a video demonstration of the VID programmer here.

Schematic of KeyStone DSP and VID-enabled synchronous buck converter

 

LM10011 and Analog PWM Power Stage

The LM10011 captures the VID information presented on the DSP’s VCNTL interface and sets the output of a current DAC connected to the feedback (FB) pin of the power stage circuit. In 6-bit mode, 64 current settings with 940-nA resolution and better than 1% accuracy are available. In this example, CVDD is arbitrated by the DSP to a level between 0.9V and 1.1V with step resolution of 6.4mV. Level translators or glue logic are not required, and resistor RSET dictates the CVDD voltage at startup. The LM10011 can interface to any voltage-, current-, or DCAP-mode PWM regulator with a FB input.

Feel free to leave your comments on what other issues you face while designing your DSP power solution.

Resources:

  • Read “Optimizing DSP Power Budget by Adjusting Regulator” in EDN.
  • Review an LM10011 design in the Powerlab™ reference design library.
  • Order the LM10011 evaluation module.
  • Choose power solutions for processors and FPGAs here.
  • Start a design now with WEBENCH® Power Designer.
  • Hello Timothy,

    I am looking to power the Keystone 1 (C6655) using the LM10011. The EVM uses the digital version for voltage scaling, but the app note on Keystone 1 mentions I can use either Analog or Digital. Do you have a particular preference for CC6655? Also, the buck that goes with LM10011, I see we did testing with the high current LM21215. In my application, I need only 5A out (maximum, for the most part 3A). The app note again mentions any buck with compatible feedback can be used, so I assume any buck with external compensation. Do you have any suggestions and do you know if any testing was done with lower current bucks?

    Thank you very much for your help,

    Natallia  

  • Hi Natallia,

    Yes, the LM10011 will work with any converter with a FB pin. As mentioned above, current from the LM10011’s IDAC_OUT pin is injected into the FB node to adjust VOUT as required.

    Choice of a suitable 5A buck regulator depends on the input voltage operating range as well as efficiency and size requirements. For example, the LM21305 current-mode controlled 5A synchronous regulator operates from 3V to 18V input, VOUT can be as low as 0.6V, and it has switching frequency synchronization, precision enable, PGOOD, etc.

    You may also want to check out design example 2 provided in the LM27403 datasheet. Finally, you can review the hardware design guide for Keystone devices at www.ti.com/.../sprabi2c.pdf

    Regards,

    Tim

  • Hi Tim,

    I came across this article while looking for info on the LM21215.

    I have a graphics board and I am using 4x of the LM21215A-A1.  AFAICT, the only difference between the 21215 and the A-1 is the frequency.  We chose the A-1 thinking that the plain one would go away.  BTW, the handy apps info that you have for the 21215 web page is not duplicated in the A-1 page.  I am using an LT 4-phase clock generator to try to minimize switching noise and that seems to work pretty well.  

    I will try you LM10011 idea as right now I am using an 74T07 solution to implement a 2-bit voltage selection.  A bit mickey-mouse but it works.

    I am doing a rev to our board and thought I should just re-run all the comp values since I thought there might have been some improvements. What a mistake!!

    Webench is really flaky - sometimes is saves the new values that I plug in and sometimes it doesn't.  Also, it usually forgets custom values of inductors and caps.  I don't understand why Webench doesn't have a real "save" button that will truly save a design as it sits. And, restore it correctly without recalculating things and blowing up the values.

    Also, I think there might be a bug:  if I run a load simulation with Vin 4.75 to 5.5 and Vout set to 1V@15A, the transients are on the order of 10s of mV, and if I add more caps they smooth out as you would expect.  BUT, if I set Vout to 0.9V, the transients are 100s of mV and they don't respond to added caps the same way.  And, if you add a ton, say, 20x100uf (or fewer but higher cap values), the program generally fails.  In these cases I tell the custom cap to be ceramic.

    Also, if I enter a voltage value of, say, 1.13, the simulator just goes nuts.  It acts like I told it 1000s of volts.

    What do you think?

    Thanks!!

    Victor

  • Hi,

    One other question:  can the LM21215 quick start spreadsheet be used with the LM21215A-1?  If not, why not?

    Thanks!!

    Victor

  • And, yet one more question:  playing around with the quick start calculator, I see that if I want to minimize the transient voltage droop that once I go beyond about 250uF, the phase margin goes infinite at somewhere beyond 200KHz.  I don't think that is something I want.

    If I start reducing the crossover frequency, I have to set the desired to 19KHz to get the phase margin to go to 0 instead of off the chart.  My expected values are then 22KHz and 68 degrees.

    So, this would seem to imply that in order for me to get the transient droop down (700uf >> 24mV droop) is not a realistic approach because it forces the crossover frequency really low.

    I tried hand tweaking the values of Cc1, Rc1, Cc2, Rc2, and Cc3 and with a lot of fiddling and for 5Vin, 0.9V@15A and 700uF and 0.68uH, I got the values 1.1nF, 34K, 39pF, 0.18K and 2200pF.  This gave me 74KHz crossover and 50 degree phase, AND a 23mV transient droop.  

    I am not sure why the spreadsheet couldn't figure this out automatically, but oh well.  

    What do you think?

    Thanks!!

    Victor

  • Hello Victor,

    As you already know, the LM21215 has fixed frequency and adjustable current limit while the LM21215A-1 has a SYNC input. Note that the LM21215A-1 is quite similar to the LM21212-1 with the exception of a higher current limit and a different nominal operating frequency (with no SYNC signal).  Even though the LM21215A-1 quickstart tool isn't yet available, while noting these differences you can refer to the LM21212-1 file at www.ti.com/.../lm21212-1design-calc to conveniently calculate compensation component values. Keep in mind that DCM operation exists at light loads - and this may affect transient response behavior.

    Regards,Tim

  • Hi Tim,

    OK, thanks for the tip about the LM21212.

    One thing that concerns me is that the models used in Webench and quick start clearly are different. If you do a run Webench using a particular set of parameters and then do the same thing in quick start, the resulting phase margin, ripple, and crossover  values are quite different.

    I have tried this for the 21212, 21215, and 21215-1 and I get the same sort of behavior.  I noticed the same thing with the TPS56121, which we also use, comparing between switcherpro and webench .

    Quickstart (or switcherpro) is a lot easier to use, but if Webench gives more correct results, then I will use that.  

    What say you?

    Thanks!!

  • Hi Tim,

    If you have an email, I could send you the webench and quickstart outputs for identical circuits and you could see what I mean.

  • Victor, please open and submit an e2e post for this.Regards, Tim