674x questions on power and benchmarking

Hi Steve

Steve Sager said:

1. Is the power  vs.  core frequency curve linear? (customer will run very slow for most low end

applications and they would like the power to drop accordingly)

The active power will follow the CV^2F formula in general, and you will see power scale with frequency. However, you would need to be careful on how you gauge Power vs Frequency. If you talk about "Total Device Power" , its going to be made up of  Active power (CV2F) and Leakage Power ( f(V,T, process), and within active power, it is a sum total of the core logic power and IO power, and within the c6748 you have different modules and IOs running at different frequency ( e.g. c674x DSP @ 200 MHz (c6742), EDMA, infrastructure components at half that speed, and some other modules at 1/4th that speed etc) so maybe for coarse estimates you could scale the entire power with device frequency,  for higher granularity overall power will scale slightly differently. I hope this makes sense? I have also attached a trimmed down and slightly modified version of a TI Tech day presentation that was presented by a colleague of mine on OMAPL1x/c674x power, you might find the last couple of slides helpful to address the questions from your customer. The data in slides was preliminary, we plan to have power estimation spreadsheets for OMAPL1x/c674x devices in coming months, and hopefully they will help addressing these questions in a more precise manner.

Steve Sager said:

Can the OMAP 138 kit be used to benchmark the 6742 properly?  What other options do we have?

I would think this is still your best option of evaluation an assessment. OMAPL138 is our super set offering (ARM+DSP), c6748 is our DSP only version  (very similar peripheral set and operating points) , and c6742 as a subset of c6748, with reduced peripheral set, and lower performance. The OMAPL138 EVM bundle comes with both the OMAPL138 and a c6748 SOM, so you could use either of them for benchmarking. If you have access to c6748 SOM (DSP Only), thats probably better.

Steve Sager said:

I found the benchmark code on the main TI site.  Are there actual output performance numbers available?  (looking for FIR filter specifically)

Can you please specify where you found the benchmark code and what collateral on TI website you were referring to?

Have you also seen the c674x DSPLIB http://wiki.davincidsp.com/index.php?title=C674x_DSPLIB

The software package available for DSPLIB, includes an xls spreadsheet with performance data on all the kernels, including some FIR filter implementations.

Hope this helps.
Regards

Mukul

Hi There,

Below is the link from the TI site to teh benchmarks I was talking about.  I had not looked at the Wiki pages.

http://focus.ti.com/dsp/docs/dspplatformscontento.tsp?familyId=1622&sectionId=2&tabId=2431

thanks for all the information!

Steve