AM2634: Benchmarking F28388D vs AM2634 for FOC

Hi Sira,

Thanks a lot for sharing your results; they are very close to what I am looking for, and they have raised a few more questions in my mind. I would really appreciate it if you could help me with the following:

1. I see you used an AM263P, which includes a TMU. I will definitely consider it for my benchmark, thanks! In your opinion, is the AM263P4 the best choice for motor control and real-time applications within the Sitara Arm-based microcontrollers? From the documentation, it looks like the AM263P4 is an improved version of the AM2634. Does it make sense to say that AM263P4 > AM2634 and therefore run benchmarks only on the P version?

2. Are the f2837x results sufficient to compare the AM263P with the F28388? At the end of the day, we are mainly evaluating the C28x processor.

3. How was the benchmark organized? Was the same code used across devices, and if so, how was it adapted?

4. Considering the CPU frequencies, results show that the R5F was about 33% faster than the C28x. Did you also take into account the use of the FPU?

5. I was thinking of using the CLA on the F28388D to run the FOC, offloading the C28x for other double-precision calculations. Would such a combination still be better than using the R5F for both the FOC ISR and the double-precision tasks? On the Sitara, is there a way to replicate an architecture similar to the C28x+CLA pair?

6. Is the code you used for the benchmark shareable? It would be extremely helpful for me.

7. How did you collect these results — specifically, how did you measure the clock cycles? Did you design a repeatable benchmark? Did you use a scheduler or simple interrupts to schedule the tasks?

Thanks again for your time and support!

Best regards,

Pasquale

Hi Pasquale,

Replies in-line below

1. I see you used an AM263P, which includes a TMU. I will definitely consider it for my benchmark, thanks! In your opinion, is the AM263P4 the best choice for motor control and real-time applications within the Sitara Arm-based microcontrollers? From the documentation, it looks like the AM263P4 is an improved version of the AM2634. Does it make sense to say that AM263P4 > AM2634 and therefore run benchmarks only on the P version?

   The P version has a TMU, a hardware resolver, and a few other improvements. It definitely appears to be better for motor control. http://www.ti.com/lit/spradb3

2. Are the f2837x results sufficient to compare the AM263P with the F28388? At the end of the day, we are mainly evaluating the C28x processor.

   Correct. I don't have results specifically for the F2838x but it should be the same, atleast from RAM. From Flash, there could be differences in wait states etc. that impact performance. But my assumption is real-time code like this will be run from RAM.

3. How was the benchmark organized? Was the same code used across devices, and if so, how was it adapted?

   Yes, C code was developed and run on each device. Minor device specific updates were made.

4. Considering the CPU frequencies, results show that the R5F was about 33% faster than the C28x. Did you also take into account the use of the FPU?

   Yes, the FPU was used in both cases. If you take into account just the cycles, the C28x is 33% faster. But if you take the CPU frequency as well, the AM263P is 50% faster.

5. I was thinking of using the CLA on the F28388D to run the FOC, offloading the C28x for other double-precision calculations. Would such a combination still be better than using the R5F for both the FOC ISR and the double-precision tasks? On the Sitara, is there a way to replicate an architecture similar to the C28x+CLA pair?

   Certainly having the CLA to offload control tasks helps boost performance, since you basically have an additional core compared to the AM263P But then again, it depends on which specific part numbers are being compared. On the C2000 side, you might have 2 C28x and 2 CLAs, and on the AM263x you may have 1, 2, or even 4 R5F cores. On the Sitara side, there isn't a corresponding co-processor like we have on the C28 side. Specifically on this benchmark, the C28 is about 60% faster than the CLA, refer to http://www.ti.com/lit/spracw5

6. Is the code you used for the benchmark shareable? It would be extremely helpful for me.

   Yes, it's in C2000Ware at examples\demos\benchmark\aci_motor_benchmark

7. How did you collect these results — specifically, how did you measure the clock cycles? Did you design a repeatable benchmark? Did you use a scheduler or simple interrupts to schedule the tasks?

   The app note spracw5 above describes the benchmarking methodology.

Hi Sira,

Thanks a lot for all your answers!  

Starting from a quote from http://www.ti.com/lit/spracw5: “Many software benchmarks only focus on the processing aspect typically expressed in million instructions per second (MIPS), without full regard for the interaction between peripherals, CPU, and co-processors. Such benchmarks do not provide a full view of the real-time performance capabilities of a system.”, I defined five candidate configurations to compare for a motor-control use case (FOC real-time loop + a separate, computationally heavy algorithm that ideally runs in double precision):

1. TI28388D — C28 standalone (FPU + TMU) running both tasks

2. TI28388D — C28 + CLA (C28 = trajectory; CLA = FOC)

3. Sitara AM263P4 — single R5F running both tasks

4. Sitara AM263P4 — R5FSS0_0 = trajectory ; R5FSS0_1 = FOC (same cluster)

5. Sitara AM263P4 — R5FSS0_0 = trajectory ; R5FSS1_0 = FOC (different clusters)

Questions: 

1. From an architecture and real-time motor-control perspective, which of these configurations would you consider the best choice and why?
2. Does it make sense to run a formal benchmark across all five configurations, or is there a configuration that’s clearly preferable (making the full benchmark unnecessary)?

I tried to answer them on my own and here there is a resume of what I got:

Your feedback would be extremely appreciated! 

Best,

Pasquale