• State Resolved
  • Locked Locked
  • Replies 6 replies
  • Subscribers 99 subscribers
  • Views 925 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AM5716: AM57xx DSP+OCMC vs 66AK2G12+MSMC performance benchmarks

Stephen Turner
Genius 3476 points
Part Number: AM5716
Other Parts Discussed in Thread: 66AK2G12

We are trying to evaluate the performance of the DSP coupled with the internal RAM for the AM 57xx vs the 66AK2G12 for the same DSP clock speed. 

So: AM57xx c66x+OCMC vs 66AK2G12 c66x+MSMC

Does TI have benchmarks for this?.

thanks

  • 0
    TI__Guru* 84075 points

    Hi Stephen

    Thanks for the query. I don't think we have ready made data on this

    I have always struggled with what kind of benchmarks for memory latency make sense for a customer, given these devices go into diverse applications. Some customers want sample by sample load/strore access time , some are looking for DMA performance.

    If you had a benchmark in mind , that may help to see if we can support it.

    For AM57x, you should also be able to look at the following application note

    http://www.ti.com/lit/an/sprac46a/sprac46a.pdf

    This is for the automotive variants for this family, but see if the data in the DSP CPU section is something that you find relevant and if that is the kind of data you would like on K2G? 

    I may not be able to immediately facilitate this, but would like to take the feedback to see what we can do make your device choice easier. 

    Do note that K2G runs at 1 GHz , and only some of the AM57x may offer that. 

    Does max CPU speed matter to you? 

    Regards

    Mukul

  • 0 in reply to Mukul Bhatnagar
    Expert 1680 points

    Hi Mukul,

    I think a couple of benchmarks would be useful.

    1) create a 128 32-bit array in OCMC and copy the lower 64 words to the upper 64 words - this would be a block copy operation.

    2) do an interleaved read/add/store operation. something like

    int a[128];

for(i = 0; i < 64; i++)
  a[i] = i;

for (i = 0; i < 2 ;i++)
  for (j = 0; j < 32; j++)
    a[i + j * 2 + 64] = a[i + j * 2] + 1;

    My 2c.

  • 0 in reply to Andrew Elder
    Genius 3476 points

    I would concur with Andrew

    thanks

  • 0 in reply to Stephen Turner
    Genius 3476 points

    I found this regarding the Keystone II. 

    sprabk5b

    Looking for similar numbers for the AM57x

  • 0 in reply to Stephen Turner
    TI__Guru* 84075 points

    Thanks for the feedback Stephen and Andrew. 

    I am going to assign this to a colleague of mine to see what additional help we can provide. 

    We may be able to get some data on K2G, but may not be able to furnish equivalent data on AM57x at the moment. 

    We also have in house several data points on standard linux benchmarks like lmbench etc, which gauge more of external memory performance and i believe the numbers are comparable.

    Is either AM57x or K2G working out better for your needs apart from understanding the on chip memory access latency? Any additional decision factors that may influence the decision?

    Regards

    Mukul 

  • 0 in reply to Mukul Bhatnagar
    TI__Guru** 116170 points

    Stephen,

    We do have an internal study with benchmarks comparing AM57x OCMC to K2G MSMC in standalone DSP environment which is preliminary data that we are not able to publish on public forums. However, I am going to share broad observation to help you with your evaluation.

    • The memory read/write and copy bandwidth is comparable when the size of the data used was very large (8Mb). For small chunk of data K2G DSP with larger L2 cache performed better. Both CPU and DSP continued to show the same trend as in single core scenario. Eg: CPU has higher throughput on AM57x and DSP has higher throughput on K2G. For concurrent access from ARM and DSP the Memory throughput both min and average was similar.
    • MSMC Memory architecture on K2G maximizes DSP read throughput for the same DDR bandwidth as comparted to AM57x with OCP, L3 and DMM architecture. Since it is the same DSP core on both the devices, write throughput is almost the same. 
                Note: Data on AM572x uses dual core running at 1.5 GHz and K2G EVM at 1 GHz and DDR speeds on K2G was 1333 MTs vs 1066 MTs on AM57x
    Let us know if you want more details or if you think you would prefer to see the numbers for the loop that you provided.
    If you are more focused to have maximum DSP performance then K2G is a better option given higher L2 memory and 1:1 speed compare to A15. AM57x DSP runs max upto 750 MHz and A15 can run upto 1.5 GHz.
    Regards,
    Rahul