AM2634: Cortex R5F core GFLOPS value

Hi Kenley,

The ARM Cortex-R5F floating-point CPUs operating in lockstep offers 1.66 DMIPS/MHz. The DMIPS value of AM2634 operating at maximum CPU speed is 1.66*400Mhz = 664 DMIPS. 

The AM2634 has two dual Cortex-R5F clusters and can run in either lockstep (ASIL-D capable) or non-lockstep (ASIL-B capable) modes. For lock-step (ASIL-D) operation each cluster is capable of 664 DMIPs lock-step performance. If two clusters are all configured for lockstep operation you would have a total of 664*2 DMIPs lockstep available. If all two dual R5F clusters run in non-lockstep mode, each cluster is capable of 664*2 DMIPs and 2656 (664*4) DMIPS in total. For AM2634, each cluster can be independently configured so you can choose where you want to use lockstep mode and where you don't. 

You can estimate the MIPS using CPU speed and  stages of pipeline: 400MHz * 8 stages = 3200 MIPS. To get an accurate MIPS for a particular function, the FPU can be used to measure the cycle count used by the function, and count the number of instructions of the function in disassembly window, the MIPS=CPU frequency*#of instructions/cycle counts. I don't have a estimated number of FLOPS for Cortex-R5F. 

Hi Wang,

Thank you for the support!

Do you have any suggestion on how to decide whether AM2634 could replace the DSP of OMAP-L138 or not ?
Could we know if customer tell us exactly what kind of operation the is DSP doing for?
Or should customer test it on real evaluation to know it is possible or not ?

Thank you in advance.

Best regards,

Kenley

Hi Wang,

As additional information, customer is using their DSP to calculating trigonometric function.

Best regards,

Kenley

Hi Wang,

Do you have any updates?

Looking forward to your feedback.

Best regards,

Kenley

Hi Kenley,

Sorry for the delay in replies here. Using the AM263 for DSP operation like this isn't something we'd be able to give any assurances on so it would need to be driven by customer evaluation for their specific use cases. In general for an MCU part, AM273 might be a little closer because it has an actual DSP available as one of the cores. But I don't think we have any data on the performance of that DSP vs the OMAP one.

After having asked around about this sort of OMAP device, the general best fit would be more on the Processors side with a Cortex A core as long as the peripherals line up. I've asked someone who can speak more to that to provide some details for you too.

Best Regards,

Ralph Jacobi