AM2434: R5F core vs C6000 DSP cores

Daisuke Maeda said:

The following benchmarks show the performance of the A53 and R5F cores for AM64x, but it is not clear if VFPv3 is used for floating point operations in the R5F core.

AM64x Benchmarks
www.ti.com/.../spracv1.pdf

These benchmarks are compiled using the VFPv3-16 full implementation which is enabled on AM64x and AM243x parts.

Daisuke Maeda said:

Could you provide more information for performance comparison between R5F core and C6000 DSP cores?

In signal processing C6000 is a significantly higher end core in both performance and power consumption. Per clock cycle with SP float C66x core is 3-10x faster than R5F. Difference is ~10x in FFT, and ~4x in dot product and FIR, ~3.5x in IIR or Lattice filter. In control code, like state machines or pointer chasing measured by Dhrystone or Coremark the difference is much smaller, ~x1.5 to 2x.

  Pekka

5x makes sense. To get to my example 10x difference between C6x and R5. Somewhat depending on the size of the FFT and the level of optimization put in (for example optimize for power of 2 sizes, or optimize for what fits in R5 TCM or C6x L2 or L1D SRAM, DMA ping pong usage), the final performance steps can take effort and perhaps focus on the specific FFT needed on especially C6x. And C66x does have some advantages over C674x in 32-bit floating point.

  Pekka