i.MX 8 asymmetric read and write speeds for GPU and NPU

[aheck]

Hi all,

We are planning on using the i.MX 8M plus in our product, and we would like to use the NPU and GPU for data processing via OpenCL.  When testing the GPU and NPU I/O speed on the evaluation kit, we see a large difference between the read speed and the write speed.

Specifically, we are able to write to the GPU at ~3 GB/s, and a read back at ~0.3 GB/s.  We saw similar speeds when measured with our own C++ app monitoring OpenCL events, and this utility (https://github.com/ProjectPhysX/OpenCL-Benchmark).

Is it the expected behavior that writing to the GPU and NPU is about 10x faster than reading from them?

Thanks.

[brian14 · Accepted solution]

Hi @aheck, 

Thank you for contacting NXP Support.

No, this is not the expected behavior.
On our internal checks the expected behavior is a symmetrical read and write speed.
That asymmetric read and write speed could be influenced by the type of benchmark.
NXP uses official benchmarks provided by The Khronos Group.

Have a great day!