DDR usage issues on FVP (SSE-320)

ML Embedded Evaluation Kit
#1

Hi, currently I am using the inference_runner from MLEK on FVP to run the 900MB TFLITE model compiled by Vela. I modified the LENGTH of DDR in the /ml-embedded-evaluation-kit/scripts/cmake/platforms/mps4/sse-320/mps4-sse-320.gnu.ld file to 0x50000000. After simulation on FVP, the inference of the ethos-u operator failed and the error message was as follows:

ethosu_device_u85.c:177: NPU status=0x00000000, qread=4, cmd_end_reached=0

From the FVP document, it is seen that DDR has a total of 2G of memory. However, it is for secure and non-secure cross-operation. May I ask, is this error caused by the model being too large and crossing the S/NS area, resulting in the error message? Is there any way to successfully run a 900MB-sized large model?

#2

Hi @lllccc, thanks for bringing this to our attention. In theory this should be possible but we will need to investigate it further to understand the issue you’re seeing - we currently don’t use models of this size in our test coverage.

I wanted to mention that this discussion forum will be set to read-only from November 1st. The new home for the MLEK project is on Arm’s GitLab and we are encouraging users to direct questions etc. to the GitLab Issues page here: Issues · artificial-intelligence / ethos-u / ml-embedded-evaluation-kit · GitLab.

If you don’t receive a reply on this thread before the end of the week, please reach out to us on GitLab Issues so we can help further! Thanks