We are running a survey to help us improve the experience for all of our members. If you see the survey appear, please take the time to tell us about your experience if you can.
Hi,
We are working with a custom SoC containing an M4, an Ethos U85 (2048MACs, SRAM_512KB and default memory config), a 4MB of SRAM and 256MB of DRAM (We will implement the m85 afterwards). We are using TFLM runtime to deploy AI models. Currently, we have deployed successfully an KWS and Object detection models, we are moving to LLMs and starting slow with MobileBERT EdgeTPU version.
After initializing the model, tensor arena, scratch buffer successfully, the model is invoked, the Ethos completes the cmd stream, an IRQ is triggered but the CPU doesn't get triggered and never enters the IRQ handler which blocks the CPU at the ethos_semaphore_take(). (We are using all the weak implementations of the Ethos driver)
We didn't experience this in the other applications and wonder what could be the problem.
I couldn't include any embeddings to explain more the use case.
Looking forward to hearing from you
Ahmed
SpiNNcloud