What is ENQCMD/MOVDIR64B submission latency in Intel Data Streaming Accelerator

[yslys]

Hi all,

I would like to know in detail about the submission latency of ENQCMD/MOVDIR64B instruction for submitting 64-byte work descriptors in the context of Intel Data Streaming Accelerator. Based on what I know, it is a non-posted 64-byte MMIO write, so theoretically it should be around tens of nanoseconds.

However, I found this paper: @ https://arxiv.org/pdf/2305.02480, in which Figure 5 indicates that the submission latency is around 5us. I think this is too slow compared to a normal MMIO write. That is why I am confused.

Could anyone please tell me what is the latency starting from user issuing the instruction ENQCMD/MOVDIR64B to the device actually received the 64B work descriptor? May I know if around 17ns is a reasonable estimate? If not, what is a reasonable estimate?

Thanks in advance!