amd gpu mi300x : modelling different XCD

[lkartik2212]

Dear gem5 developers, thank you for the amd gpu models in gem5. I am trying to simulate an amd mi300x gpu and I am new to the toolchain.

In mi300x architecture(please see figure), there are 8 XCDs(Accelerator Complex Dies) with 40 CU's in each XCD.

I see configuration files here, there are 40 compute units

 # See: https://rocm.docs.amd.com/en/latest/conceptual/gpu-arch/mi300.html
    # Topology for one XCD. Number of CUs is approximately 304 / 8, rounded
    # up to 40 due to gem5 restriction of 4 CUs per SQC / scalar cache.
    args.num_compute_units = 40
    args.gpu_topology = "Crossbar"

Questions about the simulator model:

1. But the number of XCDs parameter is not there that should account for all 304 CUs. So does the model only account for 1 XCD in mi300x? or am I missing something here?

2. Also, does it model the vertically stacked LLCs in I/O die in mi300x?

Kind Regards,
Kartik

[abmerop]

The topology assumes just one XCD. When modeling multiple chiplets/stacked LLC it is generally just a matter of setting up a topology with varying latencies for different types of links (e.g., LLC link latency is reduced but chiplet-to-chiplet is higher than intra-chiplet).

[lkartik2212]

thanks for the quick response :)

[mattsinc]

Thanks for answering quicker than I could, @abmerop . Yes, @lkartik2212 this is something we're working on but don't have validated models yet. It works in the way @abmerop described though. Happy to coordinate with you if you want to take lead.