AMD: Zen vs Naples, Zen2 vs Rome

[boegel]

I noticed that we're currently using zen as codename for 1st generation AMD Epyc (also known as "Naples"), and zen2 for 2nd gen AMD Epyc (also known as "Rome").

How was the decision made to use zen rather than naples and zen2 rather than rome?
Wouldn't using naples and rome fit better compared to the Intel code names?

Changing this in place now may be painful, depending on how strict we are w.r.t. keeping the current names at this stage.
So how about adding an alias field, where we could add naples for zen and rome for zen?

Applications leveraging archspec could then decide for themselves how to leverage those aliases (ignore them, use both, prefer using the alias instead, symlinking one to the other, etc.).

Maybe alias isn't the best naming, perhaps codename is better (since "Zen" is the correct name for the 1st gen AMD Epyc microarchitecture)?

[tgamblin]

We used zen, zen2, etc. because that’s the microarchitecture name. Naples and Rome are the names of corresponding server parts, but they’re more specific. For example, there are also Matisse and Castle Peak, which are the client and high-end desktop zen2 products.

So we picked zen* for consistency, it’s just that people more frequently refer to AMD uarches (in HPC at least) based on the server product name than they do by the uarch. Intel OTOH uses incomprehensible names for specific parts.

I like the alias idea a lot, as it would let us reduce the confusion. HPE neé Cray for example uses the server names in their modules for AMD, and the uarches for Intel stuff.

I don’t quite know what to do for output, as mapping everything to the uarch is easy but deciding to map the uarch to one of many specific names is not. I think archspec would always just say zen2 even if the user said Rome initially.

[boegel]

Thanks for clarifying (I expect this question to come up from others too, so having this documented here is good).

With that in mind, archspec should always say zen2 when it's asked for the label of the uarch, that's clear.

But maybe it should provide support for returning the list of aliases for the current uarch, so whatever is using archspec could leverage that information somehow (creating a symlink from rome to zen2, for example).