Re: [PATCH v10 03/28] gpu: nova-core: Hopper/Blackwell: basic GPU identification

[Alexandre Courbot]

On Tue Apr 14, 2026 at 6:29 AM JST, John Hubbard wrote:
> On 4/13/26 2:21 PM, Timur Tabi wrote:
>> On Mon, 2026-04-13 at 14:08 -0700, John Hubbard wrote:
>>>> Can you add something to the patch description that explains why Blackwell, unlike all of the
>>>> olther architectures, is represented by two Architecture enums?
>>>
>>> How's this:
>>>
>>> Blackwell uses two Architecture variants because the hardware itself
>>> exposes two distinct architecture IDs in NV_PMC_BOOT_0: 0x1A for
>
> This is really what drove it, for me, though: the hardware itself
> saw fit to identify these two variants as architectures. It's
> encoded into the very fabric of space-time itself! haha :)
>
>>> GB10x (GB100, GB102) and 0x1B for GB20x (GB202, GB203, GB205, GB206,
>>> GB207). Open RM maps these into separate chip families with distinct
>>> HAL routing and different firmware signature sections.
>> 
>> We don't need two arch's for Ampere to handle different firmware signature sections.  Looking
>> through your patches, the vast majority do this:
>> 
>> 	Architecture::BlackwellGB10x | Architecture::BlackwellGB20x
>> 
>> Looking through your patches, it seems overkill to separate these two.  The one or two instances
>> where it makes a difference, an if-statement will suffice.

I have to agree here, if we didn't split then the architecture IDs
`0x1a` and `0x1b` would need to both map to Blackwell, which makes them
impossible to represent correctly using an enum.

And granted, there are lots of

  `Architecture::BlackwellGB10x | Architecture::BlackwellGB20x`

in the code, but that catches the attention only because the variants
share the same prefix. If `BlackwellGB20x` was named something
completely different, it would go unnoticed (just like we often include
`Ada` in the same match arms as `Blackwell`).

So while it does look a bit strange to have 2 Blackwells, it does
reflect the hardware architecture correctly and doing things differently
would result in more complexity/incorrectness. For instance until v8
there was only a single Blackwell architecture with the ID `0x1b`,
leaving `0x1a` completely unmanaged.