Re: [PATCH v8 01/31] gpu: nova-core: Hopper/Blackwell: basic GPU identification

[John Hubbard]

On 3/25/26 8:36 AM, Gary Guo wrote:
> On Wed Mar 25, 2026 at 3:52 AM GMT, John Hubbard wrote:
...
>> +    // Blackwell
>> +    GB100 = 0x1a0,
>> +    GB102 = 0x1a2,
>> +    GB202 = 0x1b2,
>> +    GB203 = 0x1b3,
>> +    GB205 = 0x1b5,
>> +    GB206 = 0x1b6,
>> +    GB207 = 0x1b7,
> 
> Sashiko asks if it'll be problematic that 0x1a and 0x1b are both Blackwell.
> https://sashiko.dev/#/patchset/20260325035242.368661-1-jhubbard%40nvidia.com
> 
> Can you confirm that it is indeed actually okay and is a false positive?
> 

Hi Gary,

Sashiko is correct, good catch. GB10x and GB20x have distinct hardware
architecture bytes (0x1a and 0x1b respectively), but I only had a single
Architecture::Blackwell = 0x1b. That means GB100 and GB102 would fail
probe with ENODEV.

I'm splitting it into Architecture::BlackwellGB10x = 0x1a and
Architecture::BlackwellGB20x = 0x1b, which also lines up with how
Open RM (ARCHITECTURE_GB100 / ARCHITECTURE_GB200) and Nouveau
(GB10x / GB20x) handle it. Will be fixed in v9.


Thanks,
John

thanks,
-- 
John Hubbard