The Release Notes for Cuda Toolkit 11.8 state:
This release introduces support for both the Hopper and Ada Lovelace GPU families.
Looking at the Programming guide instructions throughput data, there is only an entry for 9.0.
Does this mean there is no differentiation between Hopper and Ada, as there is for Ampere, (8.0 vs 8.6)?
My interpretation is that the cuda toolkit “binary blob” i.e. the CUDA toolkit download with associated drivers and libraries, is designed/intended to support both early members of the hopper family (i.e. primarily H100) and early members of the Ada Lovelace family (all currently announced members, presumably).
w.r.t. the documentation, the 9.0 SM architecture evidently does not represent Ada Lovelace GPUs that are based on the publicly announced chips such as AD102 (see the whitepaper). It seems evident (to me) that the 9.0 arch description mentions 64 FP64 cores, and no mention of those is made in the aforementioned ada lovelace whitepaper (&). I doubt that is a careless omission. So it seems evident to me that the 9.0 architecture description does not match various members of the Ada Lovelace family GPUs.
((&)In fact, in a note to figure 1, the whitepaper explicitly states that the AD102 SM contains 2 FP64 units per SM, so it is evident there is a mismatch against 9.0 SM description.)
Unfortunately without documentation I can’t go beyond that. Although supported, the Ada Lovelace SM is not yet documented to the same extent as the 9.0 SM.
Indeed CC 8.9 is exposed if you run nvcc.exe --help
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.