婆罗门
精华
|
战斗力 鹅
|
回帖 0
注册时间 2008-7-8
|
楼主 |
发表于 2026-9-6 21:19
|
显示全部楼层
Not because of the gates — those are just constants. Because the DLSS 4 snippet ships no Ampere machine code. Its 70 fatbins carry PTX sm_89 ×70, PTX sm_120 ×31 and cubin sm_89 ×31, and nothing for sm_80/sm_86. PTX is forward-compatible only, so sm_89 PTX cannot be JIT-compiled down to sm_86; the module load fails outright.
Retargeting is theoretically open — the kernels use only mma.sync m16n8k16/m16n8k8 FP16 and ldmatrix, with zero instructions newer than sm_86 (no FP8, no wgmma, no TMA), and the old hardware optical-flow dependency is gone in DLSS 4. But frame generation costs roughly a fixed amount per generated frame, and Ampere has far less FP16 tensor throughput per SM, so the generation pass would likely cost more than the frame it saves. It was investigated and deliberately dropped.
社区里的解释 |
|