dlssg_for_sm86 (github.com/sdli1995/dlssg_for_sm86, 4102 stars, created 09-07) enables NVIDIA's official DLSS frame generation on RTX 30 series (SM86) and RTX 20 series (SM75) GPUs. NVIDIA officially locks DLSS-G to RTX 40 and above; this plugin embeds the stock runtime via a proxy DLL and reports an RTX 50 architecture to the game, bypassing the check so the original frame-generation kernels run unmodified. The author's benchmark claim: since version 0.3.2, the RTX 3080 Ti's generated frames are bit-identical to the RTX 5070's — not a simulation, not an approximation, but the same output from the same kernel.
Hands-on testing (on RTX 30 series, for reference)
Before testing, my view of frame generation on 20/30 series was "a nice-to-have, not a lifeline." After actually testing it, I think the gain on 30 series is substantial.
Without frame generation, pushing image quality to high settings left you at 40-50 FPS — playable, but uncomfortable. With frame generation, that becomes 70-80 FPS at 2X, and over 100 FPS at 3X. And because frame generation is tied to Reflex, you'll notice something surprising: in some games that have Reflex off by default, enabling frame generation not only raises FPS but also significantly lowers PC latency.
In my view the biggest beneficiaries right now are the 3080 and 3080 Ti, since they have more VRAM. The 3070's 8 GB will be a serious constraint — frame generation itself consumes some VRAM.
Installation is straightforward for most games — easy enough for beginners — though a few titles need extra fiddling, such as registry edits. Also note: avoid using this in online games with anti-cheat; getting flagged as a cheater would be no fun.
Editor's fact-check (verified against the repo README before publication)
The hands-on results match the README. Several details from the README are worth expanding:
1. The "bit-identical" claim is the hardest evidence here. Version 0.3.2 rewrote the inference kernel (runtime 310.9); the 3080 Ti's output is bit-identical to the RTX 5070's, and the 2080 Ti's output is bit-identical to the 3080 Ti's — same input, and the 20/30/50 series emit the same bitstream. Four optimization levels are documented: Optimized=0 stock kernel, no acceleration; 1 accelerated but still bit-identical (default); 2 lossy image kernel, PSNR above 50 dB; 3 fully lossy-fastest. The lossless/lossy boundary is stated in the config file, not hidden in marketing.
2. There is an official VRAM table. Frame generation's own overhead depends only on output resolution, not the multiplier — 2X and 6X cost the same: ~350 MiB at 1080p, ~540 MiB at 1440p, ~810 MiB at 4K. 3070 owners can do the math: game + upscaling + 350-540 MiB, and see what remains — more precise than the vague "8 GB will be limited" feeling. That's why the author says the 3080 (10 GB) / 3080 Ti (12 GB) benefit most — the math works out.
3. Multiplier limits: 2X/3X/4X by default; runtime 310.9 supports 6X (requires in-game Dynamic MFG plugin support). Confirmed working on a 20-series 2080 Ti with bit-identical output; the author hasn't benchmarked its performance yet.
4. Different route from Nukem: this is not the "dlssg-to-fsr3" approach (substituting AMD FSR3 frame generation for DLSS-G). It runs the original kernels and merely bypasses the generational check. DLSS-G's CUDA kernels were always capable of running on 30-series optical flow accelerators — the only barrier was a single architecture check. The author states kernel optimization has reached its current optimum; future work focuses on INT8 optimization of the Transformer upscaling model, aiming to raise base framerates on 20/30 series — i.e., increase the "real" frames being interpolated.
The anti-cheat warning stands as written: single-player games are fair game, but don't use it in anti-cheat online games — getting flagged isn't worth it.