Resurrecting 2013 GPUs: Dev Ports PyTorch 2.7.1 and Native CUDA Mamba-1 to Kepler-Era Cards
TheSpicyBoi123 · reddit · 2026-11-09
A Reddit developer has ported the modern LLM training stack to legacy NVIDIA GPUs (SM 3.5–7.5, Kepler through Turing) on Windows:
- torch-on-clunkers: custom PyTorch 2.7.1 build compiled with CUDA 11.8, preserving legacy SASS
- mamba-on-clunkers: Mamba-1 with pure native CUDA kernels, validated on a 2013 Tesla K40c with max absolute error of 3.8e-06 vs the PyTorch reference
- Also includes ported causal-conv1d and Windows NCCL (DDP) for multi-GPU training across Kepler cards
- Mamba-2/3 require Triton/TileLang and only run on SM 7.5+; older cards automatically fall back to pure-CUDA Mamba-1
The author is asking the community to test on Maxwell and Pascal hardware, and plans to backport Mamba-2/3 to legacy architectures. Based on the original work by Tri Dao and Albert Gu.
More from Infra
- Startup Mirendil claims a Kimi K3-level model trained by one researcher for under $10M — Ok_Warthog6007 · 2026-11-09
- SkyFi thermal satellite imagery reveals which data centers are actually powered up — bradneuberg · 2026-11-09
- Epoch AI: 2027 chip supply could run 30–170M frontier AI agents, or 1.9B cheaper ones — Dr_Singularity · 2026-11-09
- White paper: moving beyond the AI compute arms race as big tech faces historic spending losses — ReportProfessional57 · 2026-11-09
- LLMs max out hardware: former xAI exec reports failing PSUs, tripping breakers — MParakhin · 2026-11-09
- Local LLM reality check: $3K rig with 32GB GPU handles simple tasks only — 2C_ornot2C · 2026-11-09