GeForce unlock modules double training math: NVIDIA sells one chip at two rates
julsimon · x · 2026-10-29
Julien Simon digs into the open-source Linux modules that unlocked GeForce training performance:
- How it works: the modules load before NVIDIA's driver and flip a setting, letting an RTX 4090 issue 16-bit matrix math with 32-bit sums (what mixed-precision training needs) at roughly 2x; one owner measured 1.9x per clock cycle on an RTX 3090.
- A setting, not silicon: NVIDIA's own whitepapers have long rated GeForce at half speed for 32-bit-accumulate math while the same chip in workstation cards runs full rate.
- The paradox: the math really doubles, but training jobs don't run twice as fast and token generation doesn't move at all.
- Pricing it out: after renting and benchmarking both sides of NVIDIA's line, the author concludes GeForce was already the cheapest matrix math per dollar even capped — and the capped card was the better buy all along.
Related event: Open kernel module unlocks hidden AI power on GeForce cards(3 posts)→
More from Infra
- Rogers Expands Starlink Satellite Roaming to Japan, Australia and New Zealand — XFreeze · 2026-10-30
- Corey Quinn flags macOS update silently restoring the removed eMudhra root cert — kingslyj · 2026-10-30
- NVIDIA shares guide on maximizing memory efficiency and cutting costs on Jetson — NVIDIAAI · 2026-10-30
- Rebuild NVIDIA's speaker-captions demo with one transformers serve command — yb2698 · 2026-10-30
- Cap moves transcription to its own GPU farm, up to 11x faster — JosephJacks_ · 2026-10-30
- Prime Intellect makes DeepSeek's MegaMoE trainable, up to 55% faster MoE training — samsja19 · 2026-10-30