GeForce vs. Workstation GPUs: Nvidia's Half Rate, Priced
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Two open-source Linux modules, released in late October 2026, reportedly switch off the half-rate limit NVIDIA places on GeForce RTX 20, 30 and 40 cards for 16-bit matrix math with 32-bit sums, the mode mixed-precision training uses. Owners report about 1.9x per clock on an RTX 3090 but only 10 to 34% on an RTX 4090, and the RTX 5090 is unsupported. The author's benchmarks show used GeForce cards deliver the cheapest matrix math per dollar. Some of the sourcing could not be independently checked, and some runs fail.