'Half of the RTX 3090's tensor speed was switched off,' says a user who flipped it back with one kernel module

A post claiming NVIDIA sold the RTX 3090 with half its tensor speed switched off, and that a single Linux kernel module turns it back on, drew 780 likes and 740 bookmarks on X.

X user superalesha wrote on Oct 21 that on the same cards, same driver and same 300W, BF16 tensor throughput went from 82 to 155 TFLOPS, 1.9 times faster. In his numbers, cuBLAS BF16 GEMM 8192 rose from 64 to 99 TFLOPS and TF32 from 41 to 79. Outputs matched the stock run byte for byte, he said.

"Not missing. Switched off," he wrote, arguing it was done so gaming cards wouldn't eat into workstation sales. LLMs gain little, he said, because decoding waits on memory, so he saw only about 10% on prefill. The bigger wins should come in diffusion models built on large matrix math, and in training: "every LoRA anyone ever trained on a 3090 went through exactly the path that was cut."

Asked about the 4090 and 5090, he said he didn't know. No NVIDIA response to the claim could be found.