
You can enable more Tensor Core power on your GeForce card, but don’t expect to double your AI performance.
By now you’ve probably seen the posts, the microbenchmarks, and the “RTX 4090 now hits 330 TFLOPS” screenshots. The software locks that kept consumer cards at roughly half the matrix-multiply rate of their professional counterparts on certain FP16/BF16 (and related TF32) Tensor Core paths have been cracked. On RTX 20-, 30-, and 40-series cards the enables work, they are temporary (reset on reboot), they don’t require VBIOS flashing in the main projects, and the peak numbers are real.
What those peak numbers do not tell you is how much faster your actual AI workloads will run.
What the enable actually delivers
In pure instruction throughput the gains look dramatic:
- RTX 3090 Ti: BF16 matrix multiply roughly 87.5 → 157 TFLOPS (~80 % increase). Similar jumps on TF32.
- RTX 4090: FP16/BF16 Tensor Core peaks essentially double (e.g. ~166 → 332 TFLOPS).
- Turing cards (2080 Ti and friends) and many Ampere/Ada parts show comparable microbenchmark doublings.
These are the numbers that travel well on social media. They measure how fast the Tensor Cores can chew through the specific matrix operations that were previously rate-limited. Once the software throttle is removed, the silicon is capable of the higher rate, until something else becomes the bottleneck.
Why real applications don’t double
Most local AI work is not pure Tensor Core instruction throughput.
Prompt / prefill processing (the compute-heavy phase of LLMs) is the part that benefits most. Independent reports commonly show 14-36 % faster prefill, occasionally a bit more depending on model size, batch size, and precision. That is a solid, measurable improvement (especially on longer contexts) but it is nowhere near a 2× wall-clock speedup.
Token generation (decode) is almost always limited by memory bandwidth and KV-cache movement. Enabling extra Tensor Core FLOPS does almost nothing here. If your daily use is interactive chat or continuous generation, the practical gain is often in the single-digit percentages or flat.
Image-generation and fine-tuning workloads sit somewhere in between: they see useful but still sub-linear improvements, and they still run into the card’s power and thermal limits. The enable does not raise the TDP or the power-delivery ceiling. Once the card hits its power wall or thermal throttle, the extra theoretical FLOPS stay theoretical.
Gaming performance is unaffected, which is expected, the paths being enabled are the AI/compute matrix routes, not the graphics pipelines.
Practical realities
- The tools (gpu-enable and its Ada/Turing variants) are primarily Linux kernel-module or pre-driver approaches. Windows/UEFI paths exist but are less mature. Support is most solid on the 30-series; 20- and 40-series work is still more experimental.
- Changes are not permanent. A reboot returns the card to stock behavior.
- Stability, boot reliability, and long-term risk are on you. These are unofficial modifications. Warranty implications exist.
- Power and cooling still matter. A card that was already near its limit will simply hit that limit sooner at the higher clock/throughput.
Setting the right expectation
If your workload is heavily prefill- or training-bound and you are already on Linux, the enable can give a worthwhile free performance bump, think “noticeably faster prompt processing” rather than “twice as fast overall.” If most of your time is spent generating tokens, or if you are bandwidth-starved, the difference will be modest to negligible.
The silicon was always capable of more on those Tensor Core paths. NVIDIA simply software-limited the consumer parts. Removing that limit restores the missing throughput, but it does not magically remove memory-bandwidth walls, power limits, or the fact that real models are not pure GEMM microbenchmarks.
So yes, you can improve your card’s AI output. Just don’t plan your hardware decisions around the doubled peak TFLOPS figures. The realistic gains are real, useful for the right workloads, and still well short of a full doubling of end-to-end performance.
Get the next one in your inbox
Enjoyed this? I write about local AI, self-hosting, and cloud cost ninjutsu. Join the Ninja Dispatch for occasional, high-signal notes. No fluff.

