Benchmarks: Before & After Unlock

All numbers from the official gpu-unlock README and verified community tests

These are TFLOPS (trillion floating-point operations per second) for Tensor Core matrix multiply — the operation your AI models spend 90% of their time doing. Higher = faster AI.

Every number below is a median of 5 test runs on the same card, same power limit, same driver version. No overclocking, no firmware modifications — just the register write.

RTX 30 vs RTX 40: Why the "double" claim depends on which series you have

Some tech outlets (like PBX Science) have noted that RTX 30 series gains top out around 80%, not doubling. That's accurate for Ampere cards — but it's not the whole story. The RTX 40 series (Ada Lovelace) has a much harder throttle lock, and the community fork gpu-unlock-ada consistently delivers ~100% throughput recovery — i.e. double — on the 4080 and 4090.

Short version: If you have a 30 series, expect ~80% boost. If you have a 40 series, expect ~2x. Both are real — they just lock different amounts.

RTX 40 vs RTX 30: Side-by-Side Comparison

The key question — "will my card actually double?" — depends entirely on whether you have Ampere (30 series) or Ada Lovelace (40 series) silicon. Here's the apples-to-apples look:

Card	Architecture	Stock BF16	Unlocked BF16	Improvement
RTX 3060 12GB	Ampere	~20 TFLOPS	~38 TFLOPS	+90%
RTX 3090 Ti (24GB)	Ampere	87.5 TFLOPS	157.4 TFLOPS	+80%
RTX 4080 SUPER (16GB)	Ada Lovelace	148 TFLOPS	295 TFLOPS	+99%
RTX 4090 (24GB)	Ada Lovelace	~165 TFLOPS	~332 TFLOPS	+101% 🎯

All data sourced from Zagitta/gpu-unlock README, gpu-unlock-ada fork, and community submissions on GitHub Issues / r/LocalLLaMA.

RTX 3090 Ti (24GB GDDR6X, Ampere)

Source: Zagitta's own test rig · 450W power limit · FP32 accumulation

Precision	Before (TFLOPS)	After (TFLOPS)	Improvement
FP16	87.46	118.39	+35%
BF16	87.50	157.41	+80%
TF32	42.58	82.61	+94%

RTX 4090 (24GB GDDR6X, Ada Lovelace)

Source: gpu-unlock-ada fork · FP16 matrix multiply

Precision	Before (TFLOPS)	After (TFLOPS)	Improvement
FP16 (Tensor Core)	~165	~332	+101%
BF16 (Tensor Core)	~165	~332	+101%

Real-world translation for RTX 4090: If you normally train a LoRA in 20 minutes, expect 10-13 minutes. If you normally do 30 iters per second on Stable Diffusion XL, expect 45-55 iters/s.

RTX 3060 12GB (Ampere)

Source: community, Arch Linux · gpu-unlock 1.1.0 · power limit 130W

Precision	Before	After	Improvement
FP16	~20 TFLOPS	~38 TFLOPS	+90%

What about gaming?

Basically unchanged. The unlock targets FP16/BF16 Tensor Core operations. Gaming runs primarily on FP32 compute and RT cores — neither of which are throttled. Your framerate in Cyberpunk, Hogwarts Legacy, or Valorant will not noticeably change.

If anything, you might see a slight increase in GPU temperature during gaming because the card is now free to draw slightly more power, but NVIDIA's existing power limiters still kick in. No thermal runaway. No fan explosions.

How do these numbers compare to pro cards?

This is the juiciest part. After unlock, a $1,500 RTX 4090 delivers AI performance within striking distance of a $6,000 A100 40GB (on Tensor Core throughput specifically). You're not catching up to an H100 — but you're catching up to professional cards that cost 4x more.