So what's going on here, exactly?
Let's cut through the jargon first. You spent hard-earned money on an RTX 4090 or 3090. You're running Stable Diffusion, training LoRAs, or testing local LLMs — and you know something's off. Your card is not delivering what it should.
Here's why: NVIDIA deliberately cripples AI compute on consumer GeForce cards to push you toward their expensive professional lineup. The hardware inside your RTX 4090 is basically the same silicon as an A100 — but the firmware caps its Tensor Core FP16/BF16 throughput at 50% of what the chip can actually do.
We're not conspiracy-theorizing. This was confirmed by the community project gpu-unlock, which was picked up by Tom's Hardware, Notebookcheck, and dozens of other tech outlets on October 26, 2026.
The good news?
Someone found the lock. And it's easy to pick.
Developer Zagitta released a free, open-source Linux kernel module that modifies a few GPU registers before the NVIDIA driver even loads — writes the full-speed values, resets the card, and hands control back to the stock driver like nothing happened. The Windows community quickly built a port using a UEFI shell approach (since Windows locks kernel signing).
The results? Real-world users are reporting:
- RTX 3090 Ti BF16 throughput: 87.5 → 157.4 TFLOPS (+80%)
- RTX 4090 FP16 throughput: ~165 → ~332 TFLOPS (+100%)
- No vBIOS flash, no patched driver, no permanent hardware change.
What guides do you need?
Pick your OS below:
🪟 Windows Guide
UEFI shell method, step-by-step screenshots. Works with RTX 30/40 series.
🐧 Linux Guide
Native kernel module. One-line install. The original — cleanest and safest approach.
📊 Benchmarks
Side-by-side before/after TFLOPS from the GitHub README and community tests.
⚠️ Safety & Risks
Will it brick my GPU? Void my warranty? Get me banned? Honest answers inside.
Which GPUs are supported?
30 Series (Ampere) — fully supported on Linux; Windows port has hardware-specific success reports:
40 Series (Ada Lovelace) — community fork gpu-unlock-ada covers these:
20 Series (Turing) — separate fork exists, experimental. Expect mixed results.
Why this site exists
I spent a weekend testing both the Linux kernel module and the Windows UEFI unlock on my own RTX 4090. It worked. Then I went looking for a simple guide — and couldn't find one. Every search result was either a Reddit thread, a GitHub issue, or a Tom's Hardware article that said "this exists" but not "here's exactly what to do."
So I put this together. All the steps I wish someone had laid out for me, in one place.
Not affiliated with Zagitta, NVIDIA, or any hardware brand. This is an independent community resource.