I tried to unlock my RTX 3090. I ended up testing HyperQwen instead.

It started with a request. Someone asked me if I could do "that fix" on an RTX 3090: the one going around that removes a limit NVIDIA puts on its gaming cards for AI work. I had just put a used 3090 into a PC at home, so I thought, why not? Let's try it.

So I did my homework first. And that's where things slowed down.

The unlock that had to wait

The fix is a project called gpu-unlock. NVIDIA runs certain AI maths on GeForce cards at half the rate of its professional cards, and gpu-unlock lifts that. On a 3090, the project reports dense BF16 maths going from about 75 to about 125 TFLOPS.

Sounds great. But here's the catch, at least for me:

• It's an unofficial kernel module. It loads at boot, before NVIDIA's driver, and it needs a change to your boot settings and a reboot.
• That PC isn't just a test bench. It runs some of my other AI services. Before I touch boot settings on it, I want to isolate it, test it properly, and know exactly how to roll it back.
• And it wouldn't speed up the part people notice most. By the project's own numbers, it speeds up reading the prompt by roughly 25–55%, but not the speed at which the model writes its answer.

The tech press calls it experimental. The project itself lists RTX 30 cards as stable, with 100+ user reports. Either way, it's not something I want to rush on a machine that's doing other work. So the unlock goes on the "later" list.

The next best thing: HyperQwen

A few days earlier, I had written about HyperQwen. It's a project that runs Qwen3.8-27B on a single RTX 3090 using a trick called speculative decoding. A small helper model guesses the next few words, and the big model checks all those guesses in one go. When the guesses are right, you get several words for the price of one.

Their claims: around 127 tokens per second for normal chat, and 381 tokens per second when the answer copies text from the prompt, like code edits, rewrites or translation.

Unlike the unlock, HyperQwen runs in Docker. It doesn't touch the drivers or the boot settings, and you can switch it off and walk away. So that became the experiment: does it really do what it says, on my card?

How I actually did it

I'm not a full-time developer, so I used two AI agents, each for what it's good at.

• Claude did the thinking: researching the claims, planning the tests, and writing the specs (the step-by-step instructions).
• Codex did the hands-on work on the PC, following those specs.
• I stayed in charge. Every sudo command, I typed myself. Nothing touched the system without me seeing it first.

We set some rules early on, and they saved me more than once:

• One step at a time. Run it, look at the output, then move on.
• Log everything. Every command, every output, every decision and every mistake, with a time stamp.
• No number goes into the write-up unless it came from a real run. Claude even re-checked Codex's results against the raw files before I saw them.

Results went into a private GitHub repo as we worked. The clean write-up went into a public one at the end.

Before vs After

To be fair, I compared HyperQwen against itself with the trick switched off (SPEC=off). Same model, same card, same prompts. The only difference is the speed trick.

• Normal chat: 18.4 tok/s before → 47.6–53.6 tok/s after (about 2.7× faster)
• 15 minutes non-stop: 18.4 tok/s (36 answers) before → 48.7 tok/s (105 answers) after (about 2.7× faster)
• Copy a long document: 16.6 tok/s before → 176.2 tok/s after (10.6× faster)
• Summarise the same document: 16.5 tok/s before → 38.3 tok/s after (about 2.3× faster)
• Questions about photos: 18.3–19.0 tok/s before → 41.7–61.0 tok/s after (about 2.3–3.3× faster)

That's a real, solid speed-up. And when the answer copies text, it's over ten times faster. That's exactly the kind of work HyperQwen is built for.

But you'll notice my "After" numbers are nowhere near 127 and 381. There's a reason for that, and it's the biggest lesson of the whole exercise.

The 187-watt mystery

Halfway through testing, Codex stopped and flagged something odd. The GPU was at 100% but running at only about 400 MHz, when a 3090 under load normally runs at somewhere around 1,000–1,700 MHz. It was drawing about 187 watts, half of its 370-watt limit.

The driver's own reason: "SW Power Cap". Not heat, since it never went above 60 °C. Not a hardware fault flag. Just held back.

My first suspect was the power cable. The card came without one, so I'd used the cable from my old RTX 4070: one cable that splits into both of the card's 8-pin sockets. That's a known no-no for these cards. So I rewired it with two separate cables and rebooted.

Still 187 watts. Not the cable.

So every number in my table comes from a card running at roughly half power. The fair way to read it: even on half power, HyperQwen gave me 2.5–3× more speed. On a healthy card, I'd expect the numbers to get much closer to their claims.

What's causing the cap, I don't know yet. It could be something about that motherboard, or the power supply. My next test is to move the card into a higher-spec machine with a bigger power supply. That means taking it out, moving it, and setting it all up again, so it'll take a little time. I'll report back.

What it got right (and wrong) with my photos

I also tested image understanding with four pictures: a selfie of my wife and me in front of palace gates, her portrait, a fake ID card I made, and a child's maths worksheet covered in scribbles. Claude wrote down the correct answers before anything ran, so nobody could fudge the scoring.

• People and details: excellent. The portrait description was spot on: headscarf, gold hoops, winged eyeliner, green eyes, orange lipstick, the finger-on-chin pose.
• Maths: correct. It read 50+ sums off a messy worksheet and got every answer right.
• Facts: shaky. The gates are Buckingham Palace. Across my runs it said Paris, Madrid and Edinburgh, and only one run got London right. It also explained the ID card's number format confidently, and usually named the wrong state for the code.
• Making things up: it invented tick-and-cross marking on the worksheet that simply isn't there.

And here's the thing to remember: run the same photo twice and you can get a different answer. Right or wrong is partly luck. Don't trust a single answer on facts.

What you can expect (the snags)

This did not go smoothly. My first spec didn't get through in one go, and a lot of it ended up as manual work. If you try this, expect some of these:

• The model download stalled at the last file, crawling at about 50 KB/s. Switching off Hugging Face's newer downloader with HF_HUB_DISABLE_XET=1 fixed it instantly.
• Docker wasn't ready. Ubuntu's own Docker didn't have the Compose add-on or NVIDIA's GPU support. Two installs and a restart.
• The Docker image took over an hour to download from GitHub's registry, while the same line did about 100 MB/s from Hugging Face.
• 511 restarts. After a reboot, my other AI services grabbed the GPU memory first, and HyperQwen crashed and restarted 511 times before I noticed. On a shared PC, stop the other services first.
• Thinking mode ran out of room on hard images. With a 4,096-token limit, some runs spent everything thinking and never answered.
• A false alarm. HyperQwen's doctor tool showed the wrong mode. The container was fine.

None of these are HyperQwen's fault, really. They're just what happens on a real PC that isn't a fresh lab machine.

Why I didn't test video

The model understands video too. But with images on, I had only about 1.2 GB of GPU memory to spare, and video means many frames at once. I'm fairly sure it would run out of memory, so I left it for the bigger machine.

Try it yourself

Everything is in my public repo: github.com/frozenfussion/HyperQwen.

• Doing it by hand? There's a step-by-step setup guide, with a warning at every spot where I got stuck.
• Using an AI agent? With Claude's help, I put together a trial spec, a prompt and spec you can give to Claude Code, Codex, OpenCode, DeepSeek's harness or whatever you use. It has all my fixes built in. No guarantees, but it should save you most of my roadblocks.
• Just want the data? All the raw results are there, every request with timestamps and GPU readings, plus the four photos and every answer the model gave.

My advice: before you trust any speed numbers, check your GPU isn't being held back like mine was. It's one command, and it would have saved me a lot of head-scratching.

As for the unlock, it's still on my list. Bigger machine first, then we'll see what this 3090 can really do.