# Your GPU Is Nerfed for Local AI (NVIDIA's Hidden Limit) ![[images/GPU-Unlock-Local-AI-thumbnail.jpg]] - **Video:** [Your GPU Is Nerfed for Local AI (NVIDIA's Hidden Limit)](https://www.youtube.com/watch?v=VIm73gaY_hM) - **Channel:** [Kai](https://www.youtube.com/channel/UCgUfJoPtkt7-RiiuYh4vmGA) (on-screen tag "kaiexplainsyt"; about 62,500 subscribers when read) - **Published:** Oct 27, 2026 - **Length:** 25:11 - **Views, likes and comments when read (Oct 28, 2026):** about 72,079 views, 1,182 likes, 227 comments - **Format:** voice-over on animated diagrams (dark background, charts, sliders and counters). No presenter on screen and almost no readable text, so the screenshots are mainly diagrams. - **Transcript source:** YouTube's English auto-generated captions (en-orig), downloaded with yt-dlp. See "Caption fixes" below. - **Screenshots:** 18 frames plus the thumbnail. - **Chapters (from the description):** Paid for 1,321 TOPS, got 165 0:00; Where the box number comes from 1:57; Three halvings 2:32; It's in NVIDIA's own whitepaper 4:23; Why NVIDIA locked it 5:30; gpu-unlock: a free Linux module 6:02; "Your card doubled" vs real benchmarks 7:25; Clocks drop, efficiency rises 7:57; Tokens per second 9:57; Lane 1 (GGUF) 12:03; Lane 2 (vLLM, image, video, fine-tuning) 13:17; FlashAttention at long context 13:59; A 96,000-token prompt 14:51; Your power limit is slowing you down 16:41; The real cost 18:32; Who this helps 21:00; RTX 50 series 22:40; Read the lane before the headline 23:16. - **Caveat:** hardware numbers are mostly quoted from other people's GitHub benchmarks (user names given in the video), so they are second-hand. I could confirm the vendor specs and the project's own README, but not each user's measurement. ## One-paragraph gist The narrator says NVIDIA's own documentation shows GeForce RTX 3090/4090-class cards run 16-bit AI math with a 32-bit running total at half the speed of the workstation card on the same chip. The box figure of 1,321 "AI TOPS" falls to 165 in three halvings: structured sparsity (marketing), 16-bit instead of 8-bit inputs (the model's choice), and the GeForce lock (NVIDIA's choice). A free Linux kernel module, gpu-unlock (Oct 18, 2026), sets two registers to full rate before the NVIDIA driver loads. In a raw matrix benchmark the speed doubles, but real gains are smaller because the card hits its power limit and lowers its clock. The unlock helps the "16-bit lane" (vLLM prompt processing, image and video generation, fine-tuning, long-context FlashAttention) and does nothing for token-writing speed (memory bound), GGUF/integer models, VRAM, gaming or mining. His practical advice: first check your power limit (a 200 W cap on a 3090 was slower and less efficient than 250 W), then decide which lane your software uses, and weigh the Secure Boot and unsigned-module risk. The RTX 50 series has the same lock and no fix. ## Timestamped outline ### Paid for 1,321 TOPS, got 165 (0:00-1:57) ![[images/GPU-Unlock-Local-AI-0m05s-4090-purchase.jpg]] - Opening story: buying an expensive card, loading a 70B model in vLLM on his 4090 and still waiting on long prompts. The box says 1,321 AI TOPS, but his workload ran about 165, "like ordering a full pizza and getting one and a half slices." - He promises an honest look at who this matters for, since "the headlines this week haven't been." ### Where the box number comes from, and the three halvings (1:57-4:23) ![[images/GPU-Unlock-Local-AI-2m05s-box-number-1321.jpg]] - The 1,321 figure is real but assumes everything at once. - **Step 1, sparsity:** assumes half the weights are zero in a fixed pattern; almost no models are trained that way, so 1,321 becomes about 660. - **Step 2, input precision:** the 660 assumes 8-bit inputs, but most models run in 16-bit BF16, halving to 330. ![[images/GPU-Unlock-Local-AI-2m40s-sparsity-halving.jpg]] - **Step 3, accumulator:** the tensor core can keep the running total in 16-bit (fast, rounding errors) or 32-bit (accurate; nearly all real training and inference uses it). GeForce runs the 32-bit mode at half speed, giving 165. He says the first is marketing, the second the model makers' choice, and only the third is a deliberate limit: "two registers got set to half speed before the card shipped." ### It's in NVIDIA's own whitepaper (4:23-5:30) - The Ampere whitepaper table lists the RTX 3090 at 142 teraflops with a 16-bit accumulator and 71 with 32-bit. The RTX A6000, same chip, listed at 154.8 in both modes, and the 2018 Titan RTX ran both at full speed. So nothing is "fused off." ### Why NVIDIA locked it (5:30-6:02) ![[images/GPU-Unlock-Local-AI-5m35s-tflops-per-dollar.jpg]] - NVIDIA has never explained it publicly. His reading: even locked, the gaming card is a better value per dollar, and unlocked it would eat the professional market. It is, he says, like a speed governor "printed on page 47 of the manual in 8-point font." ### gpu-unlock: a free Linux module (6:02-7:25) ![[images/GPU-Unlock-Local-AI-6m10s-gpu-unlock-released-oct-18.jpg]] - On Oct 18 (nine days before recording) a developer called Zagitta posted a free open-source Linux kernel module that loads before NVIDIA's driver, sets the two speed-selector registers to full rate and hands the card over. Nothing is flashed; the driver is untouched; removing the boot option and powering off restores stock. - Within 9 days: almost 10,000 GitHub stars, a port to the RTX 40 series (gpu-unlock-ada, based on work by TOTORONG) and a Turing port covering the 2080 Ti. Tech YouTube covered it with "your card just doubled" thumbnails. ### "Your card doubled" vs real benchmarks (7:25-7:57) ![[images/GPU-Unlock-Local-AI-7m35s-doubled-headline.jpg]] - True only for an isolated matrix benchmark: a 3090 Ti at 450 W went from 87.5 to 157.4 TFLOPS BF16. ### Clocks drop, efficiency rises (7:57-9:57) ![[images/GPU-Unlock-Local-AI-8m05s-register-and-clock.jpg]] - The card does twice the work per clock, hits its power and thermal limit sooner and drops its clock. In one GitHub user's (Galvani's) 3090 test at 250 W, the core clock fell from 1,531 to 1,087 MHz (about 30%) but the card still finished faster. - In real matrix-multiply benchmarks, FP16 gained 12-34%, and BF16 sometimes nothing. Energy per token while reading a long prompt fell by a fifth to a third. With the power limit raised from 250 W to 375 W, prompt-reading gains went from about 25% to about 50%. ### Tokens per second: why writing speed won't move (9:57-12:03) ![[images/GPU-Unlock-Local-AI-10m05s-token-generation.jpg]] - Generating each token mostly waits on memory bandwidth (936 GB/s on a 3090), which is unchanged: 112.98 to 114.24 tok/s in vLLM, "measurement noise." - No change to VRAM size, crypto mining (integer and memory work) or gaming FPS and DLSS ("nobody has measured a single frame rate difference"). ### Lane 1: Ollama, LM Studio, llama.cpp with GGUF (12:03-13:17) ![[images/GPU-Unlock-Local-AI-12m10s-lane-1-int8.jpg]] - These use 8-bit (or lower) integer math, which was never locked. A 4070 Ti Super measured 379.5 integer tensor performance before and after, and a 48 GB modded 4090 on a quantized mixture-of-experts model ran 0.983 times its prior prompt speed. The card's real strength there is VRAM capacity. ### Lane 2: vLLM, image, video, fine-tuning (13:17-13:59) ![[images/GPU-Unlock-Local-AI-13m25s-lane-2-16-bit.jpg]] - 16-bit compute is where the unlock pays: a 3090 in vLLM processed prompts 33% faster; a 3060 image generation went from 18.09 s to 11.01 s (39% better); a 3090 video render from 4:39 to 3:31 with about a quarter less energy. ### FlashAttention at long context (13:59-14:51) - Inside llama.cpp, FlashAttention uses 16-bit math with 32-bit accumulation, so it was locked. A user "Mobile Gojo" on two 3090s measured prompt reading at +3% (empty context), +10% (32K tokens) and +17% (100K tokens), relevant to coding agents with big repos. ### A 96,000-token prompt, timed (14:51-16:41) ![[images/GPU-Unlock-Local-AI-15m00s-vllm-96k-prompt.jpg]] - vLLM on a 3090 at 375 W, 96,000-token cold prompt: time to first word 105.7 s stock, 71.4 s unlocked (34 s saved). At 250 W: 126.8 s to 101.4 s. Token writing 33 to 34.6 tok/s. - He extrapolates: 20 cold reloads a day is about 11 minutes a day. ![[images/GPU-Unlock-Local-AI-16m40s-power-limit-timer.jpg]] ### Your power limit is slowing you down (16:41-18:32) ![[images/GPU-Unlock-Local-AI-16m50s-power-limit-slider.jpg]] - Galvani's test: a 3090 capped at 200 W wrote 50 tok/s at 3.97 joules per token; at 250 W it wrote 76.4 tok/s at 3.15 J. At 200 W the clock had collapsed from about 1,380 to 714 MHz. He had been capping his own card and says to check the power limit before touching the unlock, since it is "one command or slider that can be changed tonight." ### The real cost: Secure Boot and firmware (18:32-21:00) ![[images/GPU-Unlock-Local-AI-18m40s-cost-of-unlock.jpg]] - Card risk is low: the module changes no voltage, clock or power limit and thermal protections stay on. One user ("Link") ran it 8 hours on a production machine at 82-86 °C, same as stock. - The real cost: the module is unsigned, so Secure Boot must be off, a genuine security trade-off. It works by exploiting a flaw in NVIDIA's own signed firmware: a debug copy published for testing, encrypted with a publicly known test key, with a stack-overflow bug in signature checking that allows one four-byte write in the card's most trusted mode. ![[images/GPU-Unlock-Local-AI-20m00s-module-trust.jpg]] - It is a 9-day-old kernel module written mostly by one person: "read the code first." Windows needs more (a UEFI application, Secure Boot and BitLocker off, builds tied to one card, board and BIOS). NVIDIA has said nothing, and could remove the debug firmware or fix the bug. ### Who this helps, and who it doesn't (21:00-22:40) ![[images/GPU-Unlock-Local-AI-21m10s-who-it-does-not-help.jpg]] - Probably not you: Windows users with Ollama or LM Studio on 4- or 8-bit GGUF models, normal-length prompts, gaming on the same card. - Genuinely helps: Linux users running vLLM, full-precision image or video generation, fine-tuning or long-context coding agents in llama.cpp, on a dedicated machine. Prompt processing at long context is 25-50% faster with better efficiency; an image going from 18 s to 11 s. ### RTX 50 series (22:40-23:16) ![[images/GPU-Unlock-Local-AI-22m50s-rtx-5090.jpg]] - The same halving is in NVIDIA's spec for the 5090 (419 vs 209.5 teraflops, per the video). One researcher measured an even larger impact because the clock drops at 575 W. There is no fix: the exploit needs a debug copy of signed boot firmware, and nobody has found one for the RTX 50 generation, which the author called out of scope. ### Read the lane before the headline (23:16-25:11) ![[images/GPU-Unlock-Local-AI-25m00s-conclusion.jpg]] - "Your card doubled" is true in one microbenchmark; "NVIDIA has been lying" is too strong: NVIDIA printed the number in an appendix and put a different one on the box. Take-aways: check your data type (GGUF integer lane vs 16-bit lane), check your power limit now, and weigh the cost of the Secure Boot change. The 24 GB of VRAM is still the card's main value. ## Tutorial and tool notes This is an explainer, not a step-by-step tutorial, but it recommends actions. Install steps below come from the project's README, not from the video. ### Tools mentioned | Tool | What it is | Link | | --- | --- | --- | | gpu-unlock | Linux kernel module that sets the tensor-core speed registers before the NVIDIA driver loads (RTX 30 stable; 20 and 40 experimental per the README) | [github.com/Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) | | gpu-unlock-ada | Fork adding RTX 40 (4090 main target) | [github.com/JustAResearcher/gpu-unlock-ada](https://github.com/JustAResearcher/gpu-unlock-ada) | | Windows port | Community UEFI port the README points to ("YMMV") | [github.com/impacifist/Windows-GPU-UNLOCK](https://github.com/impacifist/Windows-GPU-UNLOCK) | | vLLM | LLM serving engine (16-bit lane) | [github.com/vllm-project/vllm](https://github.com/vllm-project/vllm) | | llama.cpp | GGUF inference engine | [github.com/ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) | | Ollama | Local model runner built on llama.cpp | [ollama.com](https://ollama.com) | | LM Studio | Desktop local-LLM app | [lmstudio.ai](https://lmstudio.ai) | | ComfyUI | Node-based image and video generation | [github.com/comfyanonymous/ComfyUI](https://github.com/comfyanonymous/ComfyUI) | | GPU-Z | GPU monitoring utility mentioned in the intro | [techpowerup.com/gpuz](https://www.techpowerup.com/gpuz/) | ### How the README says to install (not shown in the video) 1. `git clone https://github.com/Zagitta/gpu-unlock` then `cd gpu-unlock` and `make install`. 2. Add `gpu_unlock.targets=all` to the kernel command line (GRUB, systemd-boot and others covered in the README). 3. Power off fully and boot. Check `journalctl -k -b | grep gpu_unlock`. 4. To undo, remove the option and power off. ### Stated vs. inferred | Item | Stated in the video | My reading | | --- | --- | --- | | Power limit | 200 W cap was slower and less efficient than 250 W on a 3090 | Based on one user's test; it depends on the card model and cooler, so test your own card | | Who benefits | Linux, vLLM, image/video, fine-tuning, long context | Consistent with the project README's table | | Secure Boot | Must be off | Follows from an unsigned module (general knowledge); the README text I read does not discuss it | | Safety | Card risk low, real risk is Secure Boot and trusting new code | The video's judgment; no independent security audit seen | | Windows | Needs UEFI app and more risk | Matches the README's "I don't run Windows" and a single-board experimental port | ### Caption fixes "Zagira" and "Zagita" are Zagitta; "Todorong" is TOTORONG; "VLM" is vLLM; "AI taps" and "AI tero operations" are AI TOPS and tera operations; "Mier generation" is Ampere; "havinging" and "h havingings" are halving; "Olama" is Ollama; "Llama CPP" is llama.cpp; "jewels" is joules; "1490 spec sheet" is the 4090; "470 Thai super" is the RTX 4070 Ti Super; "GPUZ" is GPU-Z. ## Claims check | Claim (as stated) | Verdict | Source or note | | --- | --- | --- | | RTX 4090 is marketed at 1,321 AI TOPS; dense FP16 with FP32 accumulate is about 165 | Supported | [NVIDIA RTX 4090 page](https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/) and the [Ada architecture whitepaper](https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf): 1,321 is the sparse INT8/FP8 figure; 165.2 FP32-accumulate vs 330.3 FP16-accumulate dense | | 3090 is 142 TFLOPS (16-bit accumulate) vs 71 (32-bit); A6000 is 154.8 in both; Titan RTX full speed in both | Supported | [GA102 whitepaper](https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf); also tabulated in [this summary](https://jacar.es/en/gpu-unlock-geforce-fp16-bf16-tensor-unlock/) (Titan RTX 130.5 in both) | | A6000 costs about $4,650 | General knowledge | Matches its launch list price as I recall it; not checked this session | | NVIDIA never explained the lock; motive is market segmentation | Opinion | The README gives the same reason as its author's view; no NVIDIA statement found | | gpu-unlock was posted Oct 18 by "Zagira" | Supported (name corrected) | [GitHub repo](https://github.com/Zagitta/gpu-unlock) created 2026-10-18 01:28 UTC (Oct 17, 8:28 PM CT); author is Zagitta | | Almost 10,000 stars in 9 days | Supported | GitHub API on Oct 28 showed about 12,085 stars | | Ports: RTX 40 fork gpu-unlock-ada; Turing port | Supported | [gpu-unlock-ada](https://github.com/JustAResearcher/gpu-unlock-ada) (created Oct 23, 2026); README lists Turing cards (2060 to 2080 Ti, Titan RTX) as experimental | | 3090 Ti at 450 W: BF16 87.5 to 157.4 TFLOPS | Supported | README and the [jacar.es analysis](https://jacar.es/en/gpu-unlock-geforce-fp16-bf16-tensor-unlock/) | | Real matmul gains 12-34% for FP16; BF16 sometimes none | Supported, attribution differs | That source attributes 12-34% to RTX 4090 cuBLAS tests; the video ties it to a 3090 test by Galvani | | Token generation unchanged (112.98 vs 114.24 tok/s); 3090 has 936 GB/s | Supported in principle | README: decode "no change"; the 936 GB/s spec is NVIDIA's published figure (general knowledge). The tok/s pair is one user's test | | GGUF/INT8 unaffected; FlashAttention gains grow with context (+3%, +10%, +17%) | Supported | README: llama.cpp GGUF "0-15%, mostly at long context"; a [user report](https://github.com/Zagitta/gpu-unlock/issues/158) shows "+17% @100k" | | vLLM prefill 25-50% faster; image gen 18.09 to 11.01 s; 3090 video 4:39 to 3:31 | Partly checked | README: prefill +25-55%, image +30-55%, video +15-20%. The video's 24% video gain is above the README's range; the specific timings come from users and are unverified | | 96,000-token prompt: 105.7 s to 71.4 s; 250 W: 126.8 to 101.4 s | Unverified | Galvani's GitHub numbers; not located. Magnitude is plausible compared with the README's 25-55% prefill range | | 200 W vs 250 W power cap: 50 vs 76.4 tok/s; 3.97 vs 3.15 J/token | Unverified | One user's test. A 50 W lower cap giving +50% speed implies a heavily throttled card; test your own | | Module needs Secure Boot off | General knowledge | An unsigned kernel module cannot load under Secure Boot unless signed or enrolled | | It works through a stack-overflow flaw in NVIDIA's debug firmware with a public test key; one four-byte write | Unverified | The README sections I read do not describe the mechanism; the jacar.es write-up describes a "V67/PLM chain" derived from the CMP mining-card unlocker and says it writes two speed selectors. I could not confirm the narrator's description | | Windows version is a UEFI app needing Secure Boot and BitLocker off, tied to one card/BIOS | Partly supported | README points to a community port; jacar.es says the Windows port is one MSI RTX 4070 Ti profile with no binaries or installer | | Temperature 82-86 °C for 8 hours, no errors (user "Link") | Unverified | User report | | RTX 5090 has the same halving (419 vs 209.5) and no fix | Partly supported | README: Blackwell "Not supported and unlikely to happen." The [Blackwell whitepaper](https://images.nvidia.com/aem-dam/Solutions/geforce/blackwell/nvidia-rtx-blackwell-gpu-architecture.pdf) lists 209.5 dense / 419 sparse FP16; I did not confirm that 419 is the 16-bit-accumulate dense rate | | He loaded a 70B model on a 4090 in vLLM | Doubtful | A 70B model does not fit 24 GB at common 4- or 8-bit quantization (general knowledge); either a very low-bit quant, offloading or a different setup | | Description: "Qwen 3.8 27B, MTP speculative decoding" | Unverified | Not said in the narration; I could not confirm the model name or version | | "Nobody has measured any gaming FPS difference" | Matches | [jacar.es](https://jacar.es/en/gpu-unlock-geforce-fp16-bf16-tensor-unlock/): "Nobody has published gaming measurements" | Nothing in the video is financial or medical advice, but the unlock changes boot security. Treat it as a risk decision, not a settled recommendation. ## Things mentioned - [NVIDIA Ampere GA102 whitepaper](https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf) - [Zagitta/gpu-unlock](https://github.com/Zagitta/gpu-unlock) and [gpu-unlock-ada](https://github.com/JustAResearcher/gpu-unlock-ada) - RTX 4090, RTX 3090 / 3090 Ti, RTX 3060, RTX 5090, RTX A6000, Titan RTX, RTX 2080 Ti, RTX 4070 Ti Super - BF16, FP16, FP32 accumulate, INT8, GGUF, FlashAttention, structured sparsity, Secure Boot, BitLocker - vLLM, llama.cpp, Ollama, LM Studio, ComfyUI, GPU-Z (see the tools table) - Wccftech coverage of the unlock (cited by [jacar.es](https://jacar.es/en/gpu-unlock-geforce-fp16-bf16-tensor-unlock/); not named in the video) ## Biographies - **Kai (channel "kaiexplainsyt"):** creator of animated explainers on AI and hardware; I found no public bio beyond the channel page. - **Zagitta:** GitHub developer who created gpu-unlock (Oct 2026). - **JustAResearcher:** GitHub developer behind the gpu-unlock-ada fork for RTX 40. - **TOTORONG:** developer of AD103_win_unlocker, the project whose AD102 payload the Ada fork uses, per the jacar.es write-up. - **Galvani, Mobile Gojo, Link:** GitHub users whose benchmarks the video quotes; I could not find their issue threads individually. ## Tag note - Used: video-summary, ai. - Proposed (not in the approved list): **hardware** (or gpu), **local-ai**.