{"slug": "rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move", "title": "RTX 4090 Tensor Limit Unlocked, Yet Chatbot Token Speeds Barely Move", "summary": "A firmware exploit removes Nvidia's half-speed AI tensor throttle on GeForce RTX 20, 30 and 40-series GPUs, doubling the RTX 4090's peak tensor rate and lifting matrix math on the RTX 3090 Ti by up to 80%, according to Tom's Hardware, citing VideoCardz. The same reports state that real workloads gain significantly less than peak tensor rates and that chatbot token speeds stay about the same, because single-stream decoding is limited by memory bandwidth rather than tensor throughput. Three separate projects offer the generation unlocks as unofficial Linux kernel modules, with experimental Windows builds taking a different approach; none is endorsed by Nvidia.", "body_md": "Nvidia ships its GeForce cards with a cap on certain AI math. Modders say they found a way around it.\n\n[Tom’s Hardware reports](https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-20-30-and-40-series-gpus-hacked-to-remove-a-half-speed-ai-throttle-firmware-exploit-doubles-rtx-4090s-peak-tensor-rate-lifts-rtx-3090-ti-matrix-math-up-to-80-percent), [citing VideoCardz](https://theoutpost.ai/news-story/nvidia-rtx-gpus-hacked-to-remove-half-speed-ai-throttle-doubles-rtx-4090-tensor-performance-32640/), that a firmware exploit removes the half-speed AI limits on RTX 20, 30, and 40-series GPUs. Without the limit, the RTX 4090 can hit double its peak tensor rate. Matrix math on the RTX 3090 Ti improves by up to 80%. Gains on other cards vary.\n\nThen comes the sentence that matters most: real workloads gain significantly less than peak tensor rates, and chatbot token speeds stay about the same.\n\nRead that twice before you reach for a flashing tool.\n\n## How it’s being done\n\nAccording to one summary of the story, three separate projects offer generation unlocks. Each ships its own unofficial Linux module. Experimental Windows builds take a different approach. Another write-up describes the same thing as Linux kernel modules that lift Nvidia’s half-rate AI tensor limit on GeForce RTX 20, 30, and 40-series cards.\n\nThat’s about all the sources give us on mechanism. We won’t guess at the rest. No exploit internals, no project names, no install steps here. If you want those, go to the projects themselves.\n\nTwo things are worth noting at the architecture level:\n\n- **The Linux builds ship as kernel modules.** Tom’s Hardware calls the underlying trick a firmware exploit. How the modules and the card’s firmware fit together is for the projects to explain.\n- **Nothing here is blessed by Nvidia.** These are unofficial, and the Windows builds are labeled experimental.\n\n[HotHardware](https://hothardware.com/news/modders-unlock-geforce-rtx-20-30-and-40-gpus-for-up-to-2x-compute-gains) frames the result as making consumer cards nearly as capable as professional RTX GPUs for relevant AI and compute workloads. It also adds a warning: gamers shouldn’t get too excited. Extra compute doesn’t turn into extra frames.\n\n## Peak rate is not your workload\n\nA tensor core’s peak rate is a ceiling you hit only when everything else keeps up. Data has to arrive, and it has to arrive fast enough to feed the math units. For a lot of real jobs, it doesn’t.\n\nChatbot-style generation is the clearest case, and the sources say as much: token speeds stay about the same. The reason is that single-stream decoding is a memory-bandwidth problem. Each new token means streaming the model’s weights through the chip. The matrix units sit waiting on memory.\n\nWe see this in our own hardware notes. Our [ASUS ROG Astral RTX 5090](https://www.gladlabs.io/go/asus-rog-astral-nvidia-geforce-rtx) moves roughly 1,790 GB/s. Our RTX 3090 moves 936 GB/s. The integrated GPU on the Ryzen 9 9950X3D manages about 77 GB/s. For batch-1 token generation, that bandwidth ordering is what predicts the speed ordering. Tensor rate doesn’t enter into it. Doubling the math on a card that’s already waiting on memory buys you nothing.\n\nThat chart shows something related from our own logs: raw decode speed versus what the application actually receives. The gap between the two has nothing to do with tensor throughput. It lives in queueing, lock waits, and model swaps. Our pipeline loses far more time to a GPU being busy than to a GPU being slow.\n\nIf you’ve read our [local LLM GPU ranking](https://www.gladlabs.io/posts/best-nvidia-gpus-for-local-llms-in-2026-ranked-by-dde9c4f9), none of this should surprise you. VRAM decides what fits, and bandwidth decides how fast it decodes. Tensor peak comes third.\n\n## Where it could actually help\n\nDon’t write the unlock off, though. Some workloads are compute-bound, and the “up to 80%” and “2x” figures describe real ceilings.\n\n**Prompt processing.** Prefill chews through many tokens in parallel. It leans on matrix math far more than decode does. Long prompts and big contexts spend their time here. If you stuff 100K tokens of context into a model, prefill dominates your time to first token. It’s the one place a chat workload might notice. We can’t promise it will, because the sources don’t break out prefill numbers. Test it before you believe it.\n\n**Training and fine-tuning.** Larger batches push arithmetic intensity up, which moves you toward the compute-bound side. Our training runs on the RTX 5090 and 3090. We covered the platform tradeoffs in [why we still buy NVIDIA over AMD ROCm](https://www.gladlabs.io/posts/we-only-train-on-an-rtx-5090-and-3090-so-is-nvidia-502ffb23). If the unlock held up, a 3090 doing fine-tuning is the sort of job that would gain something.\n\n**Batched serving.** Run many concurrent requests and the weights get reused across them. That also raises arithmetic intensity. Our own load is one to three concurrent requests a day, so we’re nowhere near this regime. If yours is, you’re closer to where the unlock could show up.\n\n**Image and video generation.** Diffusion models are heavy on matrix math. Our render pipeline spends minutes per clip on the 5090. We can’t say whether any unlock would shorten that, since our sources don’t test it, and the 5090 is a 50-series card, outside the list of affected generations.\n\n## What we’d check before trusting it\n\nWe haven’t run any of these modules. Here’s how we’d approach one on a box we care about.\n\n### 1. Treat it as an unofficial kernel module\n\nA kernel module runs with full privilege. You’re loading code from a hobby project into the most trusted part of your machine. That’s a real decision. Our own habit is to put custom code on an isolated test machine first, and to hold anything that touches the host behind a security review. The same applies here.\n\n### 2. Benchmark your workload, not the peak\n\nThe headline numbers are peak rates. Your job isn’t a peak rate. Measure the thing you run:\n\n- Prefill time on a long prompt.\n- Decode tokens per second at batch 1.\n- Decode at the concurrency you actually serve.\n- A training step on your real batch size.\n\nRun each before and after, same driver, same clocks, same model. If decode doesn’t move, that’s the expected result, not a failed unlock.\n\n### 3. Watch power and thermals\n\nMore tensor throughput can mean more switching activity. We don’t have a number for how much extra power these modules cause, and the sources don’t give one. We’d measure it rather than assume it.\n\nPower matters in our setup. [Our workstation idles at 278 W](https://www.gladlabs.io/posts/our-dual-gpu-ryzen-9-9950x3d-workstation-idles-at-311fd66c) before any model runs, and we’ve watched the UPS cross 60% load. Spare watts aren’t free here.\n\nWe’ve written about holding efficiency up while cutting heat in our [undervolting guide](https://www.gladlabs.io/posts/undervolting-your-gpu-for-local-ai-inference-lower-4a4b9054). If the unlock raises draw, the same tools apply. A power cap with a throttle removed may still beat stock for the thermals you can live with.\n\n### 4. Plan for it breaking\n\nDriver updates change things, and unofficial modules lag. Pin your driver version once you have a working combination. Keep a stock path you can fall back to, and don’t make production depend on the patch.\n\n## The 3090 angle\n\nMost of the people who’ll care about this own older cards. Used RTX 3090s are the cheap route to 24 GB of VRAM, and that’s why they sit in so many home labs. Ours does the vision judging for our pipeline. It holds a 30B-class vision model in about 23.9 GB of VRAM.\n\nThat card is idle most of the time. In one review window it showed a 1.3% duty cycle, and the VRAM was 80% committed anyway. A tensor unlock does nothing for a card that’s waiting around. The bottleneck on that card is how often we feed it work, which is a scheduling problem.\n\nOur broader finding from running two GPUs is that the bottlenecks sit around the silicon. Lock contention, VRAM squatting between services, and eviction ordering cost us more time than the arithmetic ever did. Over a two-week window, 87% of the power we paid for went to idle, and $23.14 of the bill was idle draw against $3.40 active. Faster math on a card that mostly idles saves you almost nothing.\n\nThat doesn’t make the unlock pointless. It means the first question is whether your card is compute-bound at all. Check that before you flash anything. If your GPU sits at 100% utilization during decode, that doesn’t tell you much. Utilization counts time busy, not whether the math units are saturated. Look at where the time goes.\n\n## Why this story has legs anyway\n\nStrip out the hype and a real point remains. Nvidia built a capability into these chips and fused part of it off for segmentation. A cap like this is a product-tier decision, not a physical limit. HotHardware’s framing is that it makes GeForce parts nearly as capable as professional RTX GPUs for relevant workloads. That’s the part that makes this a story for people who build their own machines.\n\nBut for most local-LLM users, the gain is small. The sources say chatbot token speeds stay about the same. We’d bet that holds for anyone doing interactive generation on one stream, because bandwidth, not tensor rate, sets that speed.\n\nIf you do training, big batches, long-context prefill, or diffusion at volume, it’s worth a look on a test machine. Measure first, and keep the stock path as your fallback.\n\n## Bottom line\n\nThe unlock is real, according to [Tom’s Hardware](https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-20-30-and-40-series-gpus-hacked-to-remove-a-half-speed-ai-throttle-firmware-exploit-doubles-rtx-4090s-peak-tensor-rate-lifts-rtx-3090-ti-matrix-math-up-to-80-percent), HotHardware, and the other outlets covering the three Linux module projects. The doubling on the RTX 4090 is a peak figure. The 80% on the RTX 3090 Ti is a ceiling. Real workloads gain less, and chat decode barely moves.\n\nIf you run a 3090 or 4090 for inference, don’t expect faster replies. Buy bandwidth and VRAM for that. If you run a compute-heavy job, benchmark on a throwaway box, watch your power draw, and pin your driver. Don’t load an unofficial kernel module on a machine you can’t afford to lose for a number on a spec sheet. Measure your own workload before you believe anyone’s peak rate, ours included.\n\n## Sources\n\n- [https://theoutpost.ai/news-story/nvidia-rtx-gpus-hacked-to-remove-half-speed-ai-throttle-doubles-rtx-4090-tensor-performance-32640/](https://theoutpost.ai/news-story/nvidia-rtx-gpus-hacked-to-remove-half-speed-ai-throttle-doubles-rtx-4090-tensor-performance-32640/)\n- [https://hothardware.com/news/modders-unlock-geforce-rtx-20-30-and-40-gpus-for-up-to-2x-compute-gains](https://hothardware.com/news/modders-unlock-geforce-rtx-20-30-and-40-gpus-for-up-to-2x-compute-gains)\n- [https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-20-30-and-40-series-gpus-hacked-to-remove-a-half-speed-ai-throttle-firmware-exploit-doubles-rtx-4090s-peak-tensor-rate-lifts-rtx-3090-ti-matrix-math-up-to-80-percent](https://www.tomshardware.com/pc-components/gpus/nvidia-rtx-20-30-and-40-series-gpus-hacked-to-remove-a-half-speed-ai-throttle-firmware-exploit-doubles-rtx-4090s-peak-tensor-rate-lifts-rtx-3090-ti-matrix-math-up-to-80-percent)", "url": "https://wpnews.pro/news/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move", "canonical_source": "https://www.gladlabs.io/posts/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-s-46a9f65e", "published_at": "2026-10-28 13:00:00+00:00", "updated_at": "2026-10-28 13:12:02.566336+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-chips", "large-language-models"], "entities": ["Nvidia", "GeForce RTX 4090", "GeForce RTX 3090 Ti", "Tom's Hardware", "VideoCardz", "HotHardware"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move", "markdown": "https://wpnews.pro/news/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move.md", "text": "https://wpnews.pro/news/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move.txt", "jsonld": "https://wpnews.pro/news/rtx-4090-tensor-limit-unlocked-yet-chatbot-token-speeds-barely-move.jsonld"}}