News · 2026-10-25

RTX compute unlocks reach Ada and Turing, with gains narrower than the 2× headlines

Separate open-source projects have extended selected RTX tensor-compute unlocks beyond Ampere to named Ada and Turing cards. The reports show that some instruction rates can nearly double, but they do not demonstrate a universal doubling of local AI application speed. The October 23–25 development is a spread of card-specific implementations, with substantial compatibility and boot-time risks still attached.

Key facts

• An RTX 4090 owner’s matched matrix tests reported gains of 12–34% in selected half-precision cases.
• Zagitta’s original project targets RTX 30; separate projects document Ada and Turing work.
• The Ada project’s October 24 release adds probe evidence for one RTX 4070 Ti SUPER.
• Primary evidence includes the Ada repository and its RTX 4090 report.

The attraction is obvious: an installed graphics card may have more throughput available on a particular calculation path than ordinary software exposes. But the word unlock can suggest a general transformation that the evidence does not support. These projects do not add memory, increase memory bandwidth, double clock speed, or accelerate every numerical format.

The original gpu-unlock repository describes a Linux kernel module that runs before the normal NVIDIA driver takes control. It executes firmware, changes speed selectors, resets the device, and then allows the unmodified driver to initialize it. This is different from installing an ordinary application optimization, and different from flashing a replacement video-card firmware image.

The calculations affected are particular matrix operations used by neural networks. Some use half-precision inputs while accumulating results at a higher precision. Changing how quickly those instructions issue can raise their theoretical throughput. The Ada reports also examine another matrix mode and include controls in lower-precision formats that did not show material improvement.

Think of a factory that doubles the speed of one machine. Total production rises substantially only if that machine was holding up the line. If the loading dock or supply conveyor is already the constraint, the faster machine waits. Ground Truth’s explanation of Amdahl’s law describes the same limit for software.

That is why the RTX 4090 evidence needs two separate numbers. The Ada project reports roughly doubled throughput for selected instructions. A later independent owner run reports matrix-multiplication gains varying with the shape of the calculation: half-precision cases improve 12–34%, and some other cases improve less or barely at all. Neither report is an extended stability study, and neither establishes an end-to-end language-model improvement on every Ada board.

The original project’s language-model reports make another distinction: processing a prompt can improve while generating the reply stays about the same. Prompt processing performs large batches of calculations; reply generation repeatedly accesses model state to produce the next token. The prefill and decode lesson explains why those phases encounter different bottlenecks. Faster matrix instructions do not automatically make memory move faster.

The Turing evidence is narrower still. Turing-unlock reports tests on one Dell RTX 2080 Ti with physically expanded memory. Its selected matrix cases improve by roughly 30–64%, rather than uniformly doubling. KohakUnlock supplies a separate one-card demonstration and strong experimental warnings. Zagitta’s Turing support-request thread describes the status as “no stable support yet.”

Community interest is real, but discussion is not a compatibility list. In the Hacker News thread, the maintainer discusses work on newer cards while commenters raise power, stability, and bandwidth questions. Coverage has helped spread the broader two-times framing, including Wccftech’s October 24 article. The project reports are the stronger basis for deciding what actually ran.

The strongest favorable interpretation is that a reproducible hardware restriction has become a community engineering opportunity. Even a modest gain could matter for repeated workloads on equipment people already own. The strongest counterargument is that a narrow speedup may be irrelevant to common workloads and is currently purchased with privileged early-boot experimentation.

The sources reviewed contain no NVIDIA statement addressing these projects, and they do not establish NVIDIA’s reason for the limits. They also do not establish warranty outcomes or continued operation across all driver updates. The useful next milestone is a matched, same-card application benchmark with disclosed software, power, and stability measurements. A reproducible report should distinguish average latency from best-case throughput, because the fastest isolated run need not describe normal operation. Until then, the defensible news is wider experimental reach for specific compute paths, with smaller and uneven gains once the whole workload is measured.

Primary source, verified: read the paper →

Ground Truth. (2026, October 25). RTX compute unlocks reach Ada and Turing, with gains narrower than the 2× headlines. Ground Truth. https://groundtruth.day/news/rtx-unlock-spreads-to-ada-and-turing-with-narrow-gains.html

@misc{groundtruth:rtx-unlock-spreads-to-ada-and-turing-with-narrow-gains,
  title  = {RTX compute unlocks reach Ada and Turing, with gains narrower than the 2× headlines},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/rtx-unlock-spreads-to-ada-and-turing-with-narrow-gains.html}
}

Topics: hardware · local-ai · open-source · inference