Nvidia's tensor cap looks like a pricing decision, and Linux just exposed it
Open-source Linux modules double the RTX 4090's FP16 tensor throughput without adding hardware, which suggests the limit was about segmentation, not silicon.
Photo: ZMASLO / Wikimedia Commons (CC BY 3.0)
If a card can deliver twice the tensor throughput on some operations after a kernel module and a reboot, was it ever really "limited" by physics? That is the uncomfortable question Nvidia now has to answer, and we think the answer is obvious.
What actually happened
According to PC Games Hardware, three developer projects now bypass the tensor throttle on selected GeForce RTX 20, 30 and 40 cards under Linux. They build on earlier work on Nvidia's CMP mining cards. The Ada Lovelace version, gpu-unlock-ada, lifts the cap on FP16 and BF16 math with FP32 accumulation on an RTX 4090. Measured throughput goes from 165.6 to 331.6 TFLOPS.
The Ampere project, gpu-unlock, is the most mature. On an RTX 3090 Ti, the report says BF16 rises from 87.5 to 157.4 TFLOPS, a gain of nearly 80 percent. No extra compute units or memory are unlocked. The module runs before the regular driver, using signed Nvidia firmware, and the stock driver then takes over. The change disappears at reboot and has to be applied again on every boot.
The numbers say "segmentation," not "physics"
Nvidia has documented the split for years. In its Ada Lovelace whitepaper, per PCGH, the RTX 4090 is listed at 165.2 TFLOPS for FP16 with FP32 accumulation and 330.3 TFLOPS with FP16 accumulation. The unlocked result lands almost exactly on the second figure. Same chip, same silicon, and a ceiling that is evidently a configuration setting.
That looks like classic market segmentation. Professional and data-center parts built on the same GPUs get the full rate, and GeForce buyers don't. Nvidia is entitled to sell different products at different prices. But we find it hard to defend calling a register value a hardware difference. One Reddit thread put it bluntly, saying the unlock removes an artificial restriction so enterprise customers could be charged more. That is a forum opinion, not a documented fact, and Nvidia's motive isn't spelled out in the sources. But it is hard for us to read the data any other way.
We also think the timing matters. The cap matters most for BF16, which PCGH notes many current AI models use. Nvidia sold these cards in a world of gaming. Owners now run local language models and image generation on them, and in our reading a throttle that protects AI-oriented products hits them too.
Don't oversell the gain, and don't ignore the risks
Before anyone rushes to flash anything, some cold water. The gains are uneven. PCGH reports the biggest jumps in long-prompt processing and certain image models, while single-token output barely speeds up. No gaming frame-rate gains have been measured. FP8, which Nvidia promoted heavily on Ada, improved just 0.3 percent in the only test so far.
The evidence for the headline card is thin too. The 4090 success was documented on one card after a cold start, and PCGH says reliability across other units isn't properly studied. Ampere has user reports for only five models, and Turing is just a feasibility study. This is experimental software that runs below the driver and starts signed firmware with a custom payload. We'd treat it like overclocking firmware: fine on a spare machine, a poor idea on the box you need for work tomorrow. We also wouldn't count on any warranty goodwill, though the documentation doesn't say how Nvidia will respond, and we won't guess.
Our verdict
People should own the silicon they pay for. If a chip can do something safely, leaving it disabled in firmware purely to protect a higher-priced SKU is a hard sell to the buyer of a flagship card. That the community had to dig into firmware to get there is the real story. The best outcome would be Nvidia making this an official, supported option, or at least explaining where the line sits. We doubt it will, since the whole point of the cap seems to be price separation.
Until then, our advice is simple. If you run local AI on an older GeForce under Linux, watch these projects mature, test on non-critical hardware, and wait for more than one successful 4090. If you're deciding what to buy, remember that spec sheets describe what the vendor lets the chip do, not necessarily what it can do.
No topics yet: start the first one.
More stories
DuckDB 2.0 Makes the Case for Keeping Analytics Close to the Data
DuckDB 2.0 does not reject the cloud; it argues that local execution, embedded analytics and selective remote scale belong in the same data architecture.
The Moon’s First HPC Test Is Really a Test of Local Autonomy
HPE’s lunar computer matters less as a record than as a practical experiment: future missions will need to compute locally when Earth is too far away.
NVIDIA's RTX Spark: Windows Finally Gets Its AI Superchip Moment
The long-awaited NVIDIA RTX Spark has arrived, directly challenging Apple Silicon's lead in high-performance local AI processing on Windows PCs.
The AI Infrastructure Arms Race Accelerates – And Costs Escalate
Intel, AMD, and Nvidia's latest data center GPU and server rack announcements at OCP confirm an intensifying AI infrastructure battle, promising innovation but also significant cost challenges.
