Open source unlocks hidden RTX 4090 power: Tensor Core throughput doubles

One RTX 4090, no hardware swap, and throughput for certain Tensor Core operations jumps from 165.6 to 331.6 TFLOPS. That is exactly twice as fast. The open project, reported by Jiqizhixin, taps into compute headroom on GeForce cards that NVIDIA does not expose in normal operation. The driver stays the official one.

Not just the flagship

The key detail in this story is that the RTX 4090 turned out to be merely the loudest example. The same approach squeezed extra performance out of cards from different generations. The tested devices include:

• RTX 4090 (Ada Lovelace architecture): from 165.6 to 331.6 TFLOPS on certain Tensor Core operations;
• RTX 4070 Ti SUPER (Ada Lovelace): headroom found; this is a lower-tier card from the same generation;
• RTX 3090 Ti (Ampere): headroom found; this is the previous generation;
• a graphics card from 2018, i.e. the Turing era: the potential was unlocked even on this one.

The fact that the result repeats across three architectures points to a systemic cause. This is not a bug in one particular chip. Most likely, it comes down to how the stock software stack addresses the tensor units.

Where the "extra" half comes from

The official figures for NVIDIA's consumer GPUs are well known. For the RTX 4090, a value of around 165 TFLOPS corresponds to mixed-precision mode with FP32 accumulation. Professional cards built on similar silicon are rated higher in this mode. For that reason, gaming GeForce cards have long been considered artificially cut down in tensor operations. The open method shows that part of this limit can be bypassed in user code, not through firmware or a driver patch.

It is important that the gain applies to certain operations. This is not a magic doubling of everything. Real workloads such as LLM inference are also bound by memory bandwidth, cache size and kernel scheduling. The doubled Tensor Core peak will speed things up where matrix computation was the bottleneck: large batches, prefill of long prompts, and training and fine-tuning of small models.

What to do with this in practice

This is good news for owners of gaming cards. The RTX 4090 with 24 GB of memory remains the most popular card for running models locally and for experiments, and a free gain on matrix operations immediately raises its value. Startups, research groups and enthusiasts who have no budget for an H100 or B200 stand to benefit.

Before rewriting a pipeline, check three things:

• compare your own kernels and libraries with what the project does, and measure real throughput on your own workload, not just peak TFLOPS;
• make sure the setup is stable: power draw and temperature under full Tensor Core load are higher than usual;
• watch for NVIDIA's response, since the company may close the loophole in upcoming driver or CUDA versions.

The story also has a broader meaning for the industry. The gap between consumer and data-center accelerators is largely defined in software, and the open community knows how to narrow it. If the method takes hold in popular frameworks, some tasks that go to the cloud today could run on a local card. That puts pressure on NVIDIA's pricing and strengthens the case for local inference.