A developer runs one matrix multiply test before committing a weekend to a fine-tune.
On an RTX 3090 Ti capped at 450W, BF16 comes back at 87.5 TFLOPS.
That is what the spec sheet allows, and it is half of what the silicon can do.
On GeForce cards, the FP16 or BF16 multiplies that accumulate in FP32 get issued at half rate.
Most training leans on exactly that path.
The same chip sold as an RTX A6000 runs it at full speed.
The 24GB RTX 3090, the card everyone recommends for cheap AI work, ships with its tensor cores in the workstation market's slower mode.
Zagitta, a GitHub user, published gpu-unlock to change that on RTX 30-series cards.
It is a Linux kernel module that loads before Nvidia's driver, rewrites two speed-select registers and resets the card.
No VBIOS flash, no patched driver.
On the same 450W 3090 Ti the test went from 87.5 to 157.4 TFLOPS BF16, TF32 nearly doubled, and FP16 gained 35 percent.
Prompt processing runs 14 to 36 percent faster, and fine-tuning and image generation are where that lands.
An RTX 4090 doubled its FP16 issue rate, while Turing cards gained 36 to 64 percent on memory-modded 2080 Ti builds, though owners measuring real matrix workloads saw 12 to 34 percent.
The ceiling on cheap 24GB tensor work was lower than anyone planned around, and a module just moved it.
You can rent the same card at $0.15/hr with hourly billing and Buy Now Pay Later, zero upfront and payment after one month.
What is the first test you run on a new GPU before you trust it with a week of training?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w44&utm_content=evening
Originally posted on linkedin_personal