NewRTX 3050 dedicated, unlocked on Linux
Same RTX 3050, same power, a lot more AI per second
Our RTX 3050 desktops now leave the rack with the card's tensor throughput unlocked. The card is the one you know, with 6 GB GDDR6 and a 70 W limit. What changes is how fast it gets through the half-precision matrix math behind inference, fine-tuning and image generation. You rent the whole machine: Intel Core i5-9500, 32 GB RAM, 256 GB NVMe SSD, in Bucharest.
01Factory against unlocked
The numbers from our bench, in all four precisions
We booted the same RTX 3050 twice from power-off, once stock and once unlocked, and timed the same PyTorch job each way: multiplying two 8192 by 8192 matrices, over and over. Driver, kernel and the 70 W cap did not change between the runs.
TFLOPS per precision, taller is faster
PyTorch matrix multiply, NVIDIA driver 595, average of five runs
- Factory
- Unlocked
- 14.0 21.3
- 14.2 22.0
- 6.9 10.8
- 3.7 3.7
The raw tensor rate is twice the factory figure: 36.1 instead of 17.9 trillion FP16 operations a second with FP32 accumulate. A full-size multiplication lands at roughly one and a half times, since the card already runs flat out at 70 W and the faster cores share the same power budget.
Read the results as a table
| Precision | Factory | Unlocked | Difference |
|---|---|---|---|
| FP16, TFLOPS | 14.0 | 21.3 | +52% |
| BF16, TFLOPS | 14.2 | 22.0 | +54% |
| TF32, TFLOPS | 6.9 | 10.8 | +56% |
| FP32, TFLOPS | 3.7 | 3.7 | no change |
| Peak tensor rate, FP32 accumulate, TF/s | 17.9 | 36.1 | 2x |
These are our measurements on one card. A real workload also spends time outside the tensor cores, so expect your own gain to sit somewhere below these figures.
02Does your work speed up?
Check your job against this list before you order
The unlock is narrow on purpose. It touches the tensor cores and nothing else, so it helps AI work in half precision and leaves the rest of the card as it came.
Long prompts and documents for a language model
The model reads your input as one big batch of matrix math before it writes a word.
FasterLoRA and other small fine-tunes
Training steps in BF16 or FP16 on a model that fits in 6 GB GDDR6.
FasterStable Diffusion and other image models
Each denoising step is half-precision tensor work.
FasterBatch jobs: embeddings, tagging, transcription
Many inputs pushed through a model at once keep the tensor cores busy.
FasterA chatbot writing its reply word by word
Generation speed is set by how fast the card reads its memory.
SameVideo encoding with NVENC
The encoder is its own block on the chip and keeps its factory speed.
SameGames, rendering and FP32 CUDA
None of these run through the unlocked path.
Same03From order to first boot
You pick the system, we take care of the unlock
- Order
Choose Ubuntu or Debian
The machine is set up with NVIDIA's driver and the unlock already in place, and goes online about 30 minutes after approval.
- Every boot
Set fresh, then checked
The setting lives in memory, not in the card's firmware. It is applied as the system starts and verified each time.
- Fallback
Factory speed, never broken
On a start where it cannot be applied, the card just runs at its stock rate until the next restart.
Delivered unlocked, with root access to add CUDA, PyTorch and the rest of your stack.
Also available on the same machine, with the card at its factory speed: the unlock is Linux-only.
04Lab notes
What we checked before we switched it on for customers
One of these desktops spent an afternoon on our bench in Bucharest with Ubuntu 24.04 and NVIDIA driver 595. These are the results that mattered to us.
- 5 / 5starts came up unlocked, cold and warm
- 45 minat full load without a single GPU error
- 0results that differed from the first run, checked bit by bit
- 70 Wpower limit, the same in both states
- 2xthe peak tensor rate with FP32 accumulate
- 3.7FP32 TFLOPS, unchanged, as expected
05Asked often
Unlocked RTX 3050 server: short answers
Is this a different card from a normal RTX 3050?
No. It is the same RTX 3050 6GB with the same 6 GB GDDR6 and 70 W limit. The unlock changes one setting at start-up: the rate at which its tensor cores multiply in FP16, BF16 and TF32 while adding in FP32.
How much faster is it really?
In our PyTorch test, +52% in FP16, +54% in BF16 and +56% in TF32, with FP32 unchanged. A complete AI job gains less, because part of its time goes on work the tensor cores do not do.
Which operating systems get the unlocked card?
Ubuntu and Debian. If you order the machine with Windows, it runs the card at factory speed.
Is anything written to the card?
No. Nothing is flashed. The setting is applied each time the machine starts and disappears when it powers off, which is why it is checked at every boot.
What if I want the factory behaviour back?
Ask support or remove the one boot option yourself, then power the machine off and on. The card starts at its stock rate from then on.
What does it cost?
Nothing on top of the machine: €89.00 a month for the desktop with its unlocked card, 32 GB RAM, 256 GB NVMe SSD, one IPv4 address and an unmetered 1 Gbps port.
Put the unlocked RTX 3050 to work for €89.00 a month
A dedicated desktop in Bucharest, tensor cores at full rate on Ubuntu or Debian, root access and unmetered 1 Gbps.
NVIDIA, GeForce and RTX are trademarks of NVIDIA Corporation. The compute unlock is set up by us and is not an NVIDIA product.