GPU dedicated server · compute unlocked · Bucharest, EU
The RTX 3050, unlocked: up to 56% more AI math from the same card
Every RTX 3050 desktop we now deliver on Ubuntu or Debian has its tensor cores running at their full rate. On our own card that took FP16 matrix work from 14.0 to 21.3 TFLOPS, inside the same 70 W power limit, with nothing for you to install. The machine is all yours: an Intel Core i5-9500, 32 GB of memory, a 256 GB NVMe SSD and an unmetered 1 Gbps port.
- FP16 +52%
- BF16 +54%
- TF32 +56%
- Checked at every boot
Factory vs unlocked
Measured on our own RTX 3050, not copied from a spec sheet
One card in one of our machines, booted from cold twice: first at the factory setting, then unlocked. Same kernel, same driver, same 70 W power limit. The test is the work AI frameworks lean on most, a large matrix multiplication in PyTorch.
Matrix multiply throughput in TFLOPS, higher is faster
8192 × 8192 matrices, PyTorch, NVIDIA driver 595
- Factory
- Unlocked
-
FP16
14.0+52%21.3
-
BF16
14.2+54%22.0
-
TF32
6.9+56%10.8
-
FP32
3.7same3.7
The peak tensor rate doubles. Counted instruction by instruction, FP16 and BF16 with FP32 accumulate went from 17.9 to 36.1 trillion operations a second. Big real-world multiplications gain about half rather than double, because the card already works at its 70 W limit and the extra speed has to fit inside the same power.
Show the figures as a table
| Precision | Factory | Unlocked | Change |
|---|---|---|---|
| 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 | same |
| Peak tensor rate, FP32 accumulate, TF/s | 17.9 | 36.1 | 2x |
Each figure is the average of five runs of fifty multiplications, after a warm-up. Your own speed-up depends on your model, batch size and software.
Where you will feel it
Faster where the tensor cores do the work, unchanged everywhere else
The unlock raises one thing: how fast the card multiplies matrices in FP16, BF16 and TF32. Anything that spends its time there speeds up; anything that does not, runs exactly as before.
Gets faster
- Prompt processing in AI inference. Reading a long prompt or document is matrix work, so the wait before the first word comes shorter.
- Fine-tuning small models. LoRA and similar runs on models that fit in 6 GB GDDR6 finish sooner in BF16 or FP16.
- Image generation. Diffusion models spend most of each step in half-precision matrix math.
- Batch inference. Embeddings, classification and transcription jobs that push many inputs through at once.
Stays as it was
- Games and 3D rendering. Graphics work does not go through this path.
- Video encoding and decoding. NVENC and NVDEC are separate engines on the chip and keep their factory speed.
- FP32 CUDA code. Ordinary single-precision math runs at 3.7 TFLOPS either way.
- Token-by-token text generation. Once a model is writing its answer, memory speed sets the pace, not the tensor rate.
How it reaches your server
Order it on Linux and the card arrives unlocked
-
Pick Ubuntu or Debian
Choose either one on the order form. We install the system, NVIDIA's driver and the compute unlock together, so the machine is online about 30 minutes after your order is approved.
-
Applied at every start
Nothing is flashed onto the card. The unlock is set while the machine boots, before the NVIDIA driver loads, and it is checked each time the system starts.
-
Factory speed as the safety net
If a boot ever cannot apply it, the card simply comes up at its factory rate and keeps working. Nothing else on the machine changes either way.
Running Windows? The same machine is available with Windows, and there the card runs at its factory speed: the compute unlock is a Linux feature. Everything else, from the processor to the unmetered port, is identical.
How we tested it
We ran it in our own rack before offering it to you
Before a single customer machine got the unlock, we put one of these desktops through a lab test: repeated cold starts, a warm restart, every benchmark in both states and a long run under full load, watching for errors the whole time.
It held. The unlock took on every start, the card logged no errors, and every result under load came out identical to the first one, bit for bit.
- Machine
- Intel Core i5-9500, 32 GB, Bucharest
- Card
- RTX 3050 6GB, 70 W limit
- System
- Ubuntu 24.04, NVIDIA driver 595
- Starts
- Unlocked on 5 of 5
- Full-load run
- 45 minutes, 0 errors
- Results checked
- Bit-identical, 0 mismatches
- Benchmark
- PyTorch, 8192² matrices
Questions
Unlocked RTX 3050: what people ask before renting one
What exactly is unlocked?
The rate at which the card's tensor cores multiply matrices in FP16, BF16 and TF32. From the factory, this card runs that math at half the rate the chip can do when it adds results in FP32. Unlocked, it runs at the full rate. The card, its 6 GB GDDR6 and its 70 W power limit stay the same.
How much faster will my workload be?
On our card, large matrix multiplications ran 52% faster in FP16, 54% in BF16 and 56% in TF32. A whole AI job gains less than that, because it also spends time on things the unlock does not touch, such as moving data or generating tokens one at a time.
Do I need to install anything?
No. Order the machine with Ubuntu or Debian and it is delivered with NVIDIA's driver and the compute unlock already in place. You get root access and can add CUDA, PyTorch or anything else on top.
Does it work on Windows?
No. The compute unlock is applied on Linux only. The same machine with Windows runs the card at its factory speed, which is still a full RTX 3050 6GB of your own.
What happens if the unlock is not applied one day?
The card starts at its factory speed and works normally. It is checked at every start, so a restart is all it takes for the next attempt.
Can I switch it off?
Yes. It is a single boot option. Remove it, power the machine off and on again, and the card returns to its factory rate. Our support team can do it for you.
Does it cost extra?
No. The unlocked card is included in the machine's price of €89.00 a month, with 32 GB of memory, the 256 GB NVMe SSD, one IPv4 address and an unmetered 1 Gbps port.
More AI throughput from the same card, at the same €89.00 a month
A whole RTX 3050 machine in Bucharest with its tensor cores at full rate, root access and an unmetered 1 Gbps port, online about 30 minutes after approval.
NVIDIA, GeForce and RTX are trademarks of NVIDIA Corporation. The compute unlock is our own configuration of the card and is not provided by NVIDIA.