Blog
Practical guides on choosing a GPU, running open AI models and keeping compute costs under control.
12 November 2026 Blender rendering on a rented GPU: Cycles headless, several machines, Octane and Redshift
Render Blender Cycles with OptiX from the command line on an RTX 4090 or 5090, split 240 frames over four machines, and price a frame before you start.
12 November 2026 Don't pay for an idle GPU: auto-stop, tmux and stopping from code
A forgotten RTX 4090 runs 15 hours overnight. Idle auto-stop cuts that to about 3. What counts as activity, tmux for night jobs, and a one-line stop from code.
12 November 2026 Fine-tune an LLM: how much VRAM and how many GPUs for QLoRA, LoRA and full
QLoRA fits an 8B model on 16 GB and a 70B on one 80 GB card; full fine-tuning needs about 16 bytes per parameter. Memory, cards, tools, hours and cost.
12 November 2026 GPU API for batch jobs: start a server at night, run the job, stop when done
A cron-ready script in curl and Python: launch a GPU, wait until it runs, start your job or Docker image, copy results and stop, with a hard spending cap.
12 November 2026 H100 vs A100 vs RTX 4090: when a data-centre GPU is worth it for AI
80 GB vs 24 GB, 3.35 TB/s vs 1 TB/s, FP8 and NVLink. The RTX 4090 is cheapest per image; the A100 and H100 win on big models. Costs per job inside.
12 November 2026 H200 vs H100: when 141 GB is worth paying for, and when 80 GB is enough
Same Hopper compute, 141 vs 80 GB and 4.8 vs 3.35 TB/s. The H200 wins on 70B in fp8, long context and lean 4-bit 235B MoE; the H100 on most other jobs.
12 November 2026 Is my data safe on a rented GPU? Who can see it, where it runs, GDPR
An honest answer: what a host could technically see, what containers and deleted disks protect, GDPR in plain words, and 6 steps to cut the risk.
12 November 2026 Move datasets to a GPU server and back: Hugging Face, rclone, rsync, scp
Pull 100 GB in about 20 minutes instead of uploading it for 5 hours. Hugging Face, rclone from S3 or Drive, rsync, Jupyter limits and traffic cost per GB.
12 November 2026 Multi-GPU training basics: DDP, FSDP and ZeRO on a rented 2–8 GPU machine
When one card isn't enough: data vs model parallel, memory per GPU with FSDP and DeepSpeed ZeRO, NVLink vs PCIe, renting 2–8 cards and avoiding NCCL hangs.
12 November 2026 Own LLM or API: when a rented GPU is cheaper than paying per token
Four monthly scenarios in euros: one user, a team of ten, a 24/7 product bot and a 440M-token batch job. Where the API wins, where your own GPU does.
12 November 2026 Your own private ChatGPT: Open WebUI with Ollama or vLLM on a rented GPU
A ChatGPT-style chat on Qwen3 or MiMo in about 10 minutes: which card for 8B, 32B and 70B, what stays private, how to share it with a team, monthly cost.
12 November 2026 RTX 3090 vs RTX 4090 for AI: same 24 GB, what you actually pay per job
The 4090 makes images 1.6–2.5× faster but LLM text only 10–20% faster. When the RTX 3090 is cheaper per image, per token and per LoRA.
12 November 2026 RTX 5090 for AI: what 32 GB of GDDR7 runs that a 24 GB card can't
8 GB more than a 4090 and 78% more bandwidth: 32B models with long context, Wan 2.2 14B in fp8, Flux in bf16. Speed, CUDA 12.8 and when you need 48 GB.
12 November 2026 Run Llama 70B on one GPU: 4-bit or FP8, 48 GB or 80 GB, and real speeds
Llama 3.3 70B in 4-bit is 40–43 GB: one 48 GB card at 8–16k context, 80 GB for long documents or several users. Formats, commands, tokens/s.
12 November 2026 SSH into a GPU server from Windows 10 and 11: key, connect, fix errors
Five PowerShell commands, no PuTTY: create a key, find the .pub file, connect. Plus Permission denied fixes, VS Code, scp and the no-SSH Jupyter way.
12 November 2026 Text to speech on your own GPU: Qwen3-TTS, VoxCPM2 and Chatterbox voice cloning
Three free text-to-speech apps on a 12 GB card: ready voices, up to 30 languages, a cloned voice from 5 seconds, cost per hour of narration, consent rules.
12 November 2026 Train a Flux LoRA on your own photos: 20 images, 24 GB and about two hours
How many photos, which captions, ai-toolkit or FluxGym, 24 or 16 GB, 2,000 steps in 1–2.5 h on an RTX 4090, and what happens to your photos afterwards.
12 November 2026 vLLM vs Ollama vs llama.cpp: which to run your LLM with, and the commands
Ollama for one person in five minutes, llama.cpp for control, vLLM for an API with many users: speed, formats, memory behaviour and start commands.
12 November 2026 Wan 2.2 video on a rented GPU: 5B or 14B, how much VRAM and how long per clip
Wan 2.2 5B renders 5 s of 720p on a 24 GB card in 4–9 minutes; 14B needs 48 GB and looks much better. Which card, clip limits and cost per clip.
12 November 2026 Whisper transcription on a GPU: large-v3 or turbo, faster-whisper and WhisperX speakers
Whisper on a rented GPU: 3–10 GB of VRAM, 20 to 250 hours of audio per GPU-hour, cost per 100 hours, SRT subtitles and speaker labels with WhisperX.
11 November 2026 Best GPU for Flux, SDXL and ComfyUI: 16 GB vs RTX 3090, 4090, 5090 and 48 GB cards
SDXL runs on 12 GB, Flux on 16 GB, Qwen-Image and HiDream need 24 GB, Wan 2.2 14B 32–48 GB. Which card to rent for ComfyUI and when the faster one is cheaper.
11 November 2026 ComfyUI on a rented GPU: your own models, LoRAs, custom nodes and the 502 page
Run ComfyUI on a rented GPU from €0.59 an hour: when it is ready, where models and LoRAs go, how to add custom nodes and keep them for next time.
11 November 2026 Cost per image, video clip and million tokens on a rented GPU, worked out
1,000 Flux images, one 5-second Wan 2.2 clip, a million tokens from 8B to 70B, a Flux LoRA: all-in costs with live euro prices and the maths shown.
11 November 2026 CUDA driver version mismatch: fix PyTorch and ComfyUI on a rented GPU
"Driver too old", "no CUDA GPUs available", sm_120 errors: match the PyTorch wheel (cu126, cu128, cu130) to the driver in 2 minutes, or pick the right machine.
11 November 2026 CUDA out of memory: how to fix it in ComfyUI, LLMs and training
Fix CUDA out of memory by workload: FP8 and tiled VAE for Flux, context and KV cache for LLMs, 8-bit optimisers for training, and when a bigger card is cheaper.
11 November 2026 Google Colab vs a rented GPU: disconnects, compute units and the real cost per hour
Colab is free or cheap until sessions drop after a few hours and units run out. When Colab is enough, and when a GPU from €0.30 an hour is better.
11 November 2026 GPU rental billing explained: per minute, 15-minute minimum, traffic and storage
When billing starts and stops, why the price is locked at launch, what traffic and the saved folder cost, and a full session bill worked out minute by minute.
11 November 2026 How much VRAM do I need? A table for LLMs, Flux, Wan 2.2 video and training
VRAM = weights × bytes per parameter + KV cache + overhead. A 70B model in 4-bit needs about 43 GB of weights. Tables by model, quantisation and card.
11 November 2026 Rent or buy an RTX 4090 or RTX 5090 for AI: break-even hours in 2026
A used RTX 4090 costs about €2,000 and a 5090 about €4,500 (Nov 2026). Buying beats renting from roughly 90 GPU-hours a month. The full sum inside.
11 November 2026 What happens to your files when you stop a GPU server, and how to keep them
Stop wipes the server disk. Save and stop keeps /workspace for €0.03 per GB a month, on any card. When to store models, when to download again.