GPU guide
Choose the right GPU without being an expert.
The main thing is video memory (VRAM): the model and its working data must fit into the card's memory. Speed comes second. Use the calculator, compare cards, or skip all of it and just describe your job: we size it for you.
Pick in 30 seconds
How much memory do you need?
Choose a task and a model size. The figures are approximate: long context and large batches need more memory. If the model does not fit on one card, you need several cards joined by a fast link (NVLink).
Quick sizes: 7–8B small, 13B medium, 30–34B large, 70B very large. OCR, speech recognition and synthesis, and computer vision usually need only 8–24 GB.
| Task | Memory per 1B parameters | 7–8B model | 70B model |
|---|---|---|---|
| Inference, 16-bit | about 2 GB + 20% | 16–20 GB, one card | 140–170 GB: H200 or 2×H100 |
| Inference, compressed (4-bit) | about 0.6–0.7 GB | 5–6 GB, any gaming card | 40–48 GB: one card of 48 GB or more |
| LoRA / QLoRA fine-tuning | about 1–2 GB (up to 3 with long context) | 10–24 GB: RTX 4090 or 5090 | 48–80 GB: one H100 (QLoRA) |
| Full fine-tuning | about 16–20 GB | 120–160 GB: several cards | over 1 TB: a cluster |
| OCR, speech, computer vision | n/a | usually 8–24 GB | n/a |
Compare cards
Consumer and workstation cards.
Cheap cards for inference of small models and LoRA. Startups use them most. Note: NVIDIA's driver licence does not allow GeForce gaming cards in data centers, so for business customers we mark them as a trade-off in reliability and support.
| Card | Memory | Good for | Watch out for |
|---|---|---|---|
| RTX 3090 gaming, 2020 | 24 GB | Compressed models up to 13B, OCR, speech, experiments. Cheapest option. | Old and slow, runs hot, often second-hand. |
| RTX 4090 gaming, 2022 | 24 GB | Inference of 7–13B, LoRA up to 8B, image generation. | 24 GB is not enough for models from 30B. |
| RTX 5090 gaming, 2025 | 32 GB | Speech recognition and synthesis, LoRA, compressed models up to 30B. | Scarce, expensive for a gaming card. |
| RTX A6000 workstation, 2020 | 48 GB | Compressed models up to 30–40B, LoRA up to 13B. | Older architecture, slower than new cards. |
| RTX 6000 Ada workstation, 2022 | 48 GB | Same as A6000 but faster; video generation. | Pricey for 4090-level speed. |
| RTX PRO 6000 Blackwell 2025 | 96 GB | Compressed 70B inference, LoRA up to 30B, speech, video. | No NVLink; slower than H100 for training. |
Data-center cards (NVIDIA).
Built for servers and offered by local clouds and serious hosting providers. The main ones for most jobs are A100, H100 and H200.
| Card | Memory | Good for | Watch out for |
|---|---|---|---|
| T4 2018 | 16 GB | Light inference, OCR, classification. Cheapest data-center card. | Weak and old, not for training. |
| L4 2023 | 24 GB | Small-model inference, video analytics on many cameras; low power. | Weak for training. |
| A10 2021 | 24 GB | Compressed inference up to 13B. | Aging, being replaced by L4. |
| L40S 2023 | 48 GB | Inference up to 30B, image and video generation, video analytics. | No NVLink; weaker than H100 for training. |
| A100 2020 | 40 or 80 GB | Training and fine-tuning of mid-size models; universities and labs. | Previous generation, no FP8. |
| H100 2022 | 80 GB (NVL: 94 GB) | Training, fine-tuning, fast inference. | Pricey in Kazakhstan; US export controls apply. |
| H200 2024 | 141 GB | A 70B model on one card, steady workloads, companies keeping data in the country. | Expensive; capacity at local providers is tight. |
| B200 2025 | 180–192 GB | Large-model training and large-scale inference. | Scarce, expensive, needs strong cooling. |
| B300 2025–26 | 288 GB | The largest models. | Scarce; we have not seen offers in Kazakhstan. |
| DGX Spark mini PC, 2025 | 128 GB shared | Developers and prototypes in your own office. | Slow memory, not for production. |
Price depends heavily on the term. Local data-center prices in Kazakhstan (October 2026) vary roughly twofold between hourly and monthly terms: for an H100 about $12 per hour on-demand versus about $6 per hour on a monthly contract. Reference only, not a Hostmatch offer: your matched offer sets the real price.
Before you rent
Do you need to rent a GPU at all?
Sometimes the honest answer is no. Here are the alternatives and when each fits.
AMD MI300X / MI325X / MI355X
Data-center cards with 192–288 GB per card. Good for large-model inference. Software (ROCm) is behind CUDA and fewer ready-made tools exist.
Apple Mac mini / Mac Studio
Quiet, low-power local inference in an office. Slower than NVIDIA, no CUDA, not for training.
Your own hardware
Makes sense for steady long-term load and strict data rules. Big upfront cost, upkeep, and an MLOps engineer to run it.
Pay-per-request APIs
Hosted open models, billed per request instead of per card. Great for prototypes and light inference. Running your own fine-tuned model is harder, and data leaves the country.
Free cloud credits
Big clouds give $200–300 at sign-up. Fine for students and early experiments. One-time and gone quickly.
Hostmatch
Rent by the hour or day on servers in Kazakhstan, one contract, one support line. Best for short jobs and for data that must stay in the country.
Who needs what
Typical jobs and the card that fits.
| Who | Job | Card | How they usually pay |
|---|---|---|---|
| Startup with small models | LoRA fine-tuning, inference of 7–13B | RTX 4090 or 5090, 24–32 GB | Pay as you go, hours to days |
| Speech and language-model developer | Speech recognition and synthesis, LoRA | RTX 5090, RTX PRO 6000 | Pay as you go, about $15–20 per day |
| Video analytics | Recognition on cameras around the clock | L4, L40S, H100 or H200 | Monthly subscription |
| Universities and labs | Training and fine-tuning for a grant | A100, H100 | 1–2 months, often night hours |
| Public sector | Own language models, data in the country | H100, H200, clusters | Official request and contract |
| Banks with their own hardware | Inference inside the company | DGX Spark, H200 | Hardware purchase plus setup |
| Prototype on ready-made models | Light inference, OCR | No card needed: use an API | Subscription from about $9 a month |
Simple rule: one-off jobs on small models go to cheap cards with pay-as-you-go. Steady load and data that must stay in the country go to data-center cards on a monthly subscription.
Glossary
GPU and AI terms in plain English.
Nothing found. Try another word, or write to us.
Still unsure?
Skip the guesswork.
Describe the job in plain words. We pick the GPU model, the count and the time window, and send matched offers within one business day.
Memory sizes and sizing rules are common public specifications and are approximate. Check the exact numbers for your model before a long run.