Which GPU should you rent?
Every machine in the catalog is one of the NVIDIA cards below. The short version: memory decides what fits, the generation decides how fast it runs, and the number of cards decides how far you can scale.
Start with memory
Before speed, check that your job fits. A model's weights, its activations and, for LLM serving, the KV cache all have to sit in GPU memory. If they don't fit, you either split the model across several cards or it doesn't run at all.
- Up to 24–32 GB: image generation, small and quantized LLMs (7–14B), most rendering scenes.
- 48–96 GB: 30–70B models quantized, fine-tuning with LoRA, large scenes and simulations.
- 80–180 GB per card, several cards: full-precision 70B+ models, pre-training and full fine-tuning.
In the catalog, the GPU memory column already shows the total for all cards in the machine.
Data-center cards
| GPU | Memory per card | Best for |
|---|---|---|
| NVIDIA B200 | ~180 GB HBM3e | The newest generation. Training and serving the largest models. |
| NVIDIA H200 | 141 GB HBM3e | Same compute as H100 with much more and faster memory — long contexts, big batches. |
| NVIDIA H100 | 80–94 GB HBM3 | The standard for training and high-throughput inference. SXM versions are faster than PCIe. |
| NVIDIA A100 | 40 or 80 GB HBM2e | Previous generation, still strong and the cheapest way into HBM memory. |
| NVIDIA L40S | 48 GB GDDR6 | Inference and graphics in one card. |
HBM memory is several times faster than the GDDR memory in workstation and gaming cards. For LLM inference, where speed is mostly limited by memory bandwidth, that difference shows up directly in tokens per second.
Professional cards
| GPU | Memory per card | Best for |
|---|---|---|
| RTX PRO 6000 (Blackwell) | 96 GB GDDR7 | The most memory outside the data-center line. 70B models quantized on one card, rendering. |
| RTX 6000 Ada | 48 GB GDDR6 | Inference, fine-tuning and professional rendering. |
| RTX PRO 5000 / 5880 Ada | 48 GB | A cheaper step into 48 GB. |
Consumer cards
| GPU | Memory per card | Best for |
|---|---|---|
| RTX 5090 | 32 GB GDDR7 | Very fast per dollar for image and video generation and small LLMs. |
| RTX 4090 | 24 GB GDDR6X | The cheapest serious GPU. Great for experiments and Stable Diffusion-type work. |
Machines with 4 or 8 of these cards are a good fit for jobs that split into independent pieces — batch rendering, many parallel inference workers, hyper-parameter sweeps.
One card or eight?
Several cards help when your framework can use them: data-parallel training, tensor-parallel serving (vLLM, TensorRT-LLM), or many independent jobs. Cards in data-center machines are linked by NVLink, which makes splitting one model across them much more efficient than over PCIe.
If you're unsure, start with one card for a day, measure, then rent the bigger machine for the real run.