6 workloads
The right server
for what you run.
Training, fine-tuning, serving, generation, rendering or research: each workload needs something different from a GPU. Start from yours.
Six workloads, sized.
Each page explains what matters for the workload, the servers we recommend, a sizing table and the software people use.
- LLM training Pre-train or continue-train on full 8-GPU nodes: NVLink 5 moves 1.8 TB/s per GPU between cards. Best fit8× B300$32,152/mo 8× B2008× MI355X Read the guide
- Fine-tuning LoRA and QLoRA runs on 70B-class models fit on a single 96–141 GB card, with room for long sequences. Best fit1× H200$2,279/mo 1× RTX PRO 60001× H100 Read the guide
- Inference & serving Memory decides how large a model you can serve and how long its context can be: up to 288 GB on one GPU. Best fit1× MI355X$1,409/mo 1× H2001× L40S Read the guide
- Image & video Diffusion and video models run fast on GDDR7 cards with FP4 and FP8 tensor cores. Best fit1× RTX 5090$329/mo 1× RTX PRO 60001× L40S Read the guide
- 3D rendering & VFX RT cores for Blender Cycles, Octane and Redshift, with up to four cards in one server. Best fit4× RTX 5090$1,316/mo 4× RTX 40904× RTX PRO 6000 Read the guide
- Research & HPC Simulation and scientific code that needs strong FP64 next to AI throughput. Best fit4× MI355X$5,636/mo 4× H2004× H100 Read the guide
Model size → cheapest server.
The cheapest configuration that holds a model of each size, by our rule of thumb. Mixture-of-experts models count by total parameters.
| Model size | FP8 inference | 4-bit inference | QLoRA fine-tune | Full training (mixed precision, Adam) |
|---|---|---|---|---|
| 8Bparameters | 1× RTX 5080$199/mo | 1× RTX 5080$199/mo | 1× RTX 5080$199/mo | 1× MI355X$1,409/mo |
| 32Bparameters | 2× RTX 4090$538/mo | 1× RTX 4090$269/mo | 1× RTX 5090$329/mo | 4× MI355X$5,636/mo |
| 70Bparameters | 1× RTX PRO 6000$959/mo | 2× RTX 5090$658/mo | 2× RTX 5090$658/mo | 8× MI355X$11,272/mo |
| 120Bparameters | 1× MI355X$1,409/mo | 1× RTX PRO 6000$959/mo | 4× RTX 5090$1,316/mo | Beyond one server |
| 405Bparameters | 2× MI355X$2,818/mo | 1× MI355X$1,409/mo | 2× MI355X$2,818/mo | Beyond one server |
| 671Bparameters | 4× MI355X$5,636/mo | 2× MI355X$2,818/mo | 2× MI355X$2,818/mo | Beyond one server |
FP8 ≈ 1.2, 4-bit ≈ 0.65, QLoRA ≈ 0.75 and full training ≈ 18 bytes per parameter, 92% of GPU memory usable. For other sizes and precisions, use the sizing helper.
Know your model? Size it.
Enter a parameter count and a precision: we list every server that fits, cheapest first.