Start with VRAM requirements
Provide the model, inference or training use, precision, batch size and context length. Check VRAM specifications even when GPU model names match; validate with your workload when unsure.
Multiple GPUs need software support
Multiple GPUs do not automatically give an application one pooled block of memory. Check multi-GPU support, parallelism and communication requirements before choosing the count.
Do not buy on GPU count alone
Slow data loading, insufficient CPU preprocessing or unsupported parallelism can leave extra GPUs waiting. Assess CPU, RAM, storage and upload time alongside GPU capacity.
Agree on the software environment
Confirm OS, drivers, framework, disk space and service term before ordering. Use hardware specifications to shortlist options and your own model tests to assess throughput.