GPU server calculator
GPU server calculator: What does your model need?
Choose a model. Define your workload. Estimate VRAM, RAM and storage with transparent assumptions, not blanket hardware promises.
Workload and memory assumptionsAdjust the parameters to your scenario.
Preparing calculation …
Good decisions need context.
How much VRAM does an LLM need?
GPU memory holds model weights, the KV cache and runtime buffers. For uniform attention architectures, the cache calculation uses layers, KV heads, head dimensions, precision, context and concurrent requests. Longer context and more simultaneous requests increase the cache. Weights are not duplicated for every request within one model instance.
Quantization and model size
Fewer bits per weight reduce memory use. Quantized formats also contain scales, blocks and sometimes higher precision tensors. Parameter count multiplied by bit width is therefore a labeled approximation. Mixture-of-experts models require the total resident weights, not only their active parameters.
RAM, VRAM and unified memory
A PC with a discrete GPU has separate RAM and VRAM pools. CPU offloading can free GPU memory but requires RAM and may increase data transfers. Apple Silicon shares physical memory between CPU and GPU. Multiple GPUs need a supported distribution strategy; their capacities are not simply added together.
Images and audio need separate profiles
Vision models add image encoders, projectors and visual tokens. Image generation peak memory depends on pipeline, resolution, batch size and memory optimizations; step count does not simply multiply peak memory. For transcription, chunk size is distinct from total audio duration. Speech synthesis may require language, acoustic and vocoder components. If a complete profile is unavailable, the calculator uses a conservative value explicitly labeled as an estimate.
Buy hardware, rent a GPU or use an API?
Owned hardware incurs purchase, electricity and ancillary costs. GPU rental charges depend on booked hours, even when utilization is low. APIs bill documented units such as tokens, images or audio minutes. Compare equivalent workloads and sufficiently complete cost components. Memory fit does not prove that a machine can process your monthly volume in the available hours.
Method and planning reserve
The default 20 percent reserve is an editable planning assumption. Published hardware guidance takes priority; otherwise checkpoint size or parameter count, precision and task-specific working memory are combined conservatively. Software support and speed are evaluated separately. Without a matching measurement for the model, variant, hardware, software version and workload, no token rate is promised. Validate the actual workload before making a procurement decision.
Frequently asked questions
Do I need an account?
No. All results are freely accessible. Your requirements stay in this browser unless you explicitly share the configuration.
Does sufficient memory guarantee speed?
No. Speed requires a matching measurement. The tools show memory, compatibility and performance separately.
How are planning values derived?
Published hardware guidance takes priority. Otherwise, the calculator derives a conservative value from checkpoint size or parameter count, precision and task-specific working memory. The result links its basis and remains distinct from a measured benchmark.