LLM Inference: VRAM & Performance Calculator

1 min read Original article ↗

Pick a model, a quantization and a device — see whether it fits and how fast it runs.

Precision of the model weights. Lower uses less VRAM but costs quality.

KV cache precision. Dominates VRAM at long context.

Hardware

Hardware configuration

One machine, or GPUs spread over several nodes you lay out yourself.

Select your GPU or configure a custom device.

Number of devices1

Tensor-parallel replicas. Comms overhead is included.

12481632

Lets the model exceed VRAM — at host-bandwidth speed.

Workload

Batch size 1

Sequences processed per step. Raises throughput, costs KV cache.

Sequence length1,024

Tokens per sequence (prompt + generation). Drives KV cache.

Concurrent users 1

Simultaneous requests. Multiplies KV cache, splits per-user speed.

Inference simulation

Press play to watch this configuration generate at its estimated rate.

Comfortable

0 GB

of 0 GB usable

Generation speed

Per-token latency

Time to first token

Total throughput

Bottleneck

Memory allocation