AI hardware planning
GPU fit calculator for LLM inference
Add weights, KV-cache memory, overhead, and a candidate card’s capacity. The calculator tells you whether the capacity plan fits and compares a bandwidth assumption with your target.
Free · no sign-in · updates instantly
1. Enter workload and card assumptions
After precision or quantization
At peak context and concurrency
Runtime buffers and fragmentation
Use the unit consistently
Use a common unit such as GB/s
A comparison target, not a benchmark
Required VRAM0 GB
Weights + cache + overhead
Headroom0 GB
Capacity minus requirement
Capacity resultChecking…
Memory fit only
Bandwidth ratio0×
Candidate ÷ target
Limit: the bandwidth ratio is not tokens per second. Benchmark the exact model, runtime, precision, context, batch, and hardware.
Formula
required VRAM = (weights + cache) × (1 + overhead ÷ 100)
headroom = candidate capacity − required VRAM
bandwidth ratio = candidate bandwidth ÷ target bandwidth
A positive headroom means the stated capacity assumptions fit. It does not guarantee that all advertised VRAM is available to the runtime or that the workload will meet a latency target.
Read the full explanation
See how to choose a GPU for LLMs for the decision framework. Then refine the inputs with the VRAM guide and the KV-cache guide.