What it calculates
Weight memory is parameter count × precision. Runtime memory adds the chosen overhead. Token rate divides usable aggregate memory bandwidth by model-weight size.
Working now Starter lab / NVDA
Forefront Compute
Lab
Estimate model-weight memory, device fit and a bandwidth-bound token-rate ceiling for an inference setup.
Working starter tool / v0.1
Model capacity estimator
Estimate model-weight memory, device fit and a bandwidth-bound token-rate ceiling for an inference setup.
What it does not prove
The token rate is a bandwidth-only planning estimate for weight streaming, not a benchmark. It excludes KV cache growth, activations, compute limits, communication and serving overhead.
Build on the mission
Use it, inspect it, prepare the next build.
Independent starter tool. This lab is functional software, but it is intentionally narrow and does not claim to complete the full Open Compute Bench roadmap. Benchmarks must disclose methodology, software versions and hardware configuration.