VRAM GB → GiB / MiB
Your GPU box says 80 GB on the box. PyTorch OOMs at 74.5 GB. That is not a driver bug — it is the GB vs GiB gap. Vendors sell decimal gigabytes. CUDA allocates binary gibibytes. A 6.9% haircut at every scale. Close the gap before your training run hits a wall at 3 AM.
Bidirectional VRAM Converter
Put in the number printed on the GPU box. Get the number CUDA actually gives you. The difference is why your 16 GB card reports 14.9 GiB — and why your 80 GB A100 shows 74.5 GiB.
Common GPU Reference Table
Standard consumer and datacenter GPU VRAM specs with binary-equivalent values.
| GPU | GB (Dec) | GiB (Bin) |
|---|---|---|
| RTX 4060 | 8 | 7.45 GiB |
| RTX 4070 | 12 | 11.18 GiB |
| RTX 4090 | 24 | 22.35 GiB |
| RTX 5090 | 32 | 29.80 GiB |
| H100 SXM | 80 | 74.51 GiB |
| H200 SXM | 141 | 131.32 GiB |
| B200 | 192 | 178.81 GiB |
Why GB ≠ GiB Matters for ML Engineering
Memory vendors use decimal gigabytes: 1 GB = 10⁹ bytes = 1,000,000,000 bytes. Operating systems and CUDA drivers allocate in binary gibibytes: 1 GiB = 2³⁰ bytes = 1,073,741,824 bytes. The 7.4% discrepancy compounds on large models — a 70B parameter model in FP16 expects ~140 GB of weights, but if you spec a 160 GB card thinking it covers the allocation, you will hit an OOM at the CUDA driver level because the runtime sees only ~149 GiB.
This tool computes the exact binary capacity from your hardware spec so you can right-size your model sharding, gradient accumulation, and KV-cache budget before launching training or inference jobs.