Advertised VRAM is not the planning budget.
A topology passes only when estimated per-device weights, cache and workspace remain within 85% of the selected GPU's advertised memory. The remainder protects runtime variance.
Review all assumptions →Multi-GPU
calculator.
Compare 1×, 2×, 4× and 8× tensor-parallel plans against a consistent per-GPU memory ceiling.
Minimum tested topology2× NVIDIA H10040.9 GiB estimated allocation per GPU
1×81.6 GiBDoes not fit
2×40.9 GiB60% of planning budget
4×21.5 GiB32% of planning budget
8×11.7 GiB17% of planning budget
Uses an 85% capacity ceiling. It estimates memory fit, not throughput or interconnect performance; tensor-parallel communication can make a technically valid topology operationally poor.
Memory fit does not guarantee good throughput.
Interconnect, collective communication, kernel support, batch scheduling and host topology can dominate performance. Treat the result as a capacity filter before benchmarking.
Compare GPU profiles →Weights and cache shard differently.
Weights divide by GPU count in this planning model. KV cache divides only up to the model's KV-head count, while a minimum per-device workspace remains.
Inspect KV cache separately →