The numbers,
with their assumptions.
The planner estimates a working memory set for an illustrative workload. It does not allocate devices, quote capacity or predict benchmark performance.
Open the planner ↗ · Planning guide · Formula reference
A deliberately simple model
parameter memory = FP32 reference GiB × precision bits / 32
activation memory = batch × activation MiB / 1024
working set = parameter memory + activation memory + 2 GiB
available budget = device capacity × 0.85
within budget = working set ≤ available budgetParameter memory is the stored FP32 reference size. The precision setting scales that part only. Activations use the explicit per-item value. A fixed 2 GiB runtime allowance and a 15% capacity reserve make the assumptions visible.
Example calculation
The default input is 12 GiB of reference parameters at FP16, a batch of 8 and 256 MiB of activations per item. The estimate is 6 + 2 + 2 = 10 GiB. A 24 GiB example capacity leaves a 20.4 GiB planning budget after the reserve.
The exported plan
The JSON includes sample, model, all four inputs, the capacity, parameter and activation contributions, reserve fraction, estimated working set and the within_budget boolean. Values are recorded before presentation rounding.
What this model does not include
Optimizer state, execution-specific temporary buffers, host memory, transfer time and distributed communication may change actual requirements. Multiple devices are not pooled. Treat the result as a starting hypothesis for profiling.