CNN Computational Cost Calculator

Calculated results

CNN cost summary

Layer-by-layer analysis

#	Layer	Type	Input	Output	Parameters	MACs	FLOPs	Activation	Compute share

Compute distribution

The chart uses each enabled layer’s FLOP estimate.

Bottlenecks and observations

Comparison workspace

Store two result snapshots, then compare their costs.

No comparison snapshots saved.

Complete calculation report

Model setup

Input, workload, precision, and counting rules

Input height

Input width

Input channels

Batch size

Numeric precision

Workload mode

Forward passes

Dataset images

Training epochs

MAC-to-FLOP convention

Display decimals

Number format

Memory units

Training memory

Optimizer and memory assumptions

Optimizer

FP32 master weights

Activation checkpointing

Training FLOP multiplier

Includes forward and estimated backward work.

Workspace memory multiplier

Approximates temporary convolution workspace.

Hardware model

Latency, throughput, and bandwidth estimates

Device peak throughput

TFLOPS

Compute utilisation

Memory bandwidth

GB/s

Number of devices

Parallel efficiency

Fixed latency overhead

Runtime values are theoretical estimates. Real kernels, data movement, compilation, and scheduling can change performance.

Architecture builder

Create, load, reorder, and inspect layers

0 layers

Architecture preset

Import JSON

Formula used

Core CNN cost equations

A convolution output follows the standard discrete shape equation. Dilation expands the effective kernel without adding weights. Grouping reduces connected input channels per filter.

H_out = floor((H_in + P_top + P_bottom − D_h(K_h − 1) − 1) / S_h + 1)

W_out = floor((W_in + P_left + P_right − D_w(K_w − 1) − 1) / S_w + 1)

MACs = H_out × W_out × C_out × (C_in / Groups) × K_h × K_w

Parameters = K_h × K_w × (C_in / Groups) × C_out + Biases

How to use

Build and analyse a CNN

Enter the input tensor, precision, batch, and workload.
Choose hardware assumptions for theoretical runtime estimates.
Load a preset or build layers manually.
Check every layer’s dimensions and grouping values.
Select the FLOP convention used by your reference.
Calculate, inspect bottlenecks, compare snapshots, and export results.

Example data

Small classifier example

Layer	Input	Configuration	Expected output
Conv 1	32 × 32 × 3	16 filters, 3 × 3, stride 1, same	32 × 32 × 16
Max pool	32 × 32 × 16	2 × 2, stride 2	16 × 16 × 16
Conv 2	16 × 16 × 16	32 filters, 3 × 3, stride 1, same	16 × 16 × 32
Global average	16 × 16 × 32	Average each feature map	1 × 1 × 32
Dense	32 features	10 outputs	10 features

Frequently asked questions

CNN computational cost questions

What is a MAC?

A MAC combines one multiplication and one accumulation. Some references count it as one FLOP. Others count it as two FLOPs.

Why do FLOP totals differ between tools?

Tools may use different MAC conventions. They may include activations, bias, normalization, or pooling. Always compare matching assumptions.

Does parameter count predict runtime?

Not reliably. Activation sizes and data movement also matter. Hardware kernels can favour specific layer shapes.

How is grouped convolution handled?

Each filter sees fewer input channels. Input channels must divide evenly by groups. Output channels must also divide evenly.

How is depthwise convolution calculated?

The group count equals input channels. Each channel receives separate spatial filters. Pointwise mixing is counted separately when selected.

Is training memory exact?

No. Framework workspaces and saved tensors vary. The estimate exposes adjustable memory multipliers.

Why can bandwidth limit latency?

Low arithmetic intensity moves many bytes per operation. Compute units then wait for memory transfers. The calculator reports both limits.

Can this replace hardware benchmarking?

No. It provides transparent theoretical estimates. Benchmark the final exported model on target hardware.

What does activation checkpointing change?

It stores fewer forward activations during training. Backward computation may increase through recomputation. This tool adjusts stored activation memory.