ArmFit Performance Report

Model: tinycnn.pte � Generated 2026-08-14 14:51:24 UTC � Mode: dev

PROJECTED (dev-mode estimate)

Baseline (SME2 OFF)

210.29ms

SME2 ON

130.25ms

38.1%

Improvement

Faster

210.29 ms → 130.25 ms

Top Bottlenecks (baseline)

Conv2D

41.7%

Very High

DataMovement

21.2%

High

GEMM/Linear

16.3%

Medium

Elementwise

12.0%

Medium

Normalization

6.3%

Low

Other

2.5%

Low

Category	OFF (ms)	ON (ms)	Δ ms	Δ %
Conv2D	87.73	25.89	-61.84	-70.5%
DataMovement	44.53	57.22	+12.69	+28.5%
GEMM/Linear	34.19	8.39	-25.8	-75.5%
Elementwise	25.3	22.07	-3.23	-12.8%
Normalization	13.31	12.2	-1.11	-8.3%
Other	5.22	4.47	-0.75	-14.4%

Conv2D consumes 41.7% of baseline runtime and is your largest bottleneck before optimization.
SME2 reduced the cost of Conv2D, which shifted the largest remaining bottleneck to DataMovement. This is a projected (dev-mode estimate) comparison � investigate DataMovement next.
Overall latency changed by 38.1% (faster) with SME2 enabled (projected (dev-mode estimate)): 210.29 ms -> 130.25 ms.
DataMovement grew from 44.53 ms to 57.22 ms after SME2 (projected (dev-mode estimate)) � its relative share of runtime increased and it's now a bigger optimization target than before.
Note: these SME2 numbers are a dev-mode projection based on Arm's published category-level speedup ranges, not a real hardware measurement. Run with --mode arm on Arm SME2-capable hardware for real numbers before the final demo.