Kernel Benchmarks

Apollo510 EVB · Cortex-M55 + Helium MVE

heliaRT is up to 706× faster than upstream LiteRT for Micro

Across 36 operators benchmarked on real hardware, heliaRT matches or beats upstream LiteRT for Micro on every single operator — with dramatic gains on activations, reductions, and data-movement ops that LiteRT for Micro leaves unoptimized.

36 / 36 Operators equal or faster

706× Peak speedup

0 Regressions

Speedup at a Glance

50–706× — LiteRT for Micro falls back to scalar C reference · 1.3–13× — heliaRT MVE-optimized paths · ~1× — Both use CMSIS-NN / Helium MVE

Detailed Results

#	Operator	heliaRT Cycles	LiteRT Cycles	Speedup
1	CONV_2D	1,621,810	1,642,809	1.01×
2	DEPTHWISE_CONV_2D	613,204	636,011	1.04×
3	FULLY_CONNECTED	20,844	27,950	1.34×
4	TRANSPOSE_CONV	358,543	359,397	1.00×
5	AVERAGE_POOL_2D	98,581	98,590	1.00×
6	SOFTMAX	9,379	9,385	1.00×
7	ADD	218,395	218,369	1.00×
8	MUL	95,203	132,152	1.39×
9	LOGISTIC	1,015	53,208	52.4×
10	PAD	6,357	6,354	1.00×
11	RELU	98,760	985,349	10.0×
12	HARD_SWISH	65,938	870,531	13.2×
13	SUB	218,337	1,921,818	8.8×
14	CONCATENATION	38,745	93,984	2.4×
15	SPLIT	251,288	801,614	3.2×
16	STRIDED_SLICE	2,250	13,900	6.2×
17	MEAN	22,408	2,230,860	99.6×
18	REDUCE_MAX	3,808	2,688,873	706×
19	BATCH_MATMUL	214,753	290,100	1.35×
20	MAX_POOL_2D	38,405	38,405	1.00×
21	PADV2	6,377	6,450	1.01×
22	TRANSPOSE	21,820	21,701	~1.00×
23	MAXIMUM	7,811	9,861	1.26×
24	MINIMUM	7,809	9,860	1.26×
25	RELU6	8,544	197,167	23.1×
26	TANH	82,301	7,945,776	96.5×
27	LEAKY_RELU	143,929	968,936	6.7×
28	EQUAL	420,227	1,547,914	3.7×
29	GREATER	512,563	1,498,903	2.9×
30	LESS	518,818	1,512,053	2.9×
31	RESHAPE	5,425	9,596	1.77×
32	SPLIT_V	233,019	760,585	3.3×
33	PACK	10,697	98,739	9.2×
34	NOT_EQUAL	115,516	115,544	1.00×
35	FILL	2,422	33,143	13.7×
36	ZEROS_LIKE	8,581	16,840	1.96×

Why It Matters

No silent fallbacks

Upstream LiteRT for Micro only optimizes ~14 operators with CMSIS-NN. The rest — activations like RELU and LOGISTIC, reductions like MEAN and REDUCE_MAX, data-movement ops like SPLIT and CONCATENATION — fall back to scalar C reference code that ignores Helium MVE entirely. These "silent fallbacks" can dominate inference time in real models, even though they look cheap on paper.

heliaRT closes this gap with dedicated MVE kernels for 36 operators — every CMSIS-NN-optimized op plus 22 more. The result: no operator is left behind, and your model runs at full hardware capability from end to end.

What this means for your product

Lower latency

Operators like MEAN (99.6×) and REDUCE_MAX (706×) often dominate post-convolution pooling and classification stages. heliaRT eliminates these hidden bottlenecks.

Longer battery life

Fewer cycles per inference means less active time and more time in sleep. Every cycle saved is energy saved.

Drop-in compatible

heliaRT uses the same LiteRT API and .tflite models. Switch backends without changing your application code.

No regressions

Across all operators, heliaRT matches or beats LiteRT for Micro. Zero trade-offs, zero surprises.

Test Environment

Parameter	Value
Board	Apollo510 EVB (Cortex-M55 + Helium MVE)
Toolchain	arm-none-eabi-gcc v15.2.1
heliaRT	v1.16.0
LiteRT for Micro	upstream LiteRT for Micro (2.3a snapshot)
Iterations	100 (10 warmup)
Quantization	int8 (all models)

Methodology

Each operator is exercised by a single-operator int8 TFLite model (input shape [1,32,32,16] for spatial ops, appropriate shapes for non-spatial ops). All cycle counts are median values over 100 iterations after 10 warmup iterations. The same Apollo510 EVB and GCC toolchain are used for all runs. Speedup is calculated as litert_cycles / helia_cycles.

Toolchain Comparison (ATfE vs GCC)

For the first published benchmark — ATfE 22.1 vs arm-none-eabi-gcc 14.2 across the MLPerf Tiny v1.1 suite on Apollo510 — see Toolchains → Why ATfE. That section includes the full methodology, a per-model results table, and a Chart.js plot of latency, energy, and efficiency improvements.