Byte-Sized Brain — Results

Generated by bsb report from benchmarks/results/*.csv.

Quantization trade-offs (AMD64)

Per-variant detail

pipeline	modality	runtime	variant	arch	emulated	size_mb	accuracy	latency_ms_mean	size_reduction_%	latency_speedup_x	accuracy_delta
cnn_cifar10	vision	tflite	fp32	AMD64	False	8.522	0.854	2.609	0.0	1.0	0.0
cnn_cifar10	vision	tflite	int8	AMD64	False	2.615	0.681	2.651	69.3	0.98	-0.173
distilbert_imdb	nlp	onnxruntime	fp32	AMD64	False	255.549	0.846	72.272	0.0	1.0	0.0
distilbert_imdb	nlp	onnxruntime	int8	AMD64	False	64.269	0.84	43.459	74.9	1.66	-0.006
ffn_mnist	vision	tflite	fp32	AMD64	False	0.39	0.97	0.023	0.0	1.0	0.0
ffn_mnist	vision	tflite	int8	AMD64	False	0.102	0.97	0.034	73.7	0.68	0.0
rnn_imdb	sequence	tflite	fp32	AMD64	False	2.498	0.85	0.696	0.0	1.0	0.0
rnn_imdb	sequence	tflite	dynamic_range	AMD64	False	0.635	0.85	0.802	74.6	0.87	0.0

Notes

• size_reduction_%, latency_speedup_x and accuracy_delta are computed against each pipeline's own FP32 baseline, within the same architecture.
• Rows with emulated = True come from QEMU-emulated ARM64; their absolute latency is not comparable to native runs — treat it as a relative ratio only. See docs/methodology.md.