ArmNN - AshokBhat/ml GitHub Wiki

About

• An inference engine for Arm's CPUs, GPUs and NPUs
• Bridges the gap between existing NN frameworks and the underlying IP.
• Open source and free

Architecture

• Runs on top of Arm Compute Library

Performance

• [ArmNN]] vs [TFLite for RESNET-50 Inference (Stream Latency - Lower is better)

System	Stream Latency (ms)	SoC	Cores	Software
Firefly-RK3399 (firefly)	391.02	Rockchip RK3399	2xA72+4xA53	ArmNN v19.08 (Neon)
Firefly-RK3399 (firefly)	695.11	Rockchip RK3399	2xA72+4xA53	TFLite v1.15.0-rc2
Raspberry Pi 4 (rpi4)	448.31	Broadcom BCM2711B0	4xA72 @ 1.5 GHz	ArmNN v19.08 (Neon)
Raspberry Pi 4 (rpi4)	1,916.65	Broadcom BCM2711B0	4xA72 @ 1.5 GHz	TFLite v1.15.0-rc2
Linaro HiKey960 (hikey960)	494.9	HiSilicon Kirin960	4xA73+4xA53	ArmNN v19.08 (Neon)
Linaro HiKey960 (hikey960)	518.07	HiSilicon Kirin960	4xA73+4xA53	TFLite v1.15.0-rc2
Huawei Mate 10 Pro (mate10pro)	494.92	HiSilicon Kirin970	4xA73+4xA53	ArmNN v19.08 (Neon)

See also

• Frameworks: [TensorFlow]] ](/AshokBhat/ml/wiki/[PyTorch) | [MXNET]] | ONNX
• [Inference engine]]s: [[ArmNN]] ](/AshokBhat/ml/wiki/[OpenVINO) | [TensorRT]] | [Core ML]] ](/AshokBhat/ml/wiki/[[Tengine)