TVM Benchmark

Jump to bottom

Jack Yu edited this page Dec 14, 2018 · 10 revisions

Below we tested Mobilefacenet(dim128) on different platforms.

	@TVM	@NCNN	@MXNet
Macbook r2015 2.2G@ 4 Threads	6.4ms	No SIMD optimization on x86	28ms
Firefly Rk 3399@A72 * 1	69.24ms	72ms	[x]
Firefly Rk 3399@Mali GPU[fp32]	70.32ms	[x]	[x]
Firefly Rk 3399@Mali GPU[fp16]	45.20ms	[x]	[x]
Raspberry Pi 3b@A53 * 1	308.24ms	163.41ms	[x]

Tips: Now, MXNet PReLu layer will cause compiling error when using TVM on ARM cpu.