Pengutronix reposted this
New on Linumiz Tech Talk: Running a Neural Network on the NPU of the NXP i.MX 8M Plus - driven entirely by open source. What does it take to run inference on a vendor NPU with no proprietary blob? In this talk, our engineer SANTHOSH C C walks through the full open-source stack - Mesa, Etnaviv, and the Teflon delegate - that turns the VeriSilicon VIPNano-SI+ into a fully accelerated, INT8-native inference engine. The payoff: ~22x faster inference on MobileNetV1 - from 166.9 ms on the CPU down to 7.67 ms on the NPU - at 30 FPS, all on 100% mainline Mesa 24.1. Inside the talk: a) Why a dedicated NPU matters at the edge - throughput per milliwatt, not just raw speed b) CPU vs GPU vs NPU for INT8 inference c) The Gallium3D split that lets an NPU ride graphics infrastructure d) Etnaviv, the reverse-engineered mainline driver for VeriSilicon silicon e) Teflon - the INT8-native, hardware-independent TFLite delegate merged into Mesa 24.1 f) The clean-room reverse-engineering effort that made it all possible Hats off to the upstream work by Tomeu Vizoso, Philipp Zabel, and Ideas On Board that this builds on. If you're working on edge AI, embedded Linux, or open-source graphics/compute drivers, this one's for you. Watch here: https://lnkd.in/ga_sEiKr Linumiz Team: Alexpandi M Parthiban N Saravanan Sekar Karthi Krishna SANTHOSH C C Sanjay V Girinandha M Balaji R.G Thamaraimanalan M GIRUBASHINI GOVINTHARAJ Gopika Ayyappan Samurthika T Bharathiraja Nallathambi #EmbeddedAI #NPU #iMX8MPlus #OpenSource #Mesa #Etnaviv #Teflon #EdgeAI #EmbeddedLinux #TensorFlowLite #Linumiz