Running ExecuTorch Applications With Silicon Accelera...
summary
Shows a full pipeline that trains a 11 million‑parameter Conformer in PyTorch, exports a trace, applies post‑training quantization to 8‑bit weights (8/8 or 8/16 activations), uses the ExecuTorch APIs to lower and compile for the Ethos‑U85 256‑MAC target, and packages a PTE file placed in external flash for the Alif Ensemble E8 dev kit. Measured inference plus pre/post‑processing runs on real silicon; continuous audio use estimates ~43 hours (≈36 days with duty cycle).
AI-written summary · about
chapters (8)
- 0:00ExecuTorch and Alif Semi intro
- 1:06Transformer models with silicon acceleration
- 2:50ExecuTorch pipeline steps
- 5:19Alif Ensemble E8 overview
- 7:45ExecuTorch deployment benefits
- 8:48CPU cores performance
- 9:34Power efficiency estimates
- 11:42Q&A on compiler stack
related talks
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller | Corrected Audio
22mIntroducing ExecuTorch from PyTorch Edge: On-Device AI...
27mDeploying GenAI for Audio Generation on Mobile CPUs With ExecuTorch
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller
25mExecuTorch 1.0: General Availability Status for Mobile and Embedded...
33mExecutorch: Enabling and Accelerating PyTorch Models on Edge Devices
13mAI inference for the smallest processors
20mPyTorch Edge: Developer Journey for Deploying AI Models Onto Edge Devices