Profiling Whisper AI Model on RISC-V: CPU, GPU, and NPU Performance on the DC-ROMA AI PC
summary
DC-ROMA RISC-V AI PC runs Whisper.cpp across CPU, GPU and an NPU prototype on an ESIN 7702 SoC with P550 cores and up to 50 TOPS at 1.8 GHz. CPU inference on the tiny model takes ~25–30 s to transcribe 11 s of audio. GPU results are currently worse due to Vulkan / Imagination incompatibilities and Mesa driver gaps. An ND/TSMC NPU prototype is demonstrated and a local pipeline chains Whisper.cpp speech→text to an on-device LLM for real‑time English→French translation, highlighting need for better NPU access and driver work.
AI-written summary · about
chapters (7)
- 0:00Introduction and background
- 1:21DC-ROMA AI PC specs
- 2:07Whisper.cpp CPU demo
- 3:40GPU performance and challenges
- 4:24Call for software optimization
- 8:18VC-integrated demo
- 9:43Future work and conclusion
related talks
40mAdvancing AI Application Development on RISC-V Platforms
16mUnleash the superpowers of the AI PC
43mSpeeding up Whisper for speech recognition by 10x - Optimisations and walkthrough
13mKeynote: Taking RISC-V Intelligence Everywhere
05mAccelerate AI development from cloud to client
45mDeliver AI-powered experiences across cloud and edge with Windows
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
15mBuilding Gen-AI experiences using LLM and LVM