Demo: NVIDIA's use of Catapult HLS for Building ML Inference Accelerators
Talkat RISC-V Summit North America 2023 · on Wednesday, 2023-11-08· at 13:05· Expo Hall - Exhibit Hall A - Demo Theater·
summary
NVIDIA uses Catapult HLS with the Matchlib modeling methodology to write accelerator hardware in C++/SystemC, synthesize a 100% pipelined PE (vector MAC) array, and go from code to placed silicon in about 24 hours. The delivered system is a 6x6 multi-chip module of dies each with a 4x4 PE array and a small RISC-V controller, achieving trillions of ops/sec and world‑leading energy efficiency. Pre‑HLS simulation runs 30–300x faster than RTL with ≈3% accuracy and about 10× less code, enabling small ~10‑person teams instead of hundreds.
AI-written summary · about
chapters (5)
- 0:00Introduction to Catapult HLS
- 0:56Nvidia design methodology
- 3:33Benefits of high-level design
- 6:34Case study of ML accelerators
- 9:44Summary and resources
related talks
31mCFU Playground
34mBuilding FPGA-based Machine Learning Accelerators in Python
10mLightning Talk: Software Development for ML and RISC-V Vector Acceler...
23mHugging Face Accelerate: Making Device-Agnostic ML Training and Inference Easy...
26mAmazingly Fast and Incredibly Scalable Inference...
53mBuilding an AI Accelerator with High-Level Synthesis
16mHardware acceleration for on-device Machine Learning
01h 11mMachine Learning Acceleration in an Embedded World