About
Special-purpose hardware accelerators are critical for performance gains amid slowing technology scaling, but designers lack effective tools to build complex accelerators. Existing high-level synthesis (HLS) tools require intrusive source-level changes to attain high performance while most accelerator design languages excel only with simple kernels. In this talk, we present Allo [PLDI'24], a Python-embedded programming model for composable accelerator design. Allo decouples hardware customizations (compute, memory, communication, data types) from algorithms, and encapsulates them as primitives. By preserving program hierarchy, Allo integrates customizations bottom-up, enabling holistic optimizations across functions. Evaluated on HLS benchmarks and real-world applications, Allo outperforms state-of-the-art tools. Allo-generated accelerators for machine learning models achieve superior latency and energy efficiency compared to GPUs, showing scalability for complex, large-scale designs..
0:00:00 Slide 1
0:01:46 Slide 2
0:02:55 Slide 3
0:04:22 Slide 4
0:06:04 Slide 5
0:07:38 Slide 6
0:12:13 Slide 7
0:13:20 Slide 8
0:20:06 Slide 9
0:21:56 Slide 10
The talk
Credits and licence
- Institution
- CERN
- Department
- FPGA Developers' Forum (FDF)
- Speaker
- Hongzheng Chen
- Year
- 2025
- Language
- English
- Subjects
- Computer science, Art, architecture and design
- Source
- CERN
- Licence
- CERN, as CERN states it. CERN's terms
- Attribution
- © CERN. CERN provides its audiovisual media free of charge for educational and informational use, credited to CERN; this lecture is shown in CERN's own player.
- Who can watch
- CERN gives no rating. Invisible Network shows it to Tweens profiles and older, judged from what the source says about it. How courses are placed