Fixed Point Numeric Types for Hardware Description
summary
Claims a compact numeric model and a working prototype: a Julia library that implements bounded and modular integers and a fixed‑point type, uses an exponent for significance, and enforces lossless arithmetic so addition and splitting preserve numeric value and avoid unnecessary bit growth. Implementation exists with unit tests but is work‑in‑progress (needs more coverage and arbitrary‑length support). Plans include an HDL backend to emit hardware types with tighter, provable ranges rather than automatic width bloat.
AI-written summary · about
chapters (12)
- 0:00Fixed point numbers intro
- 1:04Fixed point representation
- 3:11Binary number systems
- 4:52Modulo-n integer limits
- 6:37Bounded integers for precision
- 7:20Unified fixed point design
- 9:05Special cases in fixed point
- 10:29Bit and part selection math
- 11:55Fixed point verification
- 12:58Julia implementation overview
- 13:22Summary for hardware design
- 14:25Q&A
related talks
14mEfficient Implementation of Neural Networks Usual Layers on Fixed-Point Architectures
12mMixPert: Optimizing Mixed-Precision Floating-Point Emulation on GPU Integer Tensor Cores
30mVector ISA Proposal Update
29mProgramming machine learning algorithms in hardware, sanely, using Haskell and Rust!
14mOcto: INT8 Training with Loss-aware Compensation and Backward Quantization for Tiny
38mBASE ISA
19mModular Hardware Design of Pipelined Circuits with Hazards
23mRunning an Open Source LLM on Open Source Hardware