Strength in Numbers: Unums and the Quest for Reliable Arithmetic
summary
Claim: unums (version 1) replace opaque floats with interval arithmetic that honestly returns ranges rather than pretending exact values. Demo uses the Julia UNAM library and bit-level inspection to show 0.1 + 0.2 producing a wider interval than 0.3, with unum encodings like unum 3/4 environment (sign exponent fraction and an u‑bit). Built examples work in software but show interval comparisons and hardware support remain open problems; neural‑net quantization examples illustrate where lower precision wins performance.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:16Numbers and Arithmetic
- 16:15Floating point is a broken number system.
- 27:53Part III: What does all this mean?
- 29:05We take measurements of the world.
- 35:51Declaring needs choosing options
- 36:51We need: 1. better number systems.
related talks
48mFuture-Proofing Numerical Accuracy with Formal Specifications
51mFloating-Point Arithmetic on Deterministic Systems
56mHow to Quickly Convert Floating-Point Numbers in C++
02mWhen Computers Can't Math - Floating Point Rounding Error Explained
23mModular Construction and Optimization of the UZP Sparse Format for SpMV on CPUs
49mModern Metrics
01h 04mCross-Platform Floating-Point Determinism Out of the Box
05mBinary in JavaScript