The Noise Floor of a Quantization Comparison (Cloudflare, 2611.15671)
TL;DR. When you compare a quantized model to its BF16 original, you usually change more than the weights: the serving engine, the server instance, the batch size and sometimes the kernel the library picks. This pre-registered study measures each of those non-quantization sources on 2,400 public decisions from two Cloudflare decision models (Clef-flash 9B and Clef 27B) on an NVIDIA GB10, then reports quantization effects only as the excess over that noise. Non-quantization sources flip 0 to 0.92% of decisions. 8-bit flips about 0.9-2.1%, 4-bit about 3.5-5.8%. Every 4-bit effect clears the noise floor; FP8 does not sit inside restart noise, which the authors had predicted it would. Accuracy hides most of the movement: Clef's Q4_K_M build moves net accuracy by six questions out of 31,275 while 436 individual decisions get worse.
flowchart LR
B["BF16 baseline<br/><small>one engine, one instance</small>"] --> N{"Noise sources<br/><small>restart, engine, batch, kernel</small>"}
N -->|0 to 0.92%| F["Noise floor<br/><small>decisions flipped</small>"]
B --> Q["Quantized build<br/><small>Q8, FP8, Q4_K_M, NVFP4</small>"]
Q -->|0.9 to 5.8%| E["Excess effect<br/><small>quant minus floor</small>"]
F --> E
E --> A["Accuracy view<br/><small>net change hides flips</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
class B input
class N loop
class F err
class Q core
class E exit
class A err
linkStyle 3 stroke:#2f9e44,stroke-width:2px
Key findings
- llama.cpp restarts are bit-identical on every record for both models. vLLM restarts are not: they flip 0, 5 and 20 decisions at BF16, FP8 and NVFP4 on Clef-flash, and 4, 0 and 22 on Clef.
- An engine change at BF16 flips 6 to 16 decisions, a batch-size change 8 to 15, and the Gated DeltaNet kernel that transformers selects on its own (depending on which packages are imported) flips 10 to 14.
- Quantization inside one engine moves more: 0.88% / 1.04% at Q8_0, 2.12% / 1.62% at FP8, 5.79% / 3.50% at Q4_K_M, 5.79% / 5.67% at NVFP4 (Clef-flash / Clef).
- Two of five pre-registered predictions failed, and failed again on the 27B: FP8 is not inside restart noise, and an engine change moves fewer decisions than 8-bit quantization.
- Net accuracy is the wrong metric for decision models. Q4_K_M: -0.02 points [-0.20, +0.16] net, yet 436 decisions worse. NVFP4 costs 0.95 and 0.65 points.
How it relates to prior wiki pages
- Extends the 11-15 rounding page, which showed that the operators under low-bit recipes are not exact. This paper puts a number on how much that inexactness moves real decisions, and shows it is smaller than 4-bit quantization but comparable to 8-bit.
- Confirms the SABLE bit-exact inference page (11-12) argument that serving nondeterminism is a measurement problem, not just a reproducibility nuisance.
- Complicates the 11-18 BitDecoding-FP4 and Qwen3.8 NVFP4-QAD claims of "within noise of BF16": those comparisons report accuracy, which this paper shows can hide hundreds of flipped decisions.
- Concept page: quantization.
Gaps
Two decision models, one GPU (GB10), classification-style outputs. No long-generation or agentic tasks, where flips compound across steps. The third-instance result for Clef's quantized builds is exploratory.
Raw source: HF listing · arXiv