Phiacta
ExplorePostDocsGuidesContributingAbout

Phiacta

The knowledge backend.

Contact Us
ExploreQF8 Achieves +6.6 dB SQNR Over FP8 E4M3 Across All Tested Distributions

QF8 Achieves +6.6 dB SQNR Over FP8 E4M3 Across All Tested Distributions

QF8 (38.1 dB) vs FP8 E4M3 (31.5 dB) on N(0,1) — a 4.6× reduction in quantization noise power. The advantage is consistent (+6.3 to +6.8 dB) across Gaussian, log-normal, Laplace, and 90%-sparse distributions, and across matrix sizes from 16×32×16 to 128×256×128. This is structural: 16 vs. 8 levels per octave.

ContentIssuesEditsHistoryFilesReferences4

QF8 vs FP8 E4M3: +6.6 dB SQNR Advantage

QF8 (QuakeFloat8) is an 8-bit log-domain number format that provides 16 representable levels per octave (factor-of-2 range), compared to 8 levels per octave for IEEE FP8 E4M3 (the industry-standard 8-bit floating-point format used in NVIDIA H100, AMD MI300, etc.). Both formats use the same effective storage: 8.25 bits per element (8 per-element bits + shared block exponent).

This entry presents benchmark results showing that QF8's 2× finer per-octave resolution translates to a consistent SQNR (signal-to-quantization-noise ratio) advantage of +6.6 dB — equivalent to a 4.6× reduction in quantization noise power.

Key Result

FormatSQNR on N(0,1)Gap from QF8
QF8 (log-uniform, 256 levels)38.1 dB—
FP8 E4M3 (IEEE)31.5 dB6.6 dB worse

This advantage is predicted by the minimax NMSE theorem and confirmed empirically below.

Within-Binade Precision Comparison

Comparing at equal levels-per-octave (8), log-uniform still slightly wins due to truly constant relative cell width vs. FP8's piecewise-constant structure:

MetricIEEE E4M3Log-uniform (8/oct)Ratio
Average MSRE6.510×10⁻⁴6.256×10⁻⁴1.041
Peak relative error0.06250.044271.41
Worst-case MSRE (HR)1.302×10⁻³6.256×10⁻⁴2.08

But QF8's main advantage is structural: 16 levels/octave vs. 8.

Matrix Multiplication SQNR

Measured vs. float64 ground truth on random matrices:

Sizebfloat16float32FP8 E4M3 (block)QF8
16×32×1652.3 dB139.0 dB28.6 dB35.3 dB
64×128×6452.5 dB133.6 dB28.4 dB35.1 dB
128×256×12852.6 dB130.8 dB28.5 dB35.1 dB

The advantage is consistent across matrix sizes, confirming the dimension-free property.

SQNR Across Distributions

DistributionFP8 E4M3QF8Advantage
N(0, 0.02²) — typical weights31.4 dB38.2 dB+6.7 dB
N(0, 1) — post-LayerNorm activations31.6 dB37.9 dB+6.3 dB
LogN(0, 1) — gradient-like31.5 dB38.3 dB+6.8 dB
Laplace(0, 0.02) — sparse weights31.5 dB38.0 dB+6.5 dB
90% sparse + N(0,1)31.7 dB38.3 dB+6.6 dB

The +6.3 to +6.8 dB advantage is stable across all tested distributions. This is expected from the minimax theorem: log-uniform quantization achieves the same NMSE regardless of input distribution, so the advantage over FP8 is structural rather than distribution-specific.

Metadata

Type
result
Visibility
public
Published
Mar 27, 2026
Last updated
Mar 27, 2026

Tags

benchmarkFP8quakefloat8resultSQNR