prajwalpujari16/quantbench-pareto

⚡ QuantBench: LLM Quantization Quality vs. Compression Pareto Benchmark

The open-source empirical benchmark quantifying the Pareto frontier between memory compression, Signal-to-Quantization-Noise Ratio (SQNR), relative reconstruction error, and evaluation perplexity across state-of-the-art LLM quantization algorithms.

• GitHub Repository: Prajwal7-2002/quantbench
• PyPI Engine: pip install quantbench-engine
• Documentation: prajwal7-2002.github.io/quantbench
• Author: Prajwal Pujari (@Prajwal7-2002)

---

📊 Quick Start

Load the benchmark tables directly in Python with one line:

from datasets import load_dataset

# Load Pareto frontier and layer sensitivity tables
dataset = load_dataset("prajwalpujari16/quantbench-pareto")

# Inspect the Pareto trade-off curve
df_pareto = dataset["pareto_frontier"].to_pandas()
print(df_pareto[["configuration", "compression_ratio", "sqnr_db", "perplexity_delta"]])

---

🔬 Master Pareto Frontier Table

Configuration	Scheme	Bit-Width	Group Size	Compression	Memory Savings	SQNR (dB)	Delta PPL	Primary Acceleration Engine
FP16 Baseline	fp16	16.0	None	1.00×	0.0%	100.0 dB	0.000	Native Tensor Cores
INT8 SmoothQuant (α=0.5)	smoothquant	8.0	None	2.00×	50.0%	45.3 dB	+0.004	W8A8 Matrix Multiplication
INT8 AbsMax (Per-Channel)	int8	8.0	None	2.00×	50.0%	42.1 dB	+0.009	INT8 Tensor Cores (DP4A / IMMA)
FP8 E4M3FN (Simulated)	fp8	8.0	None	2.00×	50.0%	38.7 dB	+0.018	Hopper / Blackwell FP8 Cores
FP8 E5M2 (Simulated)	fp8	8.0	None	2.00×	50.0%	31.4 dB	+0.042	Hopper FP8 GEMM
INT4 GPTQ (Hessian OBS)	gptq	4.5	128	3.55×	71.8%	23.8 dB	+0.029	ExLlamaV2 / Marlin Fused Kernels
INT4 AWQ (Activation-Aware)	awq	4.5	128	3.55×	71.8%	22.4 dB	+0.035	W4A16 TensorRT-LLM / vLLM Marlin
INT4 Blockwise (G=32)	int4	6.0	32	2.67×	62.5%	20.5 dB	+0.048	Fine-grained block scaling
INT4 Blockwise (G=64)	int4	5.0	64	3.20×	68.8%	19.4 dB	+0.061	Ollama / llama.cpp Q4_K_M
INT4 Blockwise (G=128)	int4	4.5	128	3.55×	71.8%	18.6 dB	+0.074	Standard W4A16 Dequantizer
BitNet 1.58b Ternary	ternary	2.13	None	7.50×	86.7%	8.2 dB	+0.320	Addition-only integer arithmetic
Ampere 2:4 Structured Sparse	sparse24	10.0	None	1.60×	37.5%	7.1 dB	+0.410	Ampere 2:4 Sparse Tensor Cores

---

🎯 Key Empirical Insights

1. The Attention vs. MLP Degradation Asymmetry:

   • Attention Query ($Q$) and Key ($K$) projections suffer 1.8 to 2.4 dB higher SQNR degradation than Feed-Forward (MLP) layers under identical quantization schemes.
   • Compressing Query/Key projections directly disturbs the attention entropy distribution ($\text{softmax}(QK^T / \sqrt{d})$), causing outsized perplexity spikes. This empirical finding provides direct justification for Activation-Aware Weight Quantization (AWQ) preserving salient channels.

2. The 3.5× "Sweet Spot" at INT4:

   • Blockwise INT4 with group size $G=128$ yields $3.55\times$ weight compression ($71.8%$ memory savings) while maintaining $>18.6\text{ dB}$ SQNR and minimal perplexity impact ($\Delta\text{PPL} \le +0.074$).
   • Applying second-order error compensation (GPTQ) or salient channel scaling (AWQ) recovers another $3.8 - 5.2\text{ dB}$ of signal fidelity.

3. BitNet 1.58b Memory Ceiling:

   • Packing ternary weights ${-1, 0, +1}$ using base-3 arithmetic ($3^4 = 81 < 256$, packing 4 weights per byte) achieves an unprecedented $7.50\times$ compression ratio ($86.7%$ VRAM reduction).

---

🧮 Mathematical Definitions

1. Signal-to-Quantization-Noise Ratio (SQNR)

Measures the ratio of original tensor energy to the noise energy introduced by discrete quantization: $$\text{SQNR (dB)}=10\cdot {\log }_{10}\left(\frac{\mathrm{\parallel }W{\mathrm{\parallel }}_2^2}{\mathrm{\parallel }W-\widehat{W}{\mathrm{\parallel }}_2^2}\right)\backslash text\{SQNR\; (dB)\}\; =\; 10\; \backslash cdot\; \backslash log\_\{10\}\; \backslash left(\; \backslash frac\{\backslash |W\backslash |\_2^2\}\{\backslash |W\; -\; \backslash hat\{W\}\backslash |\_2^2\}\; \backslash right)$$

• $\text{SQNR} \ge 40\text{ dB}$: Perceptually lossless (INT8 / SmoothQuant).
• $20\text{ dB} \le \text{SQNR} < 40\text{ dB}$: High fidelity; minor loss (INT4 AWQ / GPTQ).
• $\text{SQNR} < 15\text{ dB}$: Noticeable degradation requiring domain calibration.

2. Relative Frobenius Reconstruction Error

$${ϵ}_F=\frac{\mathrm{\parallel }W-\widehat{W}{\mathrm{\parallel }}_F}{\mathrm{\parallel }W{\mathrm{\parallel }}_F}\backslash epsilon\_F\; =\; \backslash frac\{\backslash |W\; -\; \backslash hat\{W\}\backslash |\_F\}\{\backslash |W\backslash |\_F\}$$

---

🛠️ Reproducing Locally

To reproduce this benchmark or inspect sub-byte packing layouts directly on your machine:

# Install the zero-dependency quantbench engine
pip install quantbench-engine

# Run the complete Pareto sweep
quantbench pareto

# Inspect sub-byte ternary packing (4 weights packed per uint8 byte)
quantbench pack-inspect --format ternary

---

📜 Citation

@misc{pujari2026quantbench,
  author = {Prajwal Pujari},
  title = {QuantBench: LLM Quantization Quality vs Compression Pareto Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Data repository},
  howpublished = {\\url{https://huggingface.co/datasets/prajwalpujari16/quantbench-pareto}}
}