EDEN, a 2021 vector quantization method, consistently outperforms TurboQuant, its 2026 successor. The algorithms share similar approaches using random rotation and scalar quantization, but EDEN's analytical scale optimization gives it a critical edge. This matters for compressing neural network embeddings and KV caches, where efficient quantization directly impacts model performance and memory usage.

TurboQuant essentially functions as EDEN with fixed scaling (S=1), while EDEN derives optimal scale factors analytically. In tests across dimensions 16-4096 and bit-widths 1-4, EDEN-biased reduces MSE by up to 2.25% over TurboQuant-mse. The gap persists at practical dimensions used in real applications, demonstrating that proper scaling isn't just theoretically important but practically significant.

For unbiased compression needed in distributed training and attention mechanisms, EDEN-unbiased substantially outperforms TurboQuant-prod. The single-pass design with optimized scaling allows EDEN with b bits to match or exceed TurboQuant's performance with b+1 bits, effectively saving more than a full bit per coordinate while maintaining accuracy.