---
id: "framework-turboquant-process"
type: "framework"
source_timestamps: ["04:05:00", "06:10:00"]
tags: ["algorithm", "compression", "steps"]
related: ["concept-turboquant", "concept-polar-quantization", "concept-qjl", "concept-vector-quantization"]
steps_count: 2
sources: ["s49-killed-ram-limits"]
sourceVaultSlug: "s49-killed-ram-limits"
originDay: 49
---
# Turboquant Compression Process

The [[concept-turboquant]] algorithm achieves lossless compression of the [[concept-kv-cache]] through a specific **two-step mathematical process**, avoiding the overhead of traditional [[concept-vector-quantization]].

## Step 1 — Polar Quantization

Rotate the data into a polar coordinate system (radius and angle) to make the data structure highly predictable and **eliminate the need for per-block normalization instructions**. This solves the 'extra bag of folding instructions' problem inherent to vector quantization.

- **Radius** = signal strength of the vector
- **Angle** = directional / semantic component

Details: [[concept-polar-quantization]].

## Step 2 — Quantized Johnson-Lindenstrauss (QJL)

Apply a **data-oblivious mathematical error-checker** using a single bit to correct the tiny residual rounding errors introduced during the polar rotation, ensuring **perfect losslessness**.

Details: [[concept-qjl]] and [[concept-data-oblivious-algorithm]].

## End Result

- KV cache representation reduced from 32 bits → **3 bits per token** (or 2.5 bits with outlier channel allocation)
- **6x memory reduction**, **8x speedup**, **zero accuracy loss** — see [[claim-turboquant-performance]]
- Works universally across model architectures because both steps are data-oblivious