Course contentsShow
← Machine Learning and Deep Learning
Lesson 2634 of 3,53857. Model Compression: QuantizationFree lesson

Symmetric vs Asymmetric Quantization

Learn symmetric quantization (zero point at zero) versus asymmetric (arbitrary zero point) and when to use each for PTQ.

Symmetric vs Asymmetric Quantization

What You'll Learn

You'll understand the difference between symmetric and asymmetric quantization schemes and when each is appropriate for post-training quantization.

The Two Quantization Approaches

When mapping floating-point values to integers, you need to decide where zero lands in your quantized range—this choice defines symmetric versus asymmetric quantization.

Symmetric Quantization

In symmetric quantization, the floating-point zero maps exactly to integer zero. The quantization range is balanced around zero, like a number line centered at the origin.

Analogy: Think of a thermometer centered at 0°C—equal ranges above and below freezing.

  • Zero-point parameter = 0
  • Scale factor handles the mapping
  • Range: [-127, 127] for int8
  • Pros: Simpler math, faster hardware operations, no zero-point addition needed
  • Cons: Wastes representation space if data isn't centered around zero

Asymmetric Quantization

In asymmetric quantization, zero can map to any integer value. This allows the quantized range to fit the actual data distribution more tightly.

Analogy: Like a thermometer that starts at 10°C and goes to 40°C—optimized for your climate, not wasting scale on impossible temperatures.

  • Zero-point parameter can be any value
  • Uses full integer range: [-128, 127] for int8
  • Pros: Better utilizes quantization range, especially for activations (often ReLU outputs ≥ 0)
  • Cons: Extra computation for zero-point offset

When to Use Each for PTQ

  • Symmetric: Use for weights (typically centered distributions) and when hardware acceleration matters
  • Asymmetric: Use for activations with skewed distributions (like post-ReLU values) when accuracy is critical

Key Takeaway: Symmetric quantization (zero-point = 0) simplifies computation but may waste range; asymmetric quantization (arbitrary zero-point) maximizes range utilization at the cost of slightly more complex operations—choose based on your data distribution and hardware constraints.