Course contentsShow
← Computer Science
Lesson 1179 of 2,87227. Pipelining and PerformanceFree lesson

Two-Bit Saturating Counter Predictor

Improved predictor using four states to require two mispredictions before changing direction; hysteresis benefits.

Two-Bit Saturating Counter Predictor

What You'll Learn: You'll discover how a two-bit predictor adds "confidence" to branch predictions, reducing mispredictions by requiring two mistakes before switching direction.

The Problem with One-Bit Predictors

Remember the one-bit branch predictor from the previous lesson? It flips its prediction after every single misprediction. This means if a branch slightly misbehaves even once, the predictor immediately changes its mind — which can be too reactive, especially for branches that are usually predictable but occasionally deviate.

Adding Hysteresis with Two Bits

A two-bit saturating counter uses four states instead of two, creating a form of hysteresis (resistance to change). Think of it like a thermostat that doesn't turn on/off with every tiny temperature fluctuation — it needs a bigger change to switch states.

The Four States

  1. Strongly Not Taken (00)
  2. Weakly Not Taken (01)
  3. Weakly Taken (10)
  4. Strongly Taken (11)

The predictor treats "strong" states (00, 11) as predictions and requires two consecutive mispredictions to flip from "taken" to "not taken" (or vice versa).

How It Works

  • When branch is taken: Counter increments (saturates at 11)
  • When branch is not taken: Counter decrements (saturates at 00)
  • Prediction: Based on the high-order bit (0 = predict not taken, 1 = predict taken)

Why This Helps

If a loop normally iterates 10 times, the one-bit predictor mispredicts twice (last iteration + first iteration of next loop). The two-bit predictor only mispredicts once — the final iteration doesn't immediately flip the prediction from "strongly taken" to "not taken"; it moves to "weakly taken," so the next loop entry still predicts correctly.

Key Takeaway: The two-bit saturating counter predictor adds confidence through four states, requiring two mispredictions before changing prediction direction, dramatically improving accuracy for loops and mostly-predictable branches.