Real-World Pipelines: Comparing ARM and Intel, and Speeding Up Matrix Multiply

Why Real Processors Differ From the Textbook Pipeline

The five-stage pipeline covered earlier in this series is a simplified teaching model. Real commercial processors implement pipelining with design choices shaped by their intended use case, whether that priority is power efficiency, raw single-threaded speed, or a balance of both.

ARM Cortex-A53: Designed for Power Efficiency

The ARM Cortex-A53, commonly used in mobile devices, uses a relatively short, in-order pipeline with limited multiple-issue capability. This design choice favors low power consumption and reasonable performance per watt, since mobile devices are far more constrained by battery life and heat dissipation than by achieving the absolute highest possible clock-for-clock performance.

Intel Core i7: Designed for Raw Performance

The Intel Core i7, by contrast, uses a much deeper pipeline, aggressive out-of-order execution, and extensive dynamic branch prediction, all covered conceptually earlier in this series. These design choices prioritize maximum single-threaded performance, at the cost of significantly higher power consumption and more complex, larger silicon area dedicated to the processor core.

Design comparison:
ARM Cortex-A53: shorter pipeline, in-order execution,
                narrower issue width, lower power draw

Intel Core i7:  deeper pipeline, out-of-order execution,
                wider issue width, higher power draw

Neither approach is universally superior; each represents a deliberate tradeoff aligned with the device's intended purpose, directly reflecting the power wall discussed earlier in this series.

Applying These Ideas: Speeding Up Matrix Multiply

Matrix multiplication, discussed earlier in this series in the context of subword parallelism, benefits further from instruction-level parallelism. Because the individual multiply-and-add operations across different output elements are independent of each other, a superscalar processor with multiple ALUs can execute several of these operations in the very same clock cycle.

Sequential execution:
result[0] = a[0]*b[0] (cycle 1)
result[1] = a[1]*b[1] (cycle 2)

With instruction-level parallelism:
result[0] and result[1] computed
simultaneously in the same cycle,
using two separate ALUs

Combining this with the subword parallelism discussed earlier compounds the benefit further, since each of the parallel instructions issued in a single cycle can itself process multiple data values through SIMD-style operations.

Why Studying Real Chips Matters

Comparing these two real processors demonstrates that pipelining and instruction-level parallelism are not fixed formulas applied identically everywhere, but flexible tools that designers tune according to specific goals, and that a computation as common as matrix multiplication benefits from nearly every hardware technique covered throughout this chapter working together.