Great question! Cycle stealing is a clever trick DMA uses to move data without hogging the system bus full-time, but it does come with a small unavoidable cost to CPU performance. Let’s break down the key reasons in plain terms:

• Bus Contention & Forced Idle Time
  DMA controllers "steal" bus cycles when the CPU isn’t using the system bus (like when it’s executing internal register-only instructions that don’t need memory access). But here’s the catch: if the CPU does need the bus right when DMA wants to steal a cycle, DMA usually has higher priority (critical peripheral data—like from a disk or network card—can’t be lost). The CPU has to pause its current operation cold and wait for DMA to finish the transfer. Those idle waiting cycles directly cut into the CPU’s usable execution time, making it run slower than if there were no DMA activity at all.

• Cumulative Impact of Frequent Stealing
  For high-bandwidth tasks (think copying a large video file or streaming data from a fast Ethernet card), DMA will steal cycles very frequently. Each individual wait might only be a handful of clock cycles, but the cumulative effect adds up quickly. Over time, the CPU spends more and more time waiting for the bus instead of crunching through instructions, leading to a measurable drop in overall performance that you might even notice in real-time.

• Pipeline Stalls in Modern CPUs
  Modern CPUs rely on deep instruction pipelines to maximize efficiency—they’re constantly fetching, decoding, and executing instructions in parallel. When a bus wait forces the CPU to pause, the entire pipeline stalls. Restarting the pipeline takes extra cycles to refill with new instructions, adding even more overhead beyond just the direct wait time. This stall effect amplifies the slowdown compared to older, simpler CPUs that don’t rely on pipelining.

It’s worth emphasizing that cycle stealing is still way more efficient than the alternative (having the CPU handle every data transfer via interrupts, which would tie up the CPU far more). But it’s not entirely free—hence the slight but noticeable hit to CPU speed.

内容的提问来源于stack exchange，提问作者박상혁