a hardware accelerator
Suppose you do your taxes by hand every year. You could keep doing the arithmetic yourself, or you could buy a dedicated calculator that does exactly the columns of figures you need, instantly. The calculator does not replace you — you still organise the work — but you hand off the heavy number-crunching to it and get the result back. A hardware accelerator is that dedicated calculator for a computer: a separate block of silicon that the main CPU offloads a specific, heavy task to, because the accelerator does that one task much faster and more cheaply in energy than the CPU could.
Concretely, an accelerator sits alongside the CPU and is driven by it. The CPU prepares the data and the parameters in memory, tells the accelerator 'go' (often by writing to a control register), and then either waits or goes off to do other work while the accelerator runs. When the accelerator finishes it raises an interrupt or sets a status flag, and the CPU collects the results. The accelerator may be a GPU, a neural-network unit, a video encoder, an encryption engine, or a network/storage offload chip. What makes it an accelerator rather than just 'another processor' is that it is specialized: its datapath is shaped around one task, so it has many more of the right arithmetic units and far less general-purpose overhead.
The honest point is that an accelerator only helps if the offloaded work is large enough to overcome the cost of getting data to it and back. Moving inputs to the accelerator's memory, starting it, and copying results out all take time and energy. If the task is tiny, that setup overhead dwarfs the speedup and the CPU would have been faster doing it itself. Accelerators shine on big, repetitive, parallel jobs; they are a poor fit for small, one-off, branchy work.
Encrypting a 4-gigabyte disk image: a CPU loop might take many seconds, but a dedicated AES crypto engine in the chip can stream through it far faster while the CPU is free to do other things. Encrypting one short password, by contrast, is so small that calling the engine costs more than just letting the CPU do it.
Offloading pays off only when the task is big enough to amortize the cost of handing it over.
An accelerator does not make a computer faster in general — it speeds up only the specific workload it was built for, and often the real bottleneck becomes feeding it data, not the computation itself.