Photo by Thorium on Unsplash

In this article we are going to discuss Python Multiprocessing vs Threading.

You run the following code and watched five seconds of work finish in one:

import time
from concurrent.futures import ThreadPoolExecutor

def slow_task(n):
    time.sleep(1)          # pretend this is a download
    return n * 2

with ThreadPoolExecutor(max_workers=5) as pool:
    results = list(pool.map(slow_task, range(5)))   # ~1 second, not 5

Threads look like magic. So the natural next thought is: "My number-crunching script is slow, I'll just throw threads at it too." Then you do, and nothing gets faster. This guide explains exactly why that happens, and shows you (with runnable code and real timings) which tool actually speeds up CPU-heavy work in Python.

Key Takeaways - Threads in standard CPython can't run Python computations in parallel because of the Global Interpreter Lock (GIL) — only one thread executes Python bytecode at a time (Python docs) - For CPU-bound work, use multiprocessing / ProcessPoolExecutor: each process gets its own interpreter and its own GIL, so they run truly in parallel across cores. - For I/O-bound work (downloads, disk, database calls), threads are perfect, the GIL is released while a thread waits. - In my own 8-core benchmark, threads gave a 1.0× speedup on CPU work (nothing) but ~8× on I/O work; processes gave ~4.6× on the same CPU work. - Python 3.14 (Oct 2025) makes a free-threaded "no-GIL" build officially supported — but it's opt-in, not the default, so multiprocessing is still the beginner-safe answer today.

The One Question That Decides Everything: Is Your Work CPU-Bound or I/O-Bound?

Before you pick a tool, classify your work. This single distinction decides everything that follows.

• I/O-bound work spends most of its time waiting. Example for a network response, a file to load, a database query. Your CPU is idle during the wait. The time.sleep(1) in the snippet above is a stand-in for exactly this kind of waiting.
• CPU-bound work spends most of its time computing like parsing, hashing, resizing images, math in a loop. The CPU is pinned at 100% the whole time. Nothing is waiting; there's simply a lot to calculate.

Quick self-test: If your slow code is waiting on something outside Python (a server, a disk, an API), it's I/O-bound. If it's a busy loop doing arithmetic or data processing with no sleep and no network, it's CPU-bound.

Why does this matter? Because threading only helps when there's waiting to overlap. When the work is pure computation, there's no idle time to fill — and, as you'll see next, a lock inside Python stops threads from computing simultaneously anyway.

Why Threads Don't Speed Up CPU-Bound Work (The GIL, Explained Simply)

Here's the rule that trips up every beginner: standard CPython lets only one thread run Python code at a time. The mechanism enforcing that is the Global Interpreter Lock, or GIL.

Think of the GIL as a single "talking stick." No matter how many threads you start, a thread can only execute Python bytecode while it holds the stick, and there's exactly one stick. Threads take turns; they don't run at the same time. The GIL exists because CPython's memory management (reference counting) isn't thread-safe, and a single lock is the simplest way to keep it correct (Real Python: What Is the Python GIL?, retrieved 2026-09-15).

For I/O-bound work this is fine — a thread releases the GIL while it waits for a download or disk read, so other threads run during the wait. That's why the opening snippet works. But for CPU-bound work there is no waiting: every thread wants the stick constantly, they hand it back and forth, and your program runs at roughly single-threaded speed — sometimes even a little slower, because of the overhead of passing the stick around.

Don't take my word for it. Here's a CPU-bound benchmark you can paste and run. The code is summing squares, no sleeping, no I/O:

import time
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor

def cpu_task(n):
    total = 0
    for i in range(n):      # pure computation — the CPU never waits
        total += i * i
    return total

NUMBERS = [8_000_000] * 8   # 8 heavy chunks

def timed(label, fn):
    start = time.perf_counter()
    fn()
    print(f"{label:<34} {time.perf_counter() - start:.2f}s")

if __name__ == "__main__":
    timed("Serial (one at a time)", lambda: [cpu_task(n) for n in NUMBERS])
    timed("ThreadPoolExecutor (8 threads)",
          lambda: list(ThreadPoolExecutor(8).map(cpu_task, NUMBERS)))
    timed("ProcessPoolExecutor (8 processes)",
          lambda: list(ProcessPoolExecutor(8).map(cpu_task, NUMBERS)))

On my 8-core machine (Python 3.11), this prints almost the same result every run:

My finding: Eight threads made the CPU-bound task exactly as slow as running it one at a time. Eight processes cut it to roughly a fifth. The GIL is the whole reason for the gap.

How Multiprocessing Gets Around the GIL

If the problem is "one GIL per interpreter," the fix is simple: use more interpreters. That's what multiprocessing does. Instead of many threads sharing one interpreter (and one GIL), it starts several separate Python processes — each with its own interpreter, its own memory, and its own GIL. The operating system schedules them onto different CPU cores, and they genuinely run at the same time (Python concurrent.futures docs, retrieved 2026-09-15).

That real parallelism isn't free, though. Separate processes can't share Python objects directly, so Python has to pickle (serialize) every argument you send into a worker and every result it sends back, then ship the bytes across a pipe. Starting a process also costs more than starting a thread, and each process duplicates interpreter memory (Python⇒Speed: faster multiprocessing, retrieved 2026-09-15).

The practical consequences for a beginner:

• Multiprocessing wins when the computation per task is large relative to the cost of shipping data in and out. Summing eight million squares? Easily worth it.
• It can lose on tiny tasks, where pickling and process startup dominate — sometimes making it slower than a plain loop.
• Keep the data crossing the boundary small. Pass a filename instead of a giant list; return a number instead of a huge object.

Side-by-Side Code: The Same Task, Three Ways

The best part of concurrent.futures is that switching from threads to processes is almost a one-word change. Same map, same result handling — you swap the executor class:

from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor

# I/O-bound? Use threads.
with ThreadPoolExecutor(max_workers=8) as pool:
    results = list(pool.map(download, urls))

# CPU-bound? Change ONE word: Thread -> Process.
with ProcessPoolExecutor(max_workers=8) as pool:
    results = list(pool.map(cpu_task, numbers))

from concurrent.futures import ProcessPoolExecutor, as_completed

with ProcessPoolExecutor(max_workers=8) as pool:
    futures = [pool.submit(cpu_task, n) for n in numbers]
    for future in as_completed(futures):
        print(future.result())   # prints in completion order, not submit order