Why is division considered a more expensive operation for computers?

Division, especially for integers, takes significantly more clock cycles on a computer compared to addition or multiplication. On the machine discussed, a 32-bit integer division takes 30 cycles, while addition takes only one. This is because division circuits are more complex and often there's only one dedicated unit on the chip.

What are floating-point numbers and why are they important?

Floating-point numbers are approximations of real numbers used for calculations involving fractions or decimals. Modern computers are highly optimized for floating-point operations because they are essential for tasks like AI, gaming, and scientific simulations.

What is the penalty for a branch misprediction in a CPU?

When a CPU's branch predictor makes an incorrect guess about the program's flow, it incurs a penalty. This involves discarding work done based on the wrong guess and restarting, which can cost anywhere from 10 to 40 clock cycles. In human terms, this is analogous to being interrupted and needing time to regain focus.

Why do computers use multiple layers of cache?

Computers use multiple cache layers (L1, L2, L3) to bridge the speed gap between the fast CPU and slower RAM. These caches store frequently accessed data closer to the CPU, with L1 being the fastest and smallest, and L3 being the slowest and largest shared cache.

How does a computer's access speed to storage (SSD vs. spinning disk) compare to human timescales?

Accessing data from a SATA SSD takes about 20 microseconds, translating to 2.3 human days. A traditional spinning disk is much slower, taking approximately 10 milliseconds per sector, which equates to a staggering 3.2 years in human time for a single block read.

What is the latency involved in accessing the internet?

Even basic internet operations have significant latency. Pinging a local router takes about 400 microseconds (6 weeks in human time). Pinging Google takes 10 milliseconds (16 weeks), and pinging a distant server like one in Nottingham can take 31 milliseconds, equating to 31 years in human comparative time.

Computer Timescales Mapped onto Human Timescales - Computerphile

Computerphile

Education4 min read29 min video

Jan 7, 2025|207,749 views|7,423|516

computers computerphile computer science

Save to Pod

Key Moments

TL;DR

Computers are astronomically faster than humans, but some operations reveal surprising slowness relative to humans.

Key Insights

Computers perform basic arithmetic operations like addition, subtraction, multiplication, and division at speeds incomprehensible to humans.

While integer operations are fast, floating-point operations, crucial for real-world applications and AI, are also highly optimized and relatively quick.

Branch prediction penalties, where the computer guesses the wrong execution path, are significant but analogous to human distraction and recovery time.

Accessing memory (RAM) is a major bottleneck, taking significantly longer than on-chip operations, akin to a human taking a trip to a distant store.

Interacting with external storage (SSDs and HDDs) is orders of magnitude slower than RAM access, resembling a multi-day or multi-year task for humans.

Network latency, though seemingly small to humans, translates to weeks or even years for round trips in computer time.

Despite immense speed, the effective task time for complex operations like rendering a game frame is still substantial when mapped to human timescales.

THE MIND-BOGGLING SPEED OF BASIC ARITHMETIC

The video begins by establishing a baseline for human arithmetic speed, estimating about five seconds for a human to add two four-digit numbers. This is contrasted with a computer's ability to perform a similar operation (adding two 32 or 64-bit numbers) in a single clock cycle, which, at 2 GHz, equates to half a nanosecond. This fundamental difference highlights the vast disparity in processing power. Multiplication and division, though more complex, are also executed in mere clock cycles, with division being a more expensive operation for processors, taking around 30 cycles for 32-bit numbers.

FLOATING-POINT MATH AND HIDDEN EXPENSES

When dealing with real-world numbers that can be fractional, computers use floating-point arithmetic. While this involves approximations, modern hardware is highly optimized for these operations, especially for graphics (GPUs) and AI. Adding two floating-point numbers takes about four clock cycles, comparable to integer multiplication. Division of floating-point numbers is surprisingly faster than integer division on this machine, taking about 15 cycles. However, division, in general, is a costly operation, often hidden within commands like modulo, which can impact performance in applications like hash maps.

THE COST OF MISPREDICTION AND MEMORY ACCESS

Branch prediction is a critical technique for efficient pipeline processing in CPUs. A misprediction, where the CPU follows the wrong execution path, incurs a penalty ranging from 10 to 40 cycles. Analogously, this is like a human being interrupted and needing time to regain focus, equating to about 1 minute 40 seconds in human time. Memory access, however, presents a more significant slowdown. Accessing data in the L1 cache takes 4 cycles (20 seconds), L2 cache takes 12 cycles (1 minute), and L3 cache takes 30-50 cycles (3-4 minutes). This illustrates the escalating cost of retrieving data as it moves further from the CPU.

THE VAST GULF OF RAM AND STORAGE ACCESS

When data isn't in the caches, the computer must access RAM, which takes around 100 nanoseconds, translating to a staggering 15 minutes in human time. This highlights how crucial caching is for perceived speed. Further down the hierarchy, accessing data from a SATA SSD takes approximately 20 microseconds, which translates to 2.3 days in human equivalent time. Traditional spinning hard drives (HDDs) are even slower, taking milliseconds per access, amounting to about 3.2 years to read a single arbitrary sector.

NETWORK LATENCY AND INTER-PROCESS COMMUNICATION

Network interactions also reveal dramatic timescale differences. Pinging a router in the next room takes 400 microseconds (6 weeks). Pinging Google servers takes about 10 milliseconds (16 weeks), and pinging a server across the country can take around 31 years. Even more relatable, rendering a single frame in a video game, targeting 60 frames per second, gives the computer roughly 16 milliseconds, which is equivalent to five years of human computation time for that single frame. This vast disparity underscores the engineering marvels that make modern computing feel instantaneous.

THE ILLUSION OF MULTITASKING AND GRAND FINALE

Modern operating systems create the illusion of multitasking by rapidly switching between processes, often allocating a fixed time slice (e.g., 16 milliseconds) to each. This rapid switching, though too fast for humans to perceive, represents significant compute time per application. The video concludes by emphasizing that despite the incredible speeds of individual operations, the aggregate time taken for complex, real-world tasks when mapped onto human timescales is astronomical, making the technology we use daily a testament to clever engineering that hides these vast temporal differences.

Mentioned in This Episode

●Products

●Tools

●Companies

●Organizations

●Concepts

Comparison of Computational Task Times: Human vs. Computer

Data extracted from this episode

Task	Human Time (approx.)	Computer Cycles (approx.)	Computer Time (approx. Sha Units)
Add two 4-digit numbers	10 seconds	1 cycle	0.5 nanoseconds
Multiply two numbers	20 seconds	4 cycles	2 nanoseconds
Divide two 32-bit integers	2.5 minutes	30 cycles	15 nanoseconds
Divide two 64-bit integers	8 minutes	100 cycles	50 nanoseconds
Add two floating-point numbers	20 seconds	4 cycles	2 nanoseconds
Multiply two floating-point numbers	20 seconds	4 cycles	2 nanoseconds
Divide two floating-point numbers	1 minute	15 cycles	7.5 nanoseconds
Branch Misprediction Penalty	1 minute 40 seconds	10-40 cycles	5-20 nanoseconds
Access L1 Cache	20 seconds	4 cycles	2 nanoseconds
Access L2 Cache	1 minute	12 cycles	6 nanoseconds
Access L3 Cache	3-4 minutes	30-50 cycles	15-25 nanoseconds
Access RAM	15 minutes	Not measured in cycles	100 nanoseconds
Read from SATA SSD (512 bytes)	2.3 days	Not measured in cycles	20 microseconds
Read from Spinning Disk (512 bytes)	3.2 years	Not measured in cycles	8-10 milliseconds
Prepare one frame (60 FPS game)	5 years	Not measured in cycles	16 milliseconds
Ping local router	6 weeks	Not measured in cycles	400 microseconds
Ping Google	16 weeks	Not measured in cycles	10 milliseconds
Ping University of Nottingham	31 years	Not measured in cycles	31 milliseconds

Common Questions

For a simple task like adding two four-digit numbers, a human might take around 10 seconds, while a modern computer can do it in a single clock cycle, which is roughly half a nanosecond. This massive difference highlights the computational power computers possess for basic arithmetic.

Topics

Computer Speed Human Timescales Arithmetic Operations Floating Point Cache Hierarchy Memory Access SSD Vs HDD Network Latency Branch Prediction Clock Cycles Nanoseconds Milliseconds Analogy

Mentioned in this video

toolL2 cache

The second layer of cache, larger and slightly slower than L1 cache.

concepthash maps

An algorithm that uses division and remainders to fit values into a set of boxes, often hiding expensive operations.

conceptbranch prediction

A technique used in computer pipelines to guess the direction of a branch to avoid performance penalties.

organizationUniversity of Nottingham

A university whose web server was pinged to measure long-distance network latency.

conceptfloating point

A type of number representation used for fractional numbers, which modern computers are optimized for.

conceptmodulus operations

Operations that involve division and finding the remainder, which can be computationally expensive.

conceptGPU

Graphics Processing Unit, specialized hardware optimized for parallel processing, crucial for video games.

toolL1 cache

The fastest and smallest layer of cache, located right next to the CPU die, holding the most recently used data.

concept2 GHz

The clock speed of the laptop used for measurements, indicating half a nanosecond per clock tick.

toolRAM

Random Access Memory, the main system memory, significantly slower to access than caches.

productSATA SSD

A type of solid-state drive, a modern, faster alternative to traditional spinning hard drives for data storage.

productspinning disc

Traditional hard disk drives which use spinning platters to store data, much slower than SSDs.

productrouter

A networking device that forwards data packets between computer networks.

toolAbacus

Used as a simple tool for adding two numbers in one clock cycle in the computer analogy.

toolL3 cache

The third layer of cache, shared among multiple CPUs, larger and slower than L2 cache.

Found this useful? Build your knowledge library

Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.

Try Summify free