Despite vastly increased processing power, many modern computers exhibit higher input-to-display latency than systems from the late 1970s and early 1980s. This counterintuitive trend is primarily driven by the inherent complexity of modern operating systems and software pipelines, which introduce numerous handoffs, buffering, and scheduling delays. While high-end gaming rigs and devices like the iPad Pro can achieve competitive latencies through significant engineering, most off-the-shelf modern devices fall short of their decades-old predecessors.
Every atomic assertion extracted from the underlying record, ranked by evidence strength.
"Normal" modern computers cannot compete with thirty to forty-year-old machines in terms of latency.
"Complexity" is identified as a root cause of latency bloat.
Gains in latency reduction that bring modern machines back to the quickness of 70s and 80s machines have come from piling on more complexity, not reducing it.
For very simple tasks, people can perceive latencies down to 2ms or less.
The custom Haswell-E was tested with G-Sync on, with no observable difference in latency.
The Apple IIe effectively scans at 556 Hz.
The Ergodox keyboard claims to scan at 167 Hz.
Throughput is widely understood and measured, unlike latency.
Poor performance in aspects that are not measured is a common theme.
Google has only recently been serious about the end-to-end experience.
Measurements were done with settings as close to the default OS settings as possible.
Differences of 10ms in latency should probably not be considered significant.
The author expects to get benchmarks from a lot more old computers the next time they visit Seattle.
A display advertising 1ms switching time can actually take 10ms to fully change color once it starts changing.
Much of the complexity in modern systems provides direct or indirect benefits.
Adding a cache to reduce latency makes a system more complex.
On the Apple IIe, there are basically no handoffs, locks, or process boundaries.
The solution to poor performance caused by "excess" complexity is often to add more complexity.
CRTs from the 90s offered better resolution and higher pixel density than affordable non-laptop LCDs until relatively recently.
The Nielsen Group claims that latencies below 100ms feel equivalent and are perceived as instantaneous.
The Apple IIe CPU is a 1 MHz chip with 3.5k transistors.
Some people working on Pixel phones care about end-to-end latency.
Desktop results are measured from when the key started moving until the screen finished updating.
Fancy gaming rigs with unusually high refresh-rate displays are almost competitive with machines from the late 70s and early 80s.
The Apple IIe has significant advantages in input, output, and between input and output compared to modern computers (excluding the latest iPad Pro).
The author has a nagging feeling that the computers used today feel slower than the computers used as a kid.
Human perception has been shown to be unreliable in empirical studies.
A moderately quick page turn in a physical book appears to be about 200ms.
Color scales for latency and year are linear, while clock speed and transistor count are log scale.
The author measured the response latency of devices using a high-speed camera.
Almost every computer and mobile device that people buy today is slower than common models of computers from the 70s and 80s.
Andy Matuschak, a former iOS/UIKit engineer, provided a high-level sketch of the iOS rendering pipeline.
State-of-the-art gaming monitors can get total display latency down into the same range as old CRTs, but very few people own such displays.
A 90ms improvement in latency is observed when going from 24 Hz to 165 Hz refresh rate on the same machine.
Key overhead in iOS comes from imperfectly aligned periodic scanrates, many cross-process handoffs, and extensive locking necessitating trips into kernel-land.
Removing complexity is much harder than advocating for its reduction.
Low-latency gaming desktops and the iPad Pro can get into the same latency range as quick machines from thirty to forty years ago.
Much of the "accidental complexity" in systems is present because it is convenient.
Results are rounded to the nearest 10ms to avoid the impression of false precision.
Individual developers are improving the experience for people who use certain, very carefully coded, applications.
The computer latency tests measure the time between a keypress and the display of a character in a terminal.
Increasing latency is not only noticeable to users, but it also causes users to execute simple tasks less accurately.
The results are sorted from quickest to slowest.
The Ergodox uses a 16 MHz microcontroller with approximately 80k transistors.
The fastest machines in the computer latency tests are ancient.
Component vendors optimize for intermediate targets that are only weakly correlated to the actual user experience.
Newer machines show a wide range of latencies.
Excluding the Game Boy Color, all of the quickest mobile devices are Apple phones or tablets.
The BlackBerry Q10 is unusually quick for a non-Apple device, possibly due to having actual buttons which are easier to implement with low latency than a touchscreen.
The Palm Pilot 1000 and Kindles are at the bottom of mobile latency due to older touchscreen technology and e-ink displays, respectively.
Modern keyboards commonly scan their inputs at 100 Hz to 200 Hz.
The Apple IIe does not have to deal with context switches or buffers involved in handoffs between different processes.
Modern OSes let programmers write generic code that can deal with having other programs simultaneously running on the same machine, but this comes at a huge complexity cost and significant latency penalty.
The TI 99/4A, released in 1981 with a 3 MHz clock speed and 8k transistors, has a measured computer latency of 40ms.
The Apple IIe, released in 1983 with a 1 MHz clock speed and 3.5k transistors, has a measured computer latency of 30ms.
The Lenovo X1 Carbon 4G running Windows, released in 2016 with a 2.6 GHz clock speed and 1G transistors, has a measured computer latency of 150ms.
A change to the screen at 144 Hz will have 0ms to 7ms of extra latency as it waits for the next frame boundary before getting rendered.
At 144 Hz, each frame takes 7ms.
The expected latency from waiting for the next frame and the actual color change on a 144Hz display is 13.5ms.
A custom Haswell-E system with a 24Hz display, from 2014 with a 3.5 GHz clock speed and 2G transistors, has a measured computer latency of 140ms.