What is this about

There is a speed war on the web. Browsers compete on many fronts; security, standards support, features and speed. Most people are aware of which major browser fails on three of these, but one of them is still open for grabs. Speed.

Many organisations and users try to claim that their browser is the fastest. The Opera site claims that Opera is The Fastest Browser on Earth!. The Mozilla site claims that Firefox 1.0 empowers you to browse faster (faster than what?). Mozilla itself is quoted as being an alternative to Microsoft IE and it's faster to boot. Apple's Safari pages claim that Safari loads pages more quickly than any other Mac web browser (although they failed to show any results for Opera, and their charts fail to show results for pages that contain images). Internet Explorer users often claim that they use it because of its speed, as the alternatives take longer to start and load pages.

Opera and Firefox users are quite vocal on the subject, but none of them can actually show any research to back up their statements, ususally just saying "X takes way longer than Y to start" or "my friend uses X and it is much slower on his computer than Y is on mine". Those that did try more than one usually say "X just felt faster than Y when I tried them" (although this may refer to familiarity with the individual user interface - something that I do not cover here). I have even heard people comparing Firefox and Opera, then realised they were referring to Opera 6 (even though historically, Opera 6 is about equivalent to Netscape 4), something that was replaced with a completely new engine long before Firefox even existed.

Well, I've had enough of these unfounded arguments. I present my research. This is an honest, unbiased view (or as unbiased as I can be). Unlike most people, I not only test all the browsers, I test them all on the same computer as each other - grouped into Linux, Mac and Windows (using the same or equivalent hardware for each). This, I hope, gives an accurate comparison, not only of the browsers compared with each other, but for those that work on multiple platforms, it also gives a fairly accurate comparison of the different optimisations for each platform.

And please, those of you whose browsers did not perform as well as you had hoped, just accept it. Stop with the comments saying this article is biased. It does not matter what I use or you use. This article is not biased. The fastest browsers are the fastest browsers. I appreciate browsers for the sake of them being browsers. I feel no need to tailor tests or alter results to make a browser appear better or worse than it actually is.

The tests

In order to test each browser's speed response, I put them through a number of tests. Each test has a careful set of rules to make sure that I give unbiased results. Each set of tests are grouped by platform, and for each platform, I use just one computer to ensure that the tests compare just the browsers, and not the hardware or software they run on. Each test is done with a default browser install, without tweaking any settings (I know that many browsers perform slightly better if you tweak their network settings, but this is intended to be a test of a standard browser install - some people also suggest that using a native skin makes a browser faster, but I got virtually identical results with native and non-native skins). With Browsers that also offer email or news features, I enabled these clients, but did not have any email or news items in them (some of them may perform differently if they did, but that is not what I was trying to test).

The idea was to find out the fastest browser(s) at performing the major tasks that browser engines are expected to perform. The basic requirements were HTML, CSS, JavaScript, basic DHTML and images - if the browser engine was not capable of performing these tasks then I am not interested and it was not tested.

The tests are not perfect. Obviously nothing can be perfect. But they do give a very good idea of how each browser would cope in each type of situation. Importantly, the tests were not (intentially) tailored towards the strengths or weaknesses of any particular browser.

Cold start

This is the time it takes to do a cold load. I log out then in, and then once all background processes have completed, I run the browser as the first program. I use the default settings for the browser, without preloading or quicklaunch tools (unless the browser installs itself this way, as with Internet Explorer on Windows). The browsers are all set to show a single page on startup, and this page is a locally stored basic HTML page.

Warm start

After completing a cold start, I close the browser, then time how long it takes to start again. I restart it again another two times, and take an average.

Rendering CSS

To test CSS rendering speed I use a CSS benchmark test devised by nontroppo. The test measures the time it takes the browser to render a page consisting of almost 2500 positioned DIVs. The page is stored locally, loaded once to pre-load it, then reloaded 3 times, and the average time taken for those three renderings. The page is the first page loaded after starting the browser (after logging out and in). All browsers took significantly longer during the initial load, which is why I discount this initial load, as it does not reflect the reality of normal rendering.

Rendering table

To test table rendering speed I load a local copy of my javascript libraries page. I ignore this first load and then go back and click the link again (ensuring the page and all associated files are cached). I then use a script to time how long it takes to render the page. I take the average of three loads.

Script speed

To test scripting speed I use the fantastic benchmarking tool at 24fun.com. This is intended for testing multiple browsers on one computer to see how well they perform compared with each other. Small variations in hardware or software can cause big differences in results so it is important that all tests are done on the same computer. After logging out and logging in, I load a browser, clear its cache, enable popups (a requirement of the tests), and then run the tests. The tool tests a good variety of different things, including mathematical calculations, DHTML, string manipulation, image swapping, table manipulation, page content manipulation and window management. With most browsers I test twice and take an average but I did not do that with iCab because it took half an hour the first time, and required me to sit there and babysit it, repeatedly telling it that the script was not an infinite loop, it was just a horrendously slow browser!

Multiple images

This is done to reflect normal web site handling. I make the browser load a page containing several small images. This requires multiple server connections (or keepalive) and is a good test to see how well the browser manages its network connections, as well as how well it can cope with rendering multiple images simultaneously. After logging out and logging in, I load a browser, clear its cache, then I use Google image search to search for the word 'solid'. This ensures that any referenced pages are pre-cached, and any required DNS lookups are completed, to minimise the effects of DNS delays. Then I load use Google image search to search for three more words (pine, red, stone - don't blame me for the results that come up, I wasn't expecting them either), and I use the average load time of these three pages. (The timer starts when I hit 'search', and stops when all images have completed loading.) My connection speed is 100 Mbps and has no discernable effect on the results. I have also tested it on a 700 Kbps connection and got similar results. Google search time is typically below 0.03 seconds.

History

Something that is commonly overlooked when checking browser speed, but is very important to overall browsing speed. I use Google image search to search for pine, and I navigate through the result pages 1 to 25. I then start the timer, and go back as quickly as possible to the first result page, then forward as quickly as possible to page 25 again. Each intermediate page must be allowed to load completely - including images - before moving to the next page. (This means that any indicators that the browser provides to show that the page is loading must show the page as loaded before navigating to the next page). I use the fastest means the browser has to navigate, which typically means using keyboard shortcuts. This test demonstrates how efficiently the browser uses caching in order to improve browsing speed.

Note: I did not test plugins like Java or Flash, as these are external programs, and not the browsers themselves. They are affected much more significantly by their own version numbers or memory contraints of the Operating Syetem, and are not really in the browser's own control.

Test results

Sorted alphabetically:

• Linux browsers
• Mac browsers
• Windows browsers
• Supplementary browsers

Test results are shown by default in tabular form. If your browser's scripting engine is up to the task, you can click on the column headings to sort the table by that heading. The fastest browser is marked in each test using light green, or salmon if it was the fastest, but made mistakes trying to obtain that speed. Some results also have more information, that will appear in a tooltip when you hold your mouse over the result. These can be identified in good browsers (I have done an ugly DHTML behaviour to make this happen in Internet Explorer as well - this excuse-for-a-browser is really annoying me) by the Help cursor that the mouse should use while you hold it over those results, and by the » character.

Linux browsers

The tests are done on SuSE Linux 9.1, using KDE (except Epiphany, whose results are shown for Gnome, as that is its normal operating environment). I also tested the browsers on Gnome, but only got significantly different results in Konqueror, so that is the only browser's result I will show for both Gnome and KDE. I only concentrate on the main Gecko based browsers - there are many variants available. Mozilla (SeaMonkey), Firefox and Epiphany are the ones I will test, as these are either the official releases from Mozilla.org, or the official browser of the Gnome DTE. I could not test Galeon (one of the most popular Gnome browsers) as it was not available for the Mozilla version I had installed. However, since it is just a fancy skin for Mozilla (like Epiphany), it's times are likely to be similar to Mozilla's. Opera 8 tests were done using Opera 8 beta 1. Mozilla 1.8 tests were done using Mozilla 1.8 alpha 6. Firefox 1.5 tests were done using Firefox 1.5 beta 2. Opera 9 tests were done using Opera 9 technical preview 1.

NOTE: I have updated the page after discovering that I had done the multiple image test with Opera 8 on Linux set to 'ID as Opera' (not the default setting). This is important because Google uses the wrong HTTP header to check if the browser can handle a gzip compressed page, which Opera can. Instead of checking the Accept-Encoding header, which correctly says this, they check the User-Agent, and only use gzip if the browser claims to be Mozilla or MSIE (other browsers like Safari and Konqueror are also included in this). Setting Opera to use 'ID as MSIE' (default setting) or 'ID as Mozilla 5.0' makes Google send it the compressed version as well. This makes a huge difference, from Opera appearing to be almost the slowest at 2.50 seconds, to being clearly the fastest at 1.82 seconds.

Hardware; 800 MHz Intel Pentium 3, 256 MB RAM.

Mac browsers

I concentrate mainly on browsers for Mac OS X because no browsers for Mac OS 9 are currently in development (with the exception of iCab; IE 5 is no longer being updated. There is no standard recent release of Mozilla or related browsers for OS 9. Opera 7 is not being released for OS 9. Safari, OmniWeb and related browsers only work on OS X). Opera 8 tests were done using Opera 8 technical preview 1. Mozilla 1.8 tests were done using Mozilla 1.8 alpha 6. Firefox 1.5 tests were done using Firefox 1.5 beta 2. Opera 9 tests were done using Opera 9 technical preview 1.

The tests are done on Mac OS X 10.3.7 (Panther) with the exception of the Safari 2.0 test that is done on a 10.4 (Tiger) preview. (I tested the other browsers on Tiger as well just to make sure this was a fair test, and got virutally identical responses as with Panther.) The Safari 2.0 update shows an impressive performance improvement with CSS, DHTML and window handling. The Opera 8 update has also shown a significant improvement, nearly doubling its scripting speed, and improving load time and cache handling. (For more Opera/Mac related tests, see my Opera for Mac speed improvements article.)

Hardware; 400 MHz G4, 256 MB RAM. Supposedly the Harvard architecture of the G4 chip makes this approximately equivalent to an 800 MHz Intel Pentium chip, and therefore this computer is of comparable spec to the one used to run Windows and Linux.

Windows browsers

The tests are done on Windows XP SP2. Internet Eplorer 6 is a standard install, Internet Explorer 5 is the zip package install that can be downloaded from QuirksMode.org. (IE 5.5 crashed every time I tried to run it.) The Windows performance settings are set to the automatic setting (default). Opera 8 tests were done using Opera 8 beta 1. Mozilla 1.8 tests were done using Mozilla 1.8 alpha 6. Firefox 1.5 tests were done using Firefox 1.5 beta 2. Opera 9 tests were done using Opera 9 technical preview 1. I also tested the Moox Firefox install, "optimised" for my particular processer. I tested the new Netscape, which is basically Firefox with some added rubbish that slows it down. K-Meleon is the only other Gecko browser I tested, as is said to be the fastest one (although in fact it is slower than Mozilla and Firefox at all tasks except starting).

I also tested a few browsers that are not current browsers (Netscape 4, Escape 4, Clue, Internet Explorer 5.0, Mozilla 1.0, Opera 6.03). This is done purely for the sake of interest (and it is interesting to note that IE 5.0 is significantly faster than IE 6.0).

Hardware; 800 MHz Intel Pentium 3, 256 MB RAM.

Conclusions

No, no. That is up to you. It all depends on how you use your browser and what you use it for, and what operating system(s) you use it on. Take a look at the tables (and graphs, if your browser's scripting engine is up to the task) and work it out for yourself.

Ok, ok. Firefox and Mozilla are clearly optimised for Linux, and Opera is clearly optimised for Windows. These optimisations are mostly obvious with the loading times, although there is also a little difference in the cache handling on the different operating systems. However, Opera seems to perform admirably well on most tasks, on any platform. When it comes to page rendering (tables, CSS or images), most of the major browsers perform very fast, with very little to distinguish between them. When it comes to scripts, Opera clearly holds its head above the others, nearly twice as fast as the others. The only one that comes close is Safari 2.0, but that is tied to the Tiger release of Mac OS (currently in preview).

Opera also is a clear winner using history. In fact, on Linux it is faster than Mozilla and Firefox for all except starting time. On Mac and Windows, Opera is faster than Mozilla and Firefox for all tasks. Surprisingly, Mozilla is now faster at most tasks than Firefox (please don't send me any more emails about this line, I am well aware of why it is faster). Internet Explorer on Windows was either as fast as - or faster than Mozilla and Firefox for most tasks, with the exception of scripts, where it took over twice as long. Of course, its poor standards and security clearly make it a much less attractive prospect. The Moox Firefox install is actually slower than the standard Firefox versions distributed from Mozilla.org, even though it is supposedly optimised for my particular processor. The Mac optimised version is a bit better, but trades performance in one area for performance in another, making little or no difference overall. The performance of K-Meleon and Epiphany was similar to the performance of Mozilla and Firefox on the same platform. The new Netscape Browser preview (based on Firefox) was clearly suffering from bloat caused by the AOL add-ons.

For those of you who commented that security has nothing to do with speed, try saying the same thing after you have spent hours (or days) cleaning malware off your computer, caused by you running an insecure browser. Overall, the time you waste cleaning up malware more than counterbalances the few seconds you save with the faster engine. And as for standards support, well, speed is not much use if pages are forced to be more bloated and therefore slower to load, simply because the engine does not attempt to support current standards. For example, try seeing how much longer it takes to load a page if it is made using several levels of nested tables instead of clean CSS styling, or one that has to use Dean Edwards' IE7 script library just to get the page to render correctly.

From a humorous perspective, it is interesting to note that IE on Windows is slowing down, in the majority of tests. Browsers that use the underlying IE engine (Maxthon or Avant, etc.) generally make IE slower, although Avant manages to make an increase in scripting speed. Opera 6 was also slightly faster at rendering simple pages than Opera 7-9 on windows and linux, but this is easily outweighed by its slow and outdated script support, lack of reflow and progressive rendering, slower image and network handling, and slower history on Mac and Linux. On the other hand, Mozilla has clearly improved a tremendous amount since it's original 1.0 release (which was actually slower in most respects than its Netscape 4 ancestor), but it still has a lot of catching up to do before its scripting engine can match Opera's. And it still has competition. Opera's already faster scripting engine nearly doubled in speed in the 8.0 release, and remains fast in the 9.0 release.

One thing I do not take into account here (although it can play a significant role in real pages) is progressive rendering. For example, with Opera 6, the table was laid out faster, but nothing was displayed until the entire table was complete. With Opera 7+, the table takes longer to complete (about half a second), but it is progressively displayed, so the first part is displayed as soon as it is ready, without having to wait for the rest of the table to complete. As a result, you can actually start reading the page faster with Opera 7+. With pages that are served by slow servers (or if you have a slower connection), this can make overall browsing speed significantly faster.

The fastest browsers at starting are the less commonly used Konqueror (only on KDE), Camino and iCab 2, with Konqueror by far the fastest at a warm start. However, it has one of the slowest script implementations of all the major browsers on Linux, and even though its cache handling is a bit faster on Gnome, its slow startup time on Gnome outweighs the benefit of better caching. The only browsers that are slower at scripts are Opera 6 and Mozilla 1.0, both of which are old releases that have long since been replaced with much faster versions.

Konqueror's sister browsers; Safari 1.2 and OmniWeb are also extremely slow at scripting. IE for Mac is also slow at scripting, and although it is fast enough to start, it is slow for the other tasks. iCab and (old) OmniWeb 4.2 are the clear losers on Mac, with OmniWeb 4.2 being the slowest at tables and CSS, and failing completely to run the script tests. iCab 3 managed to produce the slowest scripting engine time, taking over half an hour - over ten times as long as any other Mac browser, and the slowest CSS time of any current browser, taking nearly 20 minutes (note, also see supplementary Amaya results below). Safari 2.0 has made an impressive advance compared with Safari 1.2, in both scripting and rendering. iCab 3 has made significant advances in its CSS and scripting capabilities, but this has been at the expense of making it perform noticably slower than before at almost all tasks. Supposedly it has not yet been optimised for speed (since it is currently in closed beta), but it has a long way to go, and it is doubtful that any optimisations will make much of a difference.

On Mac OS 9 the browsers generally perform slightly better than the same versions on Mac OS X. However, these versions have long since been left behind by newer and much faster releases that are not available on OS 9. iCab 2 is the fastest at starting but then fails to perform well enough on most of the other tests. iCab 3 may have far better capabilities, but it is slower as a result, and fails at some tasks that iCab 2 managed to perform. The other browsers are generally very slow at scripting, CSS rendering, history, and none of them are particularly fast at starting either.

The java browsers all performed very badly at scripting, most seemed to have a lot of trouble manipulating strings or performing calculations. ICEbrowser, although initially one of the slowest to start, was the clear winner at all other tasks, as well as having the most up-to-date scripting and CSS engines. Clue was clearly the worst browser, failing to run most of the tests. Well, at least I got it to start.

To summarise

So overall, Opera seems to be the fastest browser for windows. Firefox is not faster than Internet Explorer, except for scripting, but for standards support, security and features, it is a better choice. However, it is still not as fast as Opera, and Opera also offers a high level of standards support, security and features.

On Linux, Konqueror is the fastest for starting and viewing basic pages on KDE, but as soon as script or images are involved, or you want to use the back or forward buttons, or if you use Gnome, Opera is a faster choice, even though on KDE it will take a few seconds longer to start. Mozilla and Firefox give an overall good performance, but their script, cache handling and image-based page speed still cannot compare with Opera.

On Mac OS X, Opera and Safari are both very fast, with Safari 2 being faster at starting and rendering CSS, but with Opera still being distinguishably faster for rendering tables, scripting and history (especially compared with the much slower Safari 1.2). Camino is fast to start, but then it joins its sisters Mozilla and Firefox further down the list. Neither Mozilla, Firefox nor IE perform very well on Mac, being generally slower than on other operating systems.

On Mac OS 9, no single browser stands out as the fastest. In fact, my condolences to anyone who has to use one of them, they all perform badly.

These results are drawn from my tests as shown above, and are confirmed by my personal use. Please do not send me hate mail saying I have insulted your browser or whatever. If you get different results, try doing what I did. Perform a set of tests like these that cover the major areas of use. Make sure you do it unbiased, cross platform, using the same or equivalent spec on each (make sure you use the same amount of RAM in each). Once you are done, don't tell me about it. Publish your own results, as I have done, with a discussion of why you did each test and what it represents. Facts and figures are everything. Oh, and don't just say "I couldn't get it to install so I didn't bother" - you have no idea how much work I put into getting all these browsers to install on these operating systems (as my wife and the IRC community would be able to testify to). If you can't get it to install, then that is your problem, it says nothing about speed, and simply makes your results incomplete.

Updating the results

Since I get a lot of emails asking about updating the results for new browser versions, I will explain when I will and will not do updates:

• For the major browsers, I will add new results when major or important new versions come out. The results do not change much with minor releases, so I will usually only update when the browser gets a new major rendering engine version. Even so (as with Firefox 1.5), this does not always make very much of a difference. It may take me a little while to test the browsers, as I must test them on multiple platforms. Additionally, with some browsers (such as Konqueror), testing new versions requires me to upgrade my entire operating system. Please be patient.
• For the minor browsers, I will add new results only when major new versions come out, and sometimes only if the underlying engine has actually changed enough for the timings to change significantly.
• For the browser skins, such as Avant, Maxthon, and Epiphany, I will amost never retest. There is no point. They hardly change at all since they rely on the underlying browser. Their results are more likely to change depending on how the underlying browser changes - see the underlying browser (usually Mozilla or Internet Explorer) for more details. I will not add results for any more skins (such as GreenBrowser, SlimBrowser, or Galeon) - they are all very similar to each other, so look at the results for the other similar skins.
• I will not add results for tweaked installs, such as where you change the network settings, or disable splash screens. I explain my reasons for this in the tests section.
• I will not add results for any more clones (like Skipstone). They are all very similar to each other, so look at the results for the other similar clones.
• I will add results for new browsers, assuming they are capable of performing the required tasks, and assuming they are not just skins or clones.

Subscribe to my RSS feed to see when I update, and see my previous email for details of when I have made updates.

Supplementary results

I also tested a few other browsers that were not up to the required capability spec, but are used for accessibility, or general speed reasons, or often by system administrators who do not have a GUI. In several tests, Lynx and Links2 were so fast that it was nearly impossible to time, as pages were often loaded faster than buttons could be pressed. Amaya and HotJava are concept browsers, and are included just for the sake of interest. All tests were on SuSE Linux 9.1 with KDE, except HotJava which was done on Windows XP (as with the other Java browsers).

Hardware; 800 MHz Intel Pentium 3, 256 MB RAM.