The three most influential factors for Eclipse speed are:

• Using the latest version of Eclipse (2020-06 as on 26 June 2020)
  Note that David Balažic's comment (July 2014) contradicts that criteria which was working six years ago:

The "same" workspace in Indigo (3.7.2) SR2 loads in 4 seconds, in Kepler SR2 (4.3.2) in 7 seconds and in Luna (4.4.0) in 10 seconds. All are Java EE bundles. Newer versions have more bundled plugins, but still the trend is obvious. (by "same" workspace I mean: same (additionally installed) plugins used, same projects checked out from version control).

• Launching it with the latest JDK (Java 14 at the time of writing, which does not prevent you to compile in your Eclipse project with any other JDK you want: 1.4.2, 1.5, 1.6 older...)

    -vm jdk1.6.0_10\jre\bin\client\jvm.dll
  

• Configuring the eclipse.ini (see this question for a complete eclipse.ini)

    -Xms512m
    -Xmx4096m
    [...]
  

The Xmx argument is the amount of memory Eclipse will get (in simple terms). With -Xmx4g, it gets 4 GB of RAM, etc.

Note:

1. Referring to the jvm.dll has advantages:

• Splash screen coming up sooner.
• Eclipse.exe in the process list instead of java.exe.
• Firewalls: Eclipse wants access to the Internet instead of Java.
• Window management branding issues, especially on Windows and Mac.

Dec. 2020, Udo conforms in the comments

From version 4.8 (Photon) an up there was a steady speed gain after each version.
The main platform was optimized every release to load faster, enable more features for the dark theme and to add more features for newer Java versions for the Java development tools.
Especially with-in the last 3 versions the startup time was increased a lot. There should be a significant increase in start-up time with the newest version of Eclipse 2020-12.

In my experience it started a lot faster with each new version.
But: There are still plug-ins which do not follow the new way of using the Eclipse API and are therefore still slow to start.
Since the change to Java 11 as the minimum runtime version starting from Eclipse version 2020-09 at least the core system uses the newer features of the JVM. It is up to the providers of the other plug-ins to upgrade to newer APIs and to use the full power of modern CPUs (e.g. concurrent programming model).

Several reasons

Header files

Every single compilation unit requires hundreds or even thousands of headers to be (1) loaded and (2) compiled. Every one of them typically has to be recompiled for every compilation unit, because the preprocessor ensures that the result of compiling a header might vary between every compilation unit. (A macro may be defined in one compilation unit which changes the content of the header).

This is probably the main reason, as it requires huge amounts of code to be compiled for every compilation unit, and additionally, every header has to be compiled multiple times (once for every compilation unit that includes it).

Linking

Once compiled, all the object files have to be linked together. This is basically a monolithic process that can't very well be parallelized, and has to process your entire project.

Parsing

The syntax is extremely complicated to parse, depends heavily on context, and is very hard to disambiguate. This takes a lot of time.

Templates

In C#, List<T> is the only type that is compiled, no matter how many instantiations of List you have in your program. In C++, vector<int> is a completely separate type from vector<float>, and each one will have to be compiled separately.

Add to this that templates make up a full Turing-complete "sub-language" that the compiler has to interpret, and this can become ridiculously complicated. Even relatively simple template metaprogramming code can define recursive templates that create dozens and dozens of template instantiations. Templates may also result in extremely complex types, with ridiculously long names, adding a lot of extra work to the linker. (It has to compare a lot of symbol names, and if these names can grow into many thousand characters, that can become fairly expensive).

And of course, they exacerbate the problems with header files, because templates generally have to be defined in headers, which means far more code has to be parsed and compiled for every compilation unit. In plain C code, a header typically only contains forward declarations, but very little actual code. In C++, it is not uncommon for almost all the code to reside in header files.

Optimization

C++ allows for some very dramatic optimizations. C# or Java don't allow classes to be completely eliminated (they have to be there for reflection purposes), but even a simple C++ template metaprogram can easily generate dozens or hundreds of classes, all of which are inlined and eliminated again in the optimization phase.

Moreover, a C++ program must be fully optimized by the compiler. A C# program can rely on the JIT compiler to perform additional optimizations at load-time, C++ doesn't get any such "second chances". What the compiler generates is as optimized as it's going to get.

Machine

C++ is compiled to machine code which may be somewhat more complicated than the bytecode Java or .NET use (especially in the case of x86). (This is mentioned out of completeness only because it was mentioned in comments and such. In practice, this step is unlikely to take more than a tiny fraction of the total compilation time).

Conclusion

Most of these factors are shared by C code, which actually compiles fairly efficiently. The parsing step is a lot more complicated in C++, and can take up significantly more time, but the main offender is probably templates. They're useful, and make C++ a far more powerful language, but they also take their toll in terms of compilation speed.