It's actually not all that fast in a lot of cases. Boost's matrix manipulation support is particularly slow:
http://eigen.tuxfamily.org/index.php?title=Benchmark-August2...
Also a lot of the template aerobatics generate a ginormous number of symbols which massively increase the binary size and cache pressure. In a former life I ripped out a pile of heavily templated code using Boost and replaced it with a well placed virtual function or three and dramatically reduced the cache misses in the application.
I've always seen Boost (and to a large extent the STL) as a demo library to show off what's possible with templates rather than something to be used in everyday code and have worked on a number of projects which forbade both.
There's a large rift in the C++ world between what I generally label Qt and Boost camps. The Boost people see the Qt folks as Java-esque dimwits that can't be bothered to learn the full extent of C++'s power and the Qt folks see the Boost following as unpragmatic architecture astronauts.
You've picked a benchmark of the most highly optimized matrix algebra libraries, and found that uBLAS is mid-pack, even though a) it isn't particularly optimized for any platform, and b) its goal is to result in the clearest possible syntax without sacrificing adequate performance. Talk about picking nits.
"Also a lot of the template aerobatics generate a ginormous number of symbols which massively increase the binary size and cache pressure."
When it really matters (which it rarely ever does): man strip
"I've always seen Boost (and to a large extent the STL) as a demo library to show off what's possible with templates rather than something to be used in everyday code"
I worked on shipping commercial products that used boost and the STL, and I did nearly all of my speed-sensitive code with it in grad school. It's far more than a demo library. Like I said before, you don't have to use all of it to use some of it.
Second of all, those symbols will have an effect on compile times, but they should certainly be stripped out by the linker. Why would symbols that do nothing sit in the executable?
Thirdly, Boost's purpose is to iterate faster than the C++ standard library and have real-world data about what libraries work and are needed. Those libraries would then be included in next C++ standards, as it has happened with TR1. There are useful libraries inside Boost and there are research libraries or concept libraries. Don't use a library just because it's included in Boost!
I use Boost and Qt. Since TR1 I didn't need Boost as much, which means that it's achieved one of its goals.
With gcc 4.4.5 and -O2, this one small file takes 22 seconds (!) to compile and produces a 14MB output file.
I guess that's... tolerable. But it isn't "lean" in any meaningful sense.
Your point seems to just be agreement with mine: C++ bloat is tolerable. It's still bloat.
Seriously, I think it's too much trick to do meta-programming in the language. If meta-programming should be allowed, it should be easier, simpler, and for the ordinary people. Just like Ruby.
But why does the metaprogramming syntax in C have to be so convoluted, bloated and difficult? Do we really need >30 second compile times per file on recent CPUs? If a language is hard to parse both for computers and humans, something went wrong. It should be easy to read for one of both, preferably both.
Aside from your tears, what is the evidence that boost is bloated? Too slow? Too much source code? Too many files?
I've been using Boost for nearly a decade. It does a ton of useful stuff (from template metaprogramming to fast matrix algebra), it's loosely coupled, and it's insanely fast. Given it's scope, it's the least bloated library of which I know.
C++ has been evolving, but for the last decade, all of the evolutionary changes have come in the world of templates and template metaprogramming. I think this is hard on the people who have restricted themselves to the C-with-classes mentality. They're missing every modern development in the language since ~1998.