This same issue is what's holding back the clockrates in all OMAP34xx-based devices... The SoC itself is in a package that's quite good thermally, it has lots of copper with very thin insulators and the base is tight-packed with solder joints that carry away the heat produced by the processor, DSP, interfaces and all the other stuff that's integrated. But the amount of thermal power the contents of the package can churn out is somewhat large, and almost all of it has to flow away via the solder joints - there's a Flash/RAM package on top of the processor package, creating a blanket that warms up a bit even by itself (no real memory is zero power) and quite effectively prevents any heatsinking options by e.g. gluing the top of the chip to a metal part in the case (I've seen this in some teardowns on the internet, the Japanese manufacturers seem to like doing this).

The real problems, however, come from the need to get that heat away from the solder joints... And with the need to carry all the numerous signals to a meaningful place this means that the PWB under the processor is filled with signal traces, not power planes, especially in a feature-packed device such as the N900. Signal traces are good conductors of heat as they are the same copper as rest of the conducting materials in the PWB, but the ideal solution would be to have at least two essentially solid copper layers under the package as close to the surface as possible, connected with µvias made as thick as possible. This would ensure excellent heat conduction, but as a thermal connection with copper and tin-based solder is always an electrical one, it would render many signal balls useless - and many of them cannot even be grounded or connected to a certain potential. Leaving a large thermal plane floating is a terrible idea too - it will mess up the antenna performance and create nasty crosstalk issues that may be very hmm... interesting to debug. There's a very clear need for a balance between thermal performance and complexity of wiring.

Qualcomm has managed better in terms of raw thermal power - they've managed to keep the envelope sane while bumping up the clock to that oh-so-magical 1 GHz, possibly by those architecture tweaks they so advertise and further optimizing the process for lower voltages, effective power management drivers and simply selling not-that-good chips as different products. Then again, the Snapdragon platform is younger than OMAP3xxx - it just took surprisingly long for anyone to bring a mass product that uses the TI chip. Snapdragon was adopted at a faster rate, so in essence it's newer tech brought to real products sooner. (No, Qualcomm's dominance on the North American mobile RF chipset market surely has nothing to do with this :P ). TI, on the other hand, seems to be surprisingly relaxed about bringing new lower-power faster speed mobile SoCs to the market... The Cortex A9-based OMAP4xxx has been released on paper quite a while ago and so far there's almost no upcoming product announcements that use it - yet. OMAP3430 is a slightly aged product by now - there's no escaping that. The aging warhorse cannot take the faster gallop (increased clock) reliably without extra juice (higher voltage), meaning the poor thing will sweat itself to death if you're too generous with the whip.

What I'm trying to say is "overclocking a mobile device is bad, m'kay?". The effects may not be immediately noticeable - any electrical component has been designed capable for operation slightly above normal spec for some time during its lifespan. Keep it up, however, and the increased power draw may tire out the power supplies, cause the silicon parts to become ill with various diffusion-related ailments caused by too high temperatures for too long time or nasty sharp peaks in power usage and, most essentially, cause the solder joints under the package to crack - overclock the chip and you get new extremes between idle (many of the blocks are halted completely, the clock speed is throttled to a minimum, very little power used, the package cools down and becomes smaller) and full power (higher power usage than anticipated on the design table increases the raw numbers of thermal expansion, making the package larger than originally planned for and bringing the stretching of solder joints over the safe limits). The design is quite carefully optimized for a certain copper coverage under the package vs. available thermal power envelope (Why the obvious processing power usage limitations in the iPhones? Granted, no user multitasking saves battery but it also quite effectively reduces the number of times the CPU tempereature will peak in a given time interval. The things would most likely die during their normally estimated lifetimes if you allowed uncontrolled access to processing resources - there's simply not enough metal and mass in general inside them to really use all the power they hold for any extended periods without causing accelerated aging of components and burns on people's cheeks).

Exceed the safely available thermal envelope, and despite the safety margin that has been engineered in the thermal expansion WILL brick the device much, much faster than it was supposed to (no device can be made to last forever, you just engineer them to last X years with estimated ZZ% failures by then). Bump any N900 OMAP3430 spec'd at 600 MHz to even 700 and you might see it dying in just a few months if you're a heavy user - or then it may last just as long as any normally clocked device, but why take the risk? The only area where I've found my N900's raw CPU power lacking is Flash video, and even that will (according to most of the promises) be remedied in Flash 10.1 with its "toootally awsum" hardware acceleration - in N900's case, I'd guess most if not all Flash processing will be moved to the GPU and the DSP.