Ok, so - I'm not a DSP programmer, but I've done enough systems work to have a general sense of what's going on. There are a few competing factors at play.
First off, there's nothing all that special in the SHARC instruction set. A modern x86 (or even ARM!) processor has many of the same instructions and can even do vector versions of those operations. But you have to look at it from the perspective of a total system design. This is a real time embedded system. You have an analog signal in and in most cases an analog signal out. That means A/D conversion, digital processing, and D/A conversion need to all be done at
very tight latency requirements. The DSP chip is designed to get those signals into the chip, processed, and out again quickly. Its operations are focused around that purpose, and when coded properly you don't have to deal with pipeline stalls, cache misses, TLB faults, page faults, interrupts, or a mountain of other garbage that can get in the way of processing on a normal off the shelf computing platform. And while I haven't dug into it, I suspect that synchronization across two DSPs is dramatically easier and more reliable than synchronization of CPU cores.
Second, you have power and heat requirements. Those 4 GHz chips are also going to be dissipating 90W on their own, and there are other components required to complete the system. It takes heavy heat sinks and multiple fans to dissipate the heat being generated. Then there are the support components involved. You've seen the size and weight of a PC power supply, right? There are expensive VRM capacitors that pull down the +12V lines again to the 1.1V or so (at quite a few amps) that the CPU depends on. Sure, you can start scaling down the power requirements, but once you get down to levels that are appropriate for something like the 30W Axe Fx? All of a sudden you'll find that compute performance has gone to shit and is actually worse than those focused DSP chips. Mobile chips are powerful for their size, but there's a reason BIAS sounds like trash next to Axe or even an 11r.
Hell, think about the physical requirements. Have you ever had to sit in a room with rackmount servers? They're probably 20 dB louder than the grinding "loud" fan on an Axe Fx. Significant airflow is required, too. A sealed box like the Axe would never fly.
Another issue that probably doesn't apply to Axe but applies to a whole host of other applications where a DSP is used - reliability and robustness. They put these things in missiles, rockets, submarines, etc. The error rate on an Intel x86 chip (that is, computations that return invalid results) is going to be significantly higher even in normal use, and harsh environments will make things worse. General purpose platforms are not designed to sustain this type of abuse. An Intel chip lives on the ragged edge of its physical limits and those limits are easily disturbed. This is why you constantly hear news stories where space equipment, fighter jets, etc are being produced with processors that are 8+ years old. The simpler, older designs are required to provide sufficiently robust platforms, and often there's a ton of extra improvements made.
All that said, it's not immediately obvious to me why a TigerSHARC is so expensive or what makes it preferable to something like a TI C6000 DSP which is
claiming comparable performance at a fraction of the price.
