It's probably better to compare it with a FPGA that allows for rapid programming, than a multicore computer. Only you can sort of configure the whole thing as a SOC. Which has great power management. I'd love to see some updates from Greenarrays -- there hasn't been much news lately.
[1] http://www.greenarraychips.com/home/documents/greg/AP003-110...
[e] http://www.greenarraychips.com/home/documents/greg/AN007-141...
Also, "implement an SRAM interface" - perhaps, but not one particularly easy to program against, I believe.
Maybe "really, really smart dsp" is a way to look at it...
I mentioned DRAM because
* An FPGA could do it and the GA chips probably can't,
* You can't do a whole lot of embedded DSP apps without DRAM - it's not a question of "trying to be a PC", you just need memory to store stuff :-) and SRAM is too expensive.
Without DRAM, you're pretty much limited to what you call "streaming"; a whole lot of things don't work without being able to keep a working set in RAM. I don't argue that the chip is useless, just that it's useless in a huge range of use cases including many embedded ones, which is a sharp contrast to the connotations of its self-description as a "multi-computer chip"... And BTW even a really really dumb DSP chip can usually do DRAM :-)
I look at it as "an array of incredibly weak and energy-efficient cores" :-)
Anything anyone in the industry dares to call a "core" (apart from NVIDIA's ALUs misnamed as "CUDA cores" or whatever) blows a GA "computer" out of the water, in particular, in terms of memory (you'll get kilobytes or more likely tens of kilobytes of cache and/or RAM per core.) And anything anyone in the industry dares to call a "computer"... well, it's in an entirely different league from a GA "computer", for starters, well, damned peripherals are considered to be a part of a "computer".
Good luck running anything of use on 144 tiny cores only talking to their neighbors, only supporting 18b integer arithmetic and having tiny RAMs for code and data. (Do you even get a DRAM interface? How about Ethernet?)
The chip is a bloody impressive achievement - an async design, no reliance on standard EDA tools, etc. etc. And perhaps it has its uses in really - really - energy-efficient applications with very - very - limited functionality.
But the "multi-computer" business... let's say that it makes GPU marketing look honest.