Here is what our architecture looks like when using an FPGA as core fabric. Gone are the two AD75019 crosspoint switches, because audio interfaces are now fully separated from the CV ports, thanks to the use of an MD68 connector. This new design also incorporates the ES9102 audio ADC and the ES9018 audio DAC.
Clearly, we’re not looking at a “simple” synthesizer anymore. Instead, what we’re suggesting is a very advanced audio processing system, with high-fidelity audio conversion, and extremely powerful signal processing capabilities. Above, you have the Enclustra Mars ZX3, which is powered by a Xilinx Xilinx Zynq-7020 (CPU + FPGA). This module would be used as core fabric. Below, you have the Enclustra Mars AX3, powered by a Xilinx XC7A100T (CPU only). This module could be used as default digital submodule.
With such submodules, each of the 8 stereophonic channels offered by the ISHIZENO i8 would be powered by a dedicated FPGA offering 101,440 logic cells, 15,850 slices, and 240 DSP slices. This is positively mind-blowing.
One of the benefits of using the Enclustra Mars family of modules for both the core fabric and the submodules is that we could use the same tools and libraries to program both, and all developments could be made with a single base board (Mars PM3). And if we’re capable of programming the Mars ZX3 for the core fabric, we should have no problems implementing some basic DSP functions on the Mars AX3. Granted, that’s a big IF, but we should give it a shot.
And for a truly high-end system, we could use Mars ZX3 modules for both fabric and submodules. This would only give us 220 DSP slices per submodule, but it would add an ARM dual-core Cortex-A9 to each submodule. The latter could be used for running one io.js web server per submodule. This would allow each submodule to serve its own web pages and web services, through a built-in Gigabit Ethernet network, thereby removing the potential bottleneck that could be created by the primary Mars ZX3. This would require that we embed a Gigabit Ethernet switch within our module, but this should be relatively straightforward.
This design would give us 18 Cortex-A9 cores and 1,980 DSP slices. And with enough volume, the 9 Mars ZX3 would only cost $1,250. The ESS audio converters would add another $500 to the BoM, which means that a fully-loaded module could retail for $3,000.
Last but not least, at 30 × 68 mm, the Mars ZX3 module is really small, which means that we would have no problems fitting 9 of them on our backplane. And when we integrate ourselves the ES9102, ES9018, AD7606, and AD5360, we should be able to fit everything on a single PCB.