Oops, changed to 'horde'! The distinction is mostly architectural, in that the 'horde' is composed of an undifferentiated mass of CPUs (currently 8, could probably be expanded up to 128 with a cache-based memory hierarchy, which would certainly feel more 'horde-like'). The application processors can be accessed individually, whereas the horde can only be accessed as a group.

That's an interesting question that I haven't explored much - the network on the ZedRipper is a unidirectional synchronous ring operating at the full 140 MHz, with a round-trip latency of ~32 clocks or so, but the interface exposed to the Z80 is a sort of re-targetable serial port (you write an 8-bit 'destination' register, and then you push bytes to that node). The current buffer depth on the receive side is only a single byte, so the sender needs to wait until the destination node has read the byte and the credit gets returned. Turbo Pascal uses the 'Real48' format for floating point - 6 bytes per number - and I believe floating point operations take several thousand clock cycles. So in a tight loop on both sides, you might transfer a floating point number to a neighboring node in ~500 cycles.

Especially if I improved the network a bit - deeper receive buffers at a minimum, maybe a simple DMA engine - you could probably get it down to <100 cycles to forward a Real48 to a neighbor. The performance of emulated floating point on an 8-bit CPU is sufficiently bad, and the network performance is sufficiently good, that you probably could get away with some very fine-grained parallelism that way! When I'm back to commuting, I should write an n-body gravity simulator or something for it so that there is lots of numerical work to spread around, and see how much of a speedup I can get.

Thanks! The actual logic required for the Z80 core is tiny - you could probably make a dual-core version with something like this: https://www.tindie.com/products/tinyvision_ai/upduino-v30-lo...

FPGAs are super fun to play around with.

I have a real Kaypro 2 computer with a 4MHz Z80 in it, which I also use Turbo Pascal on - on the Kaypro, it's perfectly usable, but you get used to waiting a few seconds when you're loading files, compiling, etc. On the ZedRipper, when things are executing out of RAM everything is instantaneous. I think the CPU core I'm using is close to cycle-accurate, so it probably is ~35x faster than the Kaypro when executing actual code.

The actual logic requirements are pretty minimal (~1000 LUTs, maybe?), so it mostly depends on what kind of memory hierarchy you want. The ZedRipper relies entirely on internal BlockRAMs, which are a much more precious resource - I think I'm using all of the largest BlockRAMs, but only like ~3% of the logic resources in my chip. An easy solution to this is to replace the full 64KB of BlockRAM with a small cache, and then arbitrate for access to a large DRAM. So you wind up with a knob that you can turn to shift resources between more CPU cores with less non-cache memory bandwidth per core, or fewer CPU cores with more bandwidth. I went with the latter, but if I cranked the knob all the way in the other direction, I might be able to get somewhere in the range of 256-512 Z80 cores on this chip But with only ~1-2KB of cache per chip).

At that point, the performance will depend on the nature of your application - if it involves a tiny amount of RAM in terms of instructions and working set, the caching scheme will give you great performance - if each core needs full-speed, random access to its entire 64KB of memory, keeping everything on-die is the way to go.