3 ms·
60 kilohertz on 6 milliwatts sounds pretty bad (that's 100 milliwatts per megahertz and so 20 milliamps per megahertz if we assume 5 volts, while 0.06 milliamps
by kragen 2y ago
60 kilohertz on 6 milliwatts sounds pretty bad (that's 100 milliwatts per megahertz and so 20 milliamps per megahertz if we assume 5 volts, while 0.06 milliamps per megahertz is common for low-power processors) but it's actually far, far worse than it sounds because serv is bit-serial, requiring, i think, 32 cycles per instruction. so you're looking at something like 5000 times the energy consumption of existing off-the-shelf microcontrollers
the suggested price of a dollar is about 10x worse than something like the py32, ch32v003, or pmc150, which are also faster and more power-efficient
that doesn't mean this is bad research! it just means it isn't yet developed to a state where there's likely to be a market for it. it's very helpful to know that serv occupies 12600 gates, for example, and that the flex-rv process provides 720 gates per square millimeter. it's very plausible you could design something useful with it that had 600 gates, was less than a square millimeter, used 300 microwatts at 60 kilohertz, and cost five cents, for example; that's a niche that silicon photolithography is struggling to fill because of high per-chip costs. you could fit a 6502 into twice that
another potentially interesting niche is low power density; for implanting into your body you don't want hot spots that can burn your tissues (though you'd have to encapsulate the igzo behind something biocompatible)
- hajile 2y agoThe 1-bit serial version is 2100kGe (excluding registers). Their talk claims the 4-bit variant is 2.5x larger (while being 4x faster) which is around 5250kGe. Add in all the registers and you get pretty close to that 13kGe number (assuming you're using 31 registers instead of 15 in the 32e spec).
- kragen 2y agothank you for the extra context! i had assumed they were using the 1-bit serial version and that for whatever reason they just needed a lot more gates than anyone had expected from fpga simulations. i didn't realize serv had a nybble-serial option, and that's a very significant difference. so it's closer to 1000× the energy consumption of off-the-shelf microcontrollers than 5000×