On Fri, Apr 28, 2017 at 10:51:49PM -0500, ryan wrote:
On 04/28/2017 05:56 PM, Bill Kontos wrote:
Out of curiosity has anyone ever attempted to prototype a
hardware block
based on evolution principles? Doing it on an fpga is probably a bad idea since we wont be able to implement the results in more copies but this could potentially also happen in a software simulation where the input and output interfaces of the hardware block are pre defined
<snip>
I suspect that without having the feature of it being an instruction set that only works on that one chip due to it exploiting the quirks of the chip, some efficiency would be lost.
I'm imagining a system where traditional silicone grooms many FPGAs, each with a dedicated task, and the system is provided with some known-good instruction sets that work, but only slowly. So then either the OEM or the user sets up their fancy new system, and one of the steps is to plug it in and run a setup program for anywhere from a few hours to a couple of days which iterates the instructions to improve efficiency, then they can begin to use their system.
As for using this method in a software simulation, I wouldn't be surprised if some chip manufacturers already do that for certain sections of the chip, even if its only during the early design faze. I would imagine the software guiding the evolution could be instructed to cull anything that isn't working with binary, thus allowing human engineers/programmers to more easily reverse engineer the instruction set and further edit it.
Let me remind you of a real-world situation. The hardware designers were woring on the second version of their successful CPU. They attached some counters to masure hos many times the varioun instructons were being executed. THey discovered that the most common instructons were certain test and brnch instructions. So they worked hard on making sure the next model had the most efficient implementation of those test and branch instructions they could achieve.
But when they finally put the new machine together and tried it out, they foud no improvement at all.
Investigating, they discovered they had optimized the wait loop.
-- hendrik