4 ms·
The cost of routing on a chip has been climbing for decades. If we decreased the cost of routing, the demand for transistors would go up again? satisfies #1 Ma
by mikewarot 5y ago
The cost of routing on a chip has been climbing for decades. If we decreased the cost of routing, the demand for transistors would go up again? satisfies #1
Making the chip less complex, satisfies #2
What if you had a chip with no routing? Just an array of Look Up Tables? (LUTs) 4x4 bits, one bit coming from each cartesian neighbor, and one going to each, clocked in the manner of checkerboard colors to prevent race conditions.
It sounds quite silly... but it has a lot of advantages, at the cost of using the LUTs to route bits in the worst case.
move signals around to get to any I/O pin
rotate/flip the placement of a software "module"
route around a bad LUT
build a secure box around any function using a wall of LUTS
compute with LUTs on data going all directions simultaneously
easily debug by inserting probes dynamically, and re-routing around them
all the gates run all the time, unlike the transistors in RAM, better utilization of silicon
make it possible to use proofs against deployed programs in the LUTs
- eternityforest 5y agoSounds like a lot of customization. You've basically invented the FPGA, and they're mostly good for hardware accelerating specific things. A CPU is going to be better if you don't have time to convert your problem into a hardware description language. If you do have that time and expertise you can probably get good results on a regular GPU. CPUs are just way easier to program for than FPGAs. Maybe with better tooling.