For now, modern microprocessors are equipped with FMA (Fused multiply-add) instructions and other type of accelerators to perform much of the computation in one go. Still, the faster the processors, the greater the need for faster memory access. HBM2 memory and simply bigger L3 cache are some of the answers to this ongoing challenge.
@PaulSpades3 жыл бұрын
I really don't see why analogue storage helps with computing in memory versus binary storage. The whole thing starts with the premise "we're developing resistive rom and this drawing shows we can store values in it". Great, we can store values in lots of types of memory, getting the computing power local to the memory blocks with a reasonable architecture is the issue.
@t3chm0nkey2 жыл бұрын
From what I've been reading. It's better because for example you only need 1 analog bit to represent 0.01 versus multiple digital bits to represent the "0" the "." the "0" and the "1". I also adding and subtracting will be faster because again, you can just combine the voltage of two analog 0.01v + 0.01v = 0.02v. with digital you have to go back and forth with the CPU to do the math to make that happen.
@SR_M0L1NA2 жыл бұрын
🤔 estoy pensando en quizás sea posible sustituir esas resistencias por potenciometros digitales en circuitos de prototipado rápido para expirmentar con ellos, cómo un híbrido entre una fpga con resistencias, o incluir ya la parte del circuto prediseñado: con los transistores y éstos a las resistencias. El resto cómo la memoria, demux, D/A, A/D, etc el circuito se diseñaría por el ingeniero o estudiante.
@SR_M0L1NA2 жыл бұрын
Un potenciometro digital tiene normalmente de 5 a 8 bits que controlan de 31 a 255 resistencias respectivemente, es decir un 0.3922% de resolución para 8 bits, y los potenciometros analógicos ofrecen un 2%. De esta manera de pueden diseñar sistemas de IA para uso doméstico o pequeños usos a un coste bajísimo. Si se hacen de 10 bits se controlarían 1023 resistencias que tendrían una resolución del 0.0978%.