For decades, computer and smartphone circuits have steadily become smaller and more powerful, following the trend known as Moore’s Law. However, this era of consistent progress is nearing its end due to physical limits, such as the maximum number of transistors that can fit on a chip and the heat generated by densely packed components. As a result, the pace of performance improvements is slowing, even as the demand for computational power continues to grow with data-intensive technologies like artificial intelligence and machine learning.

To overcome these challenges, innovative solutions are required. One promising approach lies in photonics, which uses light instead of electricity to process information. Photonics offers significant advantages, including lower energy consumption and faster data transmission with reduced latency.

One of the most promising approaches is in-memory computing, which requires the use of photonic memories. Passing light signals through these memories makes it possible to perform operations nearly instantaneously. But solutions proposed for creating such memories have faced challenges such as low switching speeds and limited programmability.

top 6 comments

sorted by: hot top controversial new old

[–] limer@lemmy.dbzer0.com 4 points 1 month ago

I remember when these were heavily discussed in the 1980s; perhaps it will be ready one day.

Some tech always seems out of reach yet possible soon

[–] just_another_person@lemmy.world 3 points 1 month ago (4 children)

It's kind of a dumb angle though. If it ALL doesn't work at light speed, then none of it does. It's all bottlenecks up to the delivery. It's going to be exactly as slow as the slowest components.

[–] NaibofTabr@infosec.pub 13 points 1 month ago* (last edited 1 month ago)

Hmm, every computing system is a collection of bottlenecks... in most desktops the CPU has a dedicated bus for the RAM because any other device in that path would slow down the communication with the RAM.

The point is, bottlenecks can be designed around. Making the memory component faster makes it worthwhile to double or triple the memory bus bandwidth, or just reduce the amount of memory in the system while keeping the same level of functionality. And the slower components can be segregated out to their own communication paths (that's what all the different pins on the bottom of the CPU are for).

Usually the hardest part is getting the software to use the hardware properly. We've had consumer multicore processors for 2 decades now but most applications still don't do parallel processing efficiently. Hell, a lot of them are still 32 bit and can't use the 64 bit memory address space.

tl;dr: hardware guys are genius wizards, software developers mostly suck

[–] Hamartiogonic@sopuli.xyz 4 points 1 month ago* (last edited 1 month ago)

Sounds like in this case you need to switch a magnetic field on and off. That could be the bottle neck.

[–] catloaf@lemm.ee 4 points 1 month ago

And if you're doing a bunch of work on one chip, then yes it'll be much faster. Not every operation takes the whole path from disk to screen.

[–] Boomkop3@reddthat.com 2 points 1 month ago

Someone didn't read the article