Trinary computation

I trinary processing computation a thing? can ai and modern design tools rapidly accelerate the configuration and developmental roll-out?

are you talking about this?

she mentions at the end that noise is still a problem with it… well… and every other aspect of computing isn’t made for it.

But it is still interesting.

1 Like

I think you mean ternary computing? It’s been around for a long time - some of the early computer predecessors used ternary (mapping onto +'ve, 0 and -'ve voltage respectively.

Ternary computing has some advantages in terms of the theory (easier to represent some logic, and as base-3 is closer to base-e i.e. natural logarithm than base-2, it also represents numbers more efficiently). The problem is that we got really good at making binary transistors.

We might go back to Ternary with other types of systems though - if we ever get photonic computing to work, ternary would be a good fit.

4 Likes

As dgdg said, Ternary is VERY old, from the start of the computer days when they took up whole rooms. Manufacturing just favored binary with semiconductors though and there was no need at all for 3 state computing then so why bother?

Now days though with AI computation using mostly matrix arrays I bet we would find a vastly improved performance and energy efficiency for those types of computation with using a ternary computation system. It wouldnt surprise me at all if the hardware blocks of cores we currently use for these types of computations simply became a block of ternary compute cluster inside a standard CPU and the scheduler worked like normal assigning tasks that work best in 2d to standard int and FP cores while workloads best suited for 3 state computation get assigned to that section of the CPU. We have gone through evolution like this plenty of times in the past and it always becomes a hardware block in existing systems rather than overhauling the entire computer system from scratch. Backwards compatibility is king.

It’s interesting that you mention AI use cases, because Microsoft did release their “1-bit” LLM a while back, which is actually a 1-trit LLM. They found it was more energy efficient even when running on a binary CPU.

Link: Microsoft’s “1‑bit” AI model runs on a CPU only, while matching larger systems - Ars Technica

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.