Towards Memory Specialization: a Case for Long-Term and Short-Term RAM

Device physics-wise, you could probably make SRAM faster by dropping the transistor threshold voltage. It would also make it harder / slower to write. The bigger downside is that it would have higher leakage power, but if it's a small portion of all the SRAM, it might be worth the tradeoff.

For DRAM, there isn't as much "device" involved because the storage element isn't transistor-based. You could probably make some design tradeoff in the sense amplifier to reduce read times by trading off write times, but I doubt it would make a significant change.

And lower power usage. Datacenters and mobile devices will always want that.

Sometimes very yes?

If you've got 1GB of weights, those are coming through the caches on their way to execution unit somehow.

Many caches are smart enough to recognize these accesses as a strided, streaming, heavily prefetchable, evictable read, and optimize for that.

Many models are now quantized too to reduce the overall the overall memory bandwidth needed for execution, which also helps with caching.

> "The key insight motivating LtRAM is that long data lifetimes and read heavy access patterns allow optimizations that are unsuitable for general purpose memories. Primary applications include model weights in ML inference, code pages, hot instruction paths, and relatively static data pages—workloads that can tolerate higher write costs in exchange for lower read energy and improved cost per bit. This specialization addresses fundamental mismatches in current systems where read intensive data competes for the same resources as frequently modified data."

Essentially I guess they're calling for more specific hardware for LLM tasks, much like was done with all the networking equipment for dedicated packet processing with specialized SRAM/DRAM/TCAM tiers to keep latency to a minimum.

While there's an obvious need for this for traffic flow across the internet, whether or not LLMs are really going to scale like that, or there's a massive AI/LLM bubble about to pop, would be the practical issue, and who knows? The tea leaves are unclear.

The same could be said for, say, SIMD/vectorization, which 99% of ordinary application code has no use for, but it quietly provides big performance benefits whenever you resample an image, or use a media codec, or display 3D graphics, or run a small AI model on the CPU, etc. There are lots of performance microfeatures like this that may or may not be worth it to include in a system, but just because they are only useful in certain very specific cases does not mean they should be dismissed out of hand. Sometimes the juice is worth the squeeze (and sometimes not, but you can't know for sure unless you put it out into the world and see if people use it).

If you’re not in manual memory management land, then you probably don’t care about this optimization just like you barely think of stack vs heap. Maybe the compiler could guess something for you, but I wouldn’t be worrying about it in that problem space.

I'm not sure how good applications are at properly annotating it, but for most applications assets are also effectively read only.

You don't even need most of the ram usage to be able to take advantage of this. If you can reasonably predict what portion of ram usage will be heuristically read-heavy, then you can allocate your ram budget accordingly, and probably eak out a measurable performance improvement. In a world with Moore's law, this type of heterogeneous architecture has proven to not really be worth it. However that calculus chagnes once we lose the ability to throw more transistors at the problem.