I'm not sure that memory

and I'm speaking more-broadly than HBM, even

optimized for running neural nets on parallel compute hardware and memory optimized for conventional CPUs overlap all that much. I think that, setting aside the HBM question, that if we long-term wind up with dedicated parallel compute hardware running neural nets, that we may very well wind up with different sorts of memory optimized for different things.

So, if you're running neural nets, you have extremely predictable access patterns. Software could tell you what its next 10 GB of accesses to the neural net are going to be. That means that latency is basically a total non-factor for neural net memory, because the software can request it in huge batches and do other things in the meantime.

That's not the case for a lot of the memory used for, say, playing video games. Part of the reason, aside from hardware vendors using price discrimination that PCs (as opposed to servers) don't use registered memory (which makes it easier to handle more memory) is because it increases latency a little bit, which is bad when you're running software where you don't know what memory you're going to need next and have a critical path that relies on that memory.

On the other hand, parallel compute hardware doing neural nets are extremely sensitive to bandwidth. They want as much as they can possibly get, and that's where the bottleneck is today for them. Back on your home computer, a lot of software is oriented around doing operations in serial, and that's more prone to not saturate the memory bus.

I'd bet that neural net parallel compute hardware does way more reading than writing of memory, because edge weights don't change at runtime (on current models! That could change!).

searches

Yeah.

https://arxiv.org/html/2501.09605v1

AI clusters today are one of the major uses of High Bandwidth Memory (HBM). However, HBM is suboptimal for AI workloads for several reasons. Analysis shows HBM is overprovisioned on write performance, but underprovisioned on density and read bandwidth, and also has significant energy per bit overheads. It is also expensive, with lower yield than DRAM due to manufacturing complexity.

But there are probably a lot of workloads where your CPU wants to do a ton of writes.

I'd bet that cache coherency isn't a huge issue for neural net parallel compute hardware, because it's going to be a while until any value computed by one part of the hardware is needed again, until we reach the point where we can parallel-compute an entire layer at one go (which...I suppose we could theoretically do. Someone just posted something that I commented on about someone making an ASIC with Llama edge weights hard-coded into the silicon, which probably is a step in that direction). But with CPUs, a big problem is making sure that a value written by one CPU core reaches another CPU core, that the second doesn't use a stale value. That's gonna impact the kind of memory controller design that's optimal.