Hacker News Reader

Comment by ianai | Hacker News Reader

Spends a lot fewer cpu cycles waiting for data. Same is true with how much faster the main storage moves. Won’t be enough for all workflows (hence it’s worth the OPs question) but will go further than expected.

Edit-latest i9s seem to get around 90GB/s to ram while they promise the M2 ultra 800GB/s. So quarter second to access all of the ram?

Edit2-the ssd is higher end nvme speeds.

View on HN

Bandwidth doesn't tell us about how many clock cycles it's waiting for the data though (latency), does it?

For example with paging, I don't think those bandwidths make much difference and it's instead the latency which matters. Or am I wrong?

What’s going to have the faster latency? Memory chiplets on the same substrate as the cpu cores or memory dimms sitting in a housing a few inches away?

I’m sure there’s more than a few factors making latency faster for those chiplets.

https://www.anandtech.com/show/17431/apple-announces-m2-soc-...

Edit-but I don’t think this question is about latency exactly. If you’re editing a large file that fits in 196gb ram all at once then those edits should benefit by apparently an order of magnitude higher bandwidth. It’s going to be able to address all of its ram several times per second while the i9 will take much longer (roughly). That and similar speedups from the whole systems engineering should make up some for not having more ram. Think how 8 gig of ram in the m1 was widely reported to be more usable than expected when announced.

Memory chiplets on the same substrate as the cpu cores or memory dimms sitting in a housing a few inches away?

Where are you getting this information on the memory latency? What is the actual number?

I don’t think this question is about latency exactly.

You were the one the brought up latency

those edits should benefit by apparently an order of magnitude higher bandwidth

Most software is not written well enough to be memory bandwidth limited, even graphics software. It is mostly games that end up actually needing the memory bandwidth.

Reply on news.ycombinator.com