Hacker News Reader

Comment by Applejinx | Hacker News Reader

Rules change. I've got one of the 16GB RAM Macbook Pros. They've got it acting like it's L2 cache instead of RAM as we know it, so the place where you need more than 128G becomes, well, the SSD, your mass storage.

These things already don't act as if it's RAM as we're used to considering it. They could probably have shipped the first ones with 4G RAM and had 'em act relatively normal… with heavy stuff like 4K video editing. Not great, but normal. So they did 8G RAM to have 'em act more impressive… like doing said editing smoother with fewer dropped frames than far more powerful Intel machines with a lot more RAM.

View on HN

Do you have any evidence for this claim? The actual concrete performance numbers for the M1 and Zen3 don't agree with you at all:

M1: https://images.anandtech.com/doci/16252/latency-m1_575px.png

Zen3: https://images.anandtech.com/doci/16214/lat-5950.png

If anything the M1's ram has _higher_ latency than the Zen3.

(Can't reply to the sibling comment but Intel's numbers are even better: https://images.anandtech.com/doci/16214/lat-10900.png)

> like doing said editing smoother with fewer dropped frames than far more powerful Intel machines with a lot more RAM

Just noting Zen3 is not Intel.

Well exactly how are you defining latency?

Is it random accesses you want to test? Or TLB thrashing?

The R per RV prange benchmark accesses memory randomly, but in a TLB friendly pattern (TLB thrashing real world workloads aren't particularly common). The M1 gets 30ns, the 5960X is 55ns and still climbing at the max array size.

The 5950X does better on "full" random, although that's unclear if that's a faster TLB, or the testing with a different page size. With Linux for TLB thrashing workloads you can always use HUGE pages, not sure if the M1 hardware supports that though.

>They've got it acting like it's L2 cache instead of RAM....... These things already don't act as if it's RAM as we're used to considering it.

Please stop this non-sense on HN. DRAM is DRAM. L2 Cache is L2 Cache. Doesn't matter how good its 8-Channel LPDDR4X Memory Architecture sitting next to the Die in the same package it is still just DRAM.

As someone that runs Kubernetes on my work laptop, all the "the memory architecture is different, so 8GB is enough" nonsense just doesn't hold up. I need a ton of memory FULL STOP. It sounds just like the nonsense Apple was spewing about MMS messaging on the original iPhone.

>I need a ton of memory FULL STOP.

Then buy something else, full stop.

You might be confused. I wasn't asking for purchasing advice. I'm pointing out Apple's history of downplaying something that they have not yet implemented.

Apple is not selling MB Airs or 13" Pros for users who are running massive workloads like Kubernetes. It is the reason they left intel models in the current line up. The M1 models are for more pedestrian users and users with moderate video editing workloads.

It takes my M1 a little while longer to stitch together 4K videos than my 2017 MPB 16". My 2017 MBP screams and throttles down the CPU down to 60%-75% as soon as editing starts.

Yeah, I heard this too and nope - the M1 is not some magical RAM negating beast. If you actually need more RAM you still need more RAM even with the M1. I ended up returning mine (with much difficulty - me not really wanting to; Apple was more than willing to take it back) because if you need more than 16 you still need more than 16.

If I have 20GB of data and 16GB of RAM, it's simply never going to fit and will have to page to disk. That page will always be slower than RAM.

Even if the RAM were as fast as L1 cache, it still wouldn't do a thing for the SSD bottleneck.

Reply on news.ycombinator.com