MacBook Neo Deep Dive: Benchmarks, Wafer Economics, and the 8GB Gamble

14 points by makishimu


david_chisnall

I'm not sure I agree with this:

For the tasks this machine is built for (web browsing, documents, streaming, light photo editing), single-core is what matters. The Neo will feel snappy.

Pretty much all of these tasks are parallelised in modern software, and some of them are mostly run on the GPU.

The memory controller is a genuine constraint. The A18 Pro was designed for the iPhone 16 Pro, which has always shipped with 8GB. The LPDDR5x controller is configured for that package. Upgrading to 16GB would require different memory packaging and PCB routing. Apple could have designed around this, but chose not to for a first-generation budget product.

This is interesting. A lot of the performance benefits Apple gets come from sizing everything to avoid bottlenecks. This was the big lesson that they learned from the PowerPC G4: on paper, it massively outperformed contemporary Intel chips, in practice this happened only for workloads that fitted in the cache: it was almost impossible to keep the vector units fed. This applies all of the way up and down the scales. Memory bandwidth (number of memory controllers), cache topology, store-queue sizes, pipeline length and widths, and so on all contribute and it's very easy to make one of them a bottleneck so none of the others can reach their peak performance.

The issue with the A18, I suspect, is that all of this tuning was done with iPhone workloads. These have quite different characteristics to non-full-screen macOS workloads and I wonder where we're going to see limitations.

I do wonder about all of the people who keep saying that 8 GiB is a small amount. My previous MacBook Pro had 16 GiB, but I normally had 4 GiB allocated to a VM and routinely did LLVM builds on it. The older Mac where I did a lot of video editing had 3 GiB. Looking at Activity Monitor, most of the apps I'm using (including MS Office) are using under 0.5 GiBs. Unless you're running a lot of apps in parallel (and actively using them, so Sudden Termination doesn't quietly kill them in the background), I'd be quite surprised if macOS didn't work pretty well in 8 GiB, especially with NVMe storage where the performance benefits of a large buffer cache are smaller.

benton

10 minutes and a little thermal pad and you get better sustained performance as long as you are okay with the case getting much warmer.

kghose

Is this article independent of anandtech and apple?