I want a Macbook Pro with an M5 processor but the RAM kills you. It's integrated and you can't upgrade yourself so Apple charges an arm and a leg for anything beyond 16GB
There’s a couple things to understand about modern Mac’s - that is any Mac starting with the M1 in 2020.
Apple’s chips differ from X86 in a few notable ways. RAM aka Memory is one of the most important. Historically one of the most immediate bottlenecks that you will notice or “feel” when using a computer is if it begins to run low on memory.
The OS must begin to use tricks like paging out some of the in-memory data to your disk. In the days before SSD when this meant spinning platters of your hard drive, this performance hit was enormous. It could slow to a crawl and literally “freeze up” until it was able to complete the operation.
SSD’s have closed the gap an incredible amount due to their massive jumps in both latency and overall throughput. Paging out to SSD is much “smoother” overall, but can still cause hiccups, hangs, delays etc.
As fast as SSD are, you still want to avoid running out of working memory. In fact all modern OS’s are fairly aggressive with pre-caching data INTO memory. It tries to figure out what data might be needed in the near term and, assuming there is free memory available, will load that data into memory before the CPU or GPU calls for it.
Which brings up the next important factor here, which is the GPU. For a very long time now, Graphics Processing Units have been their own dedicated stand-alone processor, connected over a data bus that talks to the CPU.
Yes, there have been “integrated graphics” from both Intel and AMD that are part of the CPU but these (in general) have been much slower than having a stand alone graphics card, for a number of reasons which we won’t really get too much into here for the sake of brevity (oops too late).
In any case the take away here is that dedicated GPU cards are very (VERY) fast on their own. Once you feed the data to them, they can process it extremely quickly. BUT they are sort of an island, in that they must then turn around and send that data back to the CPU. That data bus (PCI-E) is thoroughly trounced in speed compared to how fast the card can compute.
Oh ho ho you might say, but I can play my favorite game at 100+ fps at 4K! How slow can it be! Well if games are your only concern, then It doesn’t really matter as game developers have come up with all sorts of tricks to make the best out of this less-than-ideal scenario of having very fast graphics processing capabilities saddled to a pitifully slow bus.
Numbers time just to help put this in some sort of context. modern GPU’s have internal bandwidth of hundreds of gigs/sec. The top end workstation cards are approaching or even over 1000 GB/sec
Meanwhile as soon as you process the data and need to send it back over the PCI-E 4.0 bus you are talking about a 32 GB/sec link. oof.
PCI-E 5.0 doubles that to 64 GB/sec but still a far cry from even the 300-500 GB/sec that a mid range GPU is operating at, nevermind if you have a high-end card.
From the pcie link it gets routed to the memory controllers which then sends the data to the socketed DDR5 dimms. Without delving too much into technical details here, a typical dual-channel setup will yield around 80-90 GB/sec of main system memory bandwidth.
Faster than the PCI-E link, but still a bottleneck of its own. So if you have a workflow that involves heavy graphics compute (video processing for example) then you are in a situation where you are constantly shuttling data back and forth over these two slow links (double whammy!)
Coming around to the point now I promise. So, Apple’s approach is to say you know what, we are going to build the GPU cores directly into the CPU, and then place the memory chips on that same package.
There is no separate pool of video memory plus the system memory that traditionally would have to spend all its time transferring data from one to the other. There is only 1 single pool of unified memory that the CPU+GPU can both access directly,
Beyond this change in underlying architecture is the memory bandwidth, which starts at 120GB/sec for the base M4 and roughly doubles (actually better than double) with each tier of chip (Pro, Max) topping out at 546 GB/sec.
So the end result is a very fast, responsive system because it’s processing everything in a much more efficient manner without needless copying everything over slow bus links.
As has been true since the dawn of computing, more RAM is often the single best upgrade you can make. Yes Apple charges what they charge to upgrade the memory, but given the typical lifespan of the machines it is not outrageous as it may seem on its face. and you might just be surprised how smooth “only” 16GB can be.
