Yes.

Actually, it's everything: https://en.wikipedia.org/wiki/Random-access_memory#Memory_wall

...But also, your friend is wrong. And its complicated.

First some background. There are a couple of "regular" types of RAM basically every product uses under the hood. And every one is a set of tradeoffs: bandwidth, latency, capacity, power efficiency, trace lengths, expense, and so on. A few of types:

• "Regular" DDR. This is what you see on sticks for desktops, laptops, and servers.

• LPDDR. Its power-optimized at the expense of some latency and trace length, but its specs are otherwise similar to DDR. This is what smartphones (and Macs) use.

• GDDR. Optimized for bandwidth and cost over everything. Guzzles power, terrible latency, cheap. Used on gaming GPUs.

• HBM. Low speed, requires extremely short trace lengths, but is extremely low power and low latency. Uses an exotic bus with exotic manufacturing techniques. VERY expensive. This is what you see on more exotic server accelerators.

But the packaging is also really, really important too. For example, regular DDR5 can come in:

• Unregistered RAM sticks, for your desktop.

• As registered RAM, for servers.

• As SODIMM, for laptops.

• Soldered onto laptop or mini PC motherboards.

And these all have long traces, since they have to come in sticks. They're "far" from the CPU. But laptop SODIMMs, in particular, have a poor electrical path to the CPU, and start to run into huge issues at DDR5 speeds.

...In other words, laptop DDR5 SODIMMs are particularly slow, because the old socket they sit in simply cannot keep up with their speeds.

As for LPDDR5: its basically a modification of DDR5. Its specification assumes a shorter, higher quality trace or "wire" to the CPU, hence it can only come soldered or in CAMM modules, and sitting right next to the CPU:

But electrically, it's almost the same as DDR5! It's not wider-and-slower like HBM, or more tweaked like GDDR: it's just regular, cheap DDR5, massaged to work better soldered to the motherboard or in a smartphone CPU package.

This is what Macs use.

They're not using "special Apple RAM," as is the popular perception. They're using smartphone RAM. Its a bit faster because it is "closer" to the CPU, but this is misleading, as its nothing like more exotic standards that really take advantage of that proximity. If you read this article, its closer to GDDR5 than HBM: https://graphicscardhub.com/gddr5-vs-gddr5x-vs-hbm-vs-hbm2/

In practice, Apple RAM is faster and "closer" to the CPU, but its more because of business design choices than a particularly exclusive memory specification:

• Apple uses very wide buses for their CPUs. Typical laptops are 64-bit or 128-bit, Apple stuff can be 192-bit to 512-bit. This is more expensive, and you pay for it in the sticker price.

• Regular laptops use swappable SODIMM sticks, which have a tremendous speed cost. Apple doesn't have to worry about this.

• Apple pays a premium for fast, low latency, binned LPDDR5. Regular OEMs do not.

But it doesn't have to be this way.

Some laptops already use swappable LPCAMM modules, which for all practical purposes are just as fast as Apple's RAM, and just as "close to the CPU" from an electrical perspective:

Framework was very close to implementing this for AMD's 256-bit Strix Halo CPUs. They couldn't quite work out the electrical gremlins before release. But if they had worked it out, its memory interface would basically be the same as an M-Pro CPU, with swappable RAM.

And I'm specifically avoiding technical theoretical talk because... well, it doesn't really matter here. We're not talking about clocks and wire transmission physics and protocols for fantasy CPUs, we're talking about commercial RAM that all CPU makers actually use.

TL;DR

Apple RAM is "closer to the CPU" in the existing market, but there's no technical reason PCs have to be any different.

Apple uses off-the-shelf smartphone RAM, basically. It's not that special.

An AMD Strix Halo laptop, with 2 LPCAMM modules, would be just as fast as Apple memory, and just as "close" to the CPU. Some PCs already do this on a smaller scale.

So strictly speaking, your friend is correct. But its misleading. All that really matters is the memory specification laptop/desktop manufacturers choose.

Yes.

If you really want to see some interesting data then search data eye. It's basically an infinity. Yet the exact eye itself is the dual peak and valley (DDR transmission on rise and fall off the signal).

This is why memory is positioned as close to the CPUs possible because the longer those traces are the more latency that's involved but ultimately in memory even though we've gone to much higher bandwidth latency has remained pretty similar for the last several decades.

So yeah that's why you often see the CPU ... and the ram is right next to it on the motherboard! And if you go better than that you might get some high bandwidth memory on the chip itself otherwise known as cache. Which level 1 caches can exceed 1 TB/s. Most RAM you're looking at 10s of GB/s. There's even some that can certainly hit around 100 GB/s. But that's usually the more specialty gddr. The data eyes on quad pumped memory are quite beautiful. Kind of like fractals.

The newer M-series have the RAM soldered on the board instead of removable. Typically that means they have a higher bandwidth to the RAM because of it, but you can't upgrade to more memory. It's an advantage if you're running LLMs on the unit itself, but not much use otherwise.