> Apple using a 512-bit memory bus on its Max processors is indeed the future of APUs if you ask me.
It's very expensive to have a bus that wide, which is why it's so rarely done. Desktop GPUs have done it in the past ( https://www.techpowerup.com/gpu-specs/?buswidth=512%20bit&so... ), but they all keep pulling back from it because it's too expensive.
Apple can do it because they can just pay for it and know they can charge for it, they aren't really competing with anyone. But the M3 Max is also a stonking huge chip - at 92bn transistors it's significantly bigger than the RTX 4090 (76bn transistors). Was a 512-bit bus really a good use of those transistors? Probably not. Will others do it? Probably also no, they need to be more efficient on silicon usage. Especially as node shrinks provide less & less benefit yet cost increasingly more.
512-bit is probably a bit extreme, but I can see 192-bit and 256-bit becoming more popular. At the end of the day, if you have a high-end APU, having a 128-bit bus is probably THE bottleneck to performance. It's not clear to me that it makes sense or costs any less to have two 128-bit buses on two different chips which you see on a lot of gaming laptops instead of a single 256-bit bus on one chip for the midrange market.
M3 pro only used a mere 37 billion transistors with a 192-bit bus, so you can get wider than 128-bit while being economical about it. I'd love for there to be a 512-bit Strix Halo but it probably won't happen, it probably does not make business sense.
I don't know if the comparison to GPUs necessarily tracks here because the price floor of having 8 chips of GDDR is a lot higher than having 8 chips of DDR.
Sure, which is why you see other midrange APUs use >128-bit buses, eg PS4 & PS5 are both 256-bit buses w/ GDDR.
Similarly, ultra high end APUs like the Xeon Max are doing on-package HBM for truly monstrous memory bandwidth numbers.
The thing is that nobody else is really trying to make anything like an M3 Pro, and I don't know if they even will. It's a really weird product. Big dies are expensive, hence everyone pushing towards chiplets. A really great spot to split up dies is between compute units that are largely independent - which the CPU & GPU actually are. There's a few workloads where unified memory helps, but most don't. So splitting those apart makes a ton of sense still. Then also if you push them hard to squeeze out all the performance you can, they both get very hot - so you want them physically far apart for cooling reasons. At which point you might as well just give them their own memory which can also be further specialized for their respective needs as one is latency biased and the other bandwidth biased. And now you're just back to traditional desktop architecture - it still makes just way too much sense for the high end.
It makes sense for Apple since they focus almost exclusively on laptops and you get exactly the single mix of CPU & GPU they decide, just like it does for consoles where again they have a midrange power budget & a single CPU/GPU configuration over millions of units. But as soon as different product specializations show up and different workload demands are catered to, coupling the CPU & GPU like that kinda doesn't make sense?
>The thing is that nobody else is really trying to make anything like an M3 Pro
Well, AMD is indeed making their own M3 Pro, the Strix Halo. Which is going to achieve 4050/4060 performance in laptops that will again be cheaper and more power efficient for a lack of a Nvidia GPU. This is also a net addition to their lineup.
AMD is a minority player in the laptop market, so it doesn't necessarily make sense for them to compete hard against Intel at every price segment, and they have clear competitive advantage and counter-positioning here against Intel/Nvidia/Apple.
It's very expensive to have a bus that wide, which is why it's so rarely done. Desktop GPUs have done it in the past ( https://www.techpowerup.com/gpu-specs/?buswidth=512%20bit&so... ), but they all keep pulling back from it because it's too expensive.
Apple can do it because they can just pay for it and know they can charge for it, they aren't really competing with anyone. But the M3 Max is also a stonking huge chip - at 92bn transistors it's significantly bigger than the RTX 4090 (76bn transistors). Was a 512-bit bus really a good use of those transistors? Probably not. Will others do it? Probably also no, they need to be more efficient on silicon usage. Especially as node shrinks provide less & less benefit yet cost increasingly more.