Do we even know the timings of the LPDDR5 RAM in Apple Silicon computers? Apple uses 6400MT/s RAM in the M3 generation. That speed is possible with memory modules, too. And everything else, like CAS-latency depends on the chips.
PCs can run RAM in modules at 6400MT/s and faster speeds.
That's completely irrelevant.
Let's get another thing out of the way, UMA is, as you correctly point out, about faster data sharing between CPU and GPU. It has nothing to do with how the RAM is physically attached. Those are two independent questions. They are on orthogonal axis of the design space.
That was in a separate section, it also covered Apple’s particular implementation of UMA, as well as covering Nividia’s use. It’s a broad article on UMA, but it does particularly talk about Apple’s UMA, and I linked to another article as well, which agrees on the basics of Apple’s UMA implementation. And yes I did read both articles, and clearly saw that it had separate sections where it talked about different implementations of the concept. I didn’t say that everything in the article was relevant to Apple’s implementation of UMA. Someone asked what the difference between UMA and normal RAM was, so I provided them with an article that included different implementations of UMA and how they differ from standard RAM. It wasn’t even meant as an argument for or against UMA, it was meant to be helpful for someone who was asking a question.
How is that completely irrelevant? You have this pattern of just calling things irrelevant, but someone literally asked “why does proximity matter?” And so I answered that closer proximity decreases latency. That’s the reason generally cited for close proximity.
Well, one "Key Point" right at the beginning of the article is "Unified memory is a modern memory technology that combines RAM and a hard drive into a single memory pool. This newer approach, favored by Apple Inc., is thought to improve computing efficiency and energy consumption, usually coming at a premium." That's a direct quote. You are telling me that this is a good description of Apple's UMA?
If both memory on a module and memory soldered to the SoC can run at the same frequency that is a clear indication for the decreased latency not being relevant. If it were relevant then the soldered RAM could run at a higher frequency. As it stands, you can get PCs with memory modules that run the RAM at higher frequencies than any Apple Silicon chip. Perhaps it's possible to run soldered RAM faster, but that observation is, at this point, irrelevant, because it isn't running faster.
The RAM being part of the M chips reduces latency, hence allowing faster data sharing between CPU and GPU.
We've been over this, it runs at the same frequency that memory modules easily achieve, too. Hence the minutely reduced latency does not make an appearance in the actual performance of the soldered memory of an Apple Silicon computer.
How much power does it take to increase the distance and up the speed? And can you point us to an Intel/AMD that operates its memory at those speeds while on battery?
Probably true in the case of that particular article. Other articles which may support your point (assuming that you don't think Apple's Unified RAM keeps its content when powered off) are available - that link was just a bad choice (although there's a real dearth of good links).
www.macrumors.com
Sorry for my bad article choice, I read through it and it was late here and didn’t catch those things, but looking back over it, you’re right, it’s probably a bad article choice. But there are other articles like the second one I shared with you that make my point better.
You downvoted me for pointing out that hard drives have no place in a discussion of Apple's UMA. Not sure what to think.
I guess looking at the article it does seem to mishmash some things, sorry for my article choice, it was late here and I read through it and didn’t notice that on first read through. But as you pointed out, there are other articles that make the point I was trying to make. And as to your point about the RAM being built into the M chip or not, what I was meaning and saying is it’s built into the package. It’s not etched on the same wafer, I know it’s surface mounted, but it’s part of the same package, that’s what I was talking about. 👍🏻Probably true in the case of that particular article. Other articles which may support your point (assuming that you don't think Apple's Unified RAM keeps its content when powered off) are available - that link was just a bad choice (although there's a real dearth of good links).
Examples of questionable/mangled statements in that article (none of which are really being disputed in this thread):
In other words, the linked article is a mish-mash of factoids that confuse Unified Memory/RAM, flash memory, Hybrid Storage and Hierarchical/Tiered storage - which would be entirely consistent with a LLM generating text based on an ambiguous brief like "unified memory". Unfortunately, this is the new reality - people really are using AI to bulk out websites and you have to check every article like a hawk to see that its not superficially plausible AI-generated gibberish.
As for what you are actually saying: just to be clear, the actual RAM in UMA is not built into the CPU die itself, it is in the form of conventional LPDDR memory modules surface-mount soldered to the CPU package - you can see that from the link below which shows the RAM can be upgraded (see below) for a given value of "can" (its not really a practical prospect)
M1 Mac RAM and SSD Upgrades Found to Be Possible After Purchase
Technicians in China have reportedly succeeded in upgrading the memory and storage of the M1 chip, suggesting that Apple's integrated custom...
...the main point of UMA is that both the CPU and the integrated GPU share the RAM & that this is implemented differently and more efficiently than older systems with integrated graphics that allocated a chunk of system RAM as video RAM, so the CPU doesn't have to waste time copying data between RAM and VRAM. Mounting the chips directly on the package may make things a bit faster - or, more likely, saves power. The physical RAM, though, is the exact same stuff that you get on ultraportable PCs that use LPDDR5x (on which it is still soldered direct to the logic board as close to the CPU as possible - a new "modular LPDDR" system is in the pipeline but won't be in the shops for Xmas).
I edited that response because I looked back at the article, and it does in fact say those things. Sorry for my poor choice of article, as I explained, it was late here, and I didn’t catch those things on first read-through. I shouldn’t have dismissed you as quickly as I did, and I shouldn’t have met your valid criticisms of the article with the sarcasm and impugning of your motives. I apologize for my original response, it wasn’t right of me.
Greater distance increases latency, and power required to send high speed signals to the memory. Shorter distance makes it more energy efficient and reduces latency of the connection. Look it up, this is the number one reason cited for locating UMA as close to the CPU and GPU as possible.
Let's try to come to an end here. The latency is reduced, but that doesn't translate to higher speed (as shown by the fact that the frequency isn't higher than for memory modules). Being more energy efficient sounds plausible. But the UMA architecture is independent of soldered RAM, from a performance perspective nothing would change if UMA were combined with memory modules. Finally here's a good reason for Apple to use soldered RAM that nobody came up with yet, at least that I saw: The Max chips have a 512bit memory interface. That's not easy to achieve with CAMM modules for a laptop! Some server chips use wider memory with modules, but the chips and the mainboards are massive.
Yea, the wider memory bus is certainly one of the biggest advantages to it. To my knowledge, LPDDR is actually why Apple started doing it to cut power consumption down several years ago. LPDDR is solder-only and has never had a SO-DIMM format, so any laptop that uses LPDDR has had to solder it to the board.
I've used Apple products since the 90s, and there's a lot I like about their modern products. At a high level-
Soldered components actually do benefit their customers. LPDDR RAM is a solder-only RAM type, but it has lots of benefits such as being faster and more power efficient. And the Unified Memory configuration makes it even better. It also reduces overall bulk, allowing thinner designs and more space to be used for other components such as bigger batteries. Upgrading RAM is kind of cool and all (I upgraded the RAM in my old Mid 2012 MacBook Pro), but very few people actually take the time to upgrade the RAM in their computer.
The unified memory is a different topic.
If you mean soldering it to the SOC package, the 0.5% estimate is probably not inaccurate (even that might be an overestimation, but I suppose we don't have actual numbers to compare here). There may be additional factors that increase latency beyond just the distance traveled to the chips (I'm no expert, maybe someone can chime in), but if we're comparing latency improvements that would be caused by the reduced distance of electrical transmission, we're talking about 1 CPU clock cycle for every 1.47 inches of distance saved (assuming electrical transmission at around half the speed of light).
Out of curiosity, which CPU tasks benefit from extreme RAM bandwidth? I know Nvidea manage a lot more bandwidth on their new cards than Apple manage on the Max, but that's for graphical tasks.
I could be wrong, but as I understand it, in terms of soldered memory, lots of heavier tasks such as rendering will benefit from higher RAM bandwidth. As far as I’m aware, pretty much anything that needs more RAM will benefit from the higher speed RAM. And because it is Unified Memory, it will also have sizeable benefits for graphics heavy tasks as well. And generally speaking, to be obvious, everything on the computer will benefit from the higher speed performance, everything will be snappier, even light tasks. You may argue over how “snappy” the system should be or needs to be, but it is definitely faster and more energy efficient than SO-DIMM cards.
On this, I am no expert. I do know Anandtech tried to max out the M1 Max's memory bandwidth a few years ago and could not figure out a CPU workload to do it, although they apparently did manage to get pretty close.