8GB RAM on M3 MacBook Pro ‘Analogous to 16GB’ on PCs, Claims Apple::Following the unveiling of new MacBook Pro models last week, Apple surprised some with the introduction of a base 14-inch MacBook Pro with M3 chip,…
8GB RAM on M3 MacBook Pro ‘Analogous to 16GB’ on PCs, Claims Apple::Following the unveiling of new MacBook Pro models last week, Apple surprised some with the introduction of a base 14-inch MacBook Pro with M3 chip,…
DDR5 runs at 52GB/s. Apple uses RAM that runs at up to 800GB/s (if you have enough, gets faster the more you have since it runs in parallel… but it’s never as slow as DDR5).
Huge doubt here. Apple RAM is LPDDR5. That’s Low Power DDR5.
Citing this site:
I’m looking at the Apple M2 Wikipedia page and it has the 800GB/s number you have, but that’s gotta be something like RAM speed times number of RAM unit blocks for overall bandwidth.
Apple RAM is not magically 15 times faster than DDR5.
tl;dr
The memory bandwidth isn’t magic, nor special, but generally meaningless. MT/s matter more, but Apple’s non-magic is generally higher than the industry standard in compact form factors.
Long version:
How are such wrong numbers are so widely upvoted? The 6400Mbps is per pin.
Generally, DDR5 has a 64-bit data bus. The standard names also indicate the speeds per module: PC5-32000 transfers 32GB/s with 64-bits at 4000MT/s, and PC5-64000 transfers 64GB/s with 64-bits at 8000MT/s. With those speeds, it isn’t hard for a DDR5 desktop or server to reach similar bandwidth.
Apple doubles the data bus from 64-bits to 128-bits (which is still nothing compared to something like an RTX 4090, with a 384-bit data bus). With that, Apple can get 102.4GB/s with just one module instead of the standard 51.2GB/s. The cited 800GB/s is with 8: most comparable hardware does not allow 8 memory modules.
Ironically, the memory bandwidth is pretty much irrelevant compared to the MT/s. To quote Dell defending their CAMM modules:
That contradiction makes it hard to balance speed, capacity, and upgradability. Even the upcoming Core Ultra 9 185H seems rated for 5600 MT/s-- after 2 years, we’re almost getting PC laptops that have the memory speed of Macbooks. This wasn’t Apple being magical, but just taking advantage of OEMs dropping the ball on how important memory can be to performance. The memory bandwidth is just the cherry on top.
The standard supports these speeds and faster. To be clear, these speeds and capacity don’t do ANYTHING to support “8GB is analogous to…” statements. It won’t take magic to beat, but the PC industry doesn’t yet have much competition in the performance and form factors Apple is targeting. In the meantime, Apple is milking its customers: The M3s have the same MT/s and memory technology as two years ago. It’s almost as if they looked at the next 6-12 months and went: “They still haven’t caught up, so we don’t need too much faster, yet-- but we can make a lot of money until while we wait.”
You’re comparing single channel performance to entire system performance.
That statement simply means the most highest of the high end Mac has 16 memory channels (admittedly more than EPYCs 12, but EPYC is in the ballpark). The mere mortal entry M2 has two channels, just like almost every desktop/laptop grade x86 CPU. They are not getting 800 out of only 8Gb of modules.
I’m really interested in this kinda thing, do you have sources I can read?
What I found was DDR5 runs at a max of 64 GB/s, and the M2 Pro runs at 400 GB/s.
I can’t find anything about it being faster due to running in parallel.Edit:I found it, looks like the M2 Ultra runs at 800 GB/s, cool. If I’m reading correctly, this was done by connecting two M2 ProsAlso, the PS5 allegedly has over 400GB/s bandwidth just for perspective
Note those are comparing different numbers.
The number you quoted was for a single memory channel.
A processor has as many memory channels as it feels like. So that 800 number basically means about 16 channels. The M2 plain seems to be about two channels.
For comparison, x86 desktop CPUs have long been 2 channel designs. You go up the stack and you have things like EPYC having 12 channels.
So for single socket design, apple likely has a higher max memory performance than you can do single socket in x86 (but would likely turn in lower numbers than a dual socket x86 box).
So to clarify, the M2 Ultra runs at 800 GB/s because it’s utilizing multiple memory channels, which is like running dual/quad/etc. channel RAM in an x86 PC. So at the max 64 GB/s bandwidth of DDR5 ram, you could run quad channel and get 256 GB/s right? And getting up to 12 channels of DDR5 could mean a bandwidth of 768 GB/sec?
Yeah, in that case Apple is definitely over charging. To be fair, my mb can’t run 12 channels of RAM but it also won’t cost me an arm and a leg and a kidney to have similar performance per GB
Note that I can’t think of modem four channel x86. Either they are the usual two channel (two dimms per channel is how four dimm slots are organized) or have way more (Sapphire Rapids, Bergamo)
To map the M2, the base is about the same as most x86 consumer grade, the Pro is about Threadripper, and the ultra is somewhere between single or dual socket Bergamo, at least in terms of memory bandwidth, which is a highly specific metric.
Oh gotcha, thanks for straightening me out on that. I’m still learning tech so the examples are really helpful
The PS5 uses a unified 16gb of GDDR6, which has really high bandwidth for graphics applications. Apple is full of shit about their LPDDR5.
Sorry you got down voted so hard. Your comment spawned a lot of discussion which is a good thing.