>If this is actual memory bandwidth, then Apple increased that bus speed 5-15x!
Yes because GPU requires lots of bandwidth. Top of the line Nvidia GPU has ~ 1TB/s memory bandwidth.
>I've been very irritated by the pricing, but if Apple figured out how to optimize the memory bandwidth this much then maybe it's worth the extra cost per GB.
You are not only paying for memory aka LPDDR5 chip, you are also paying for the additional memory controller. Which has an implication on yield and die size cost. Whether you think it is worth or still irritated by that pricing is of course subjective.
It was written in the assumption that Desktop M chip will have much bigger GPU and hence higher Memory bandwidth requirement.
They could use 8 Channel DDR4 which gives them 200GB/s. That is roughly 3 times the memory bandwidth of current mobile M chip. Or DDR5 and pushes it to 400GB/s. Although DDR5 is barely out.
Also worth noting, if they use 8 Channel DDR configuration, the minimal amount of RAM shipping would be 64GB.
None of the above hypothesis seems very Apple to me.
AnandTech speculates on a 128-bit DRAM bus[1], but AFAIK Apple hasn't revealed rich details about the memory architecture. It'll be interesting to see what the overall memory bandwidth story looks like as hardware trickles out.
So I've observed something from many of the tests/reviews posted so far, as well as Apple's specs: Apple claims 200GBps memory bandwidth for the M1 Pro, and 400GBps for the M1 Max. Some tests show the M1 Max saturating the memory bandwidth between 200 - 240 GBps rates.
Maybe someone who understands better than me can explain these observations:
1. DDR5 RAM is supposedly 36-38GBps bandwidth, DDR4 3200 is about 25.6GBps bandwidth. (Supposedly Alder Lake is the first chipset to support DDR5.)
2. Maybe the 200-400 GBps rate advertised is the entire bus bandwidth for the SoC? If this is actual memory bandwidth, then Apple increased that bus speed 5-15x!
3. Maybe Apple divided the RAM into additional banks with more lanes to increase the bandwidth?
I've been very irritated by the pricing, but if Apple figured out how to optimize the memory bandwidth this much then maybe it's worth the extra cost per GB.
Did I catch on to something majorly different here?
Unified memory using high bandwidth buses to support both GPU and CPU and the rest.
I believe Apple uses very specific LPDDR5 packages then the pro and max have 2x and 4x buses so you need that many packages. This is why the M1 comes in 8GB ad 16GB, M1 Pro 16GB and 32GB and the Max is 32GB and 64GB.
The current M2 is 8, 16 and 24GB, so naturally a top spec M2 Max would have 4x 24GB so you get 96GB assuming they stick with a 512 bit bus.
> I don't how/why/where these myths have been perpetuated from that Apple Silicon Macs magically needs less RAM or magically Apple Silicon Macs are better at memory management.
It depends. For the Intel macbook pros, the GPUs had their own separate memory. For the intel macbook airs, the graphics could use up to 1.5GB of the system memory (so your 8GB MBA could have as little as 6.5GB of system memory)
The unified memory model in M1 means the graphics can hopefully eat up less RAM since data from the system memory can be read directly by the graphics chip (no need to copy from system memory into graphics-reserved memory)
An M2 Ultra has 800GB/s of memory bandwidth, an Nvidia 4090 has 1008GB/s. Apple have chosen to use relatively little system memory at unusually high bandwidth.
You want to ignore the difference between DDR4 and DDR5, but the market doesn't ignore it and DDR5 is more expensive (I'm talking about non-Apple memory sticks).
Also, you want to ignore 400 Gb/s provided by Apple's 64Gb memory, but... you know.
With the integrated ram and cpu and gpu on apple silicon, however it's done it yields perf results. I do think that probably has higher cost than separately produced ram. And even separate from that, because they have that unified memory model unlike every other consumer device they can charge for it. So 64, 96 or 128 gb?
> 64GB is a pretty good deal (look at desktop DDR5 prices).
Looking DDR5 price isn't valid because no CPU is released now, and DDR5 vs DDR4 performance difference will be negligible at the beginning of DDR5 unless you use iGPU. Apple always takes huge for upgrading RAM/SSD.
> It's quite obvious that RAM capacity in the M# world is closely tied to die space occupied by the CPU/GPU cores, and... it's not entirely obvious that this is unreasonable? (unlike, say, an iPhone with 64GB flash)
That has nothing to do with the 8GB base.
There are 3 different M chips each generation, and each one has a fixed number of memory channels that goes out to seperate RAM chips specced/packaged in a way similar to smartphone RAM.
The base M3 is not capped at 8GB. Apple chose to use low capacity LPDDR5X chips in the base config. It was a business/marketing decision, not a technical one.
DDR5-5600 would be 69.21GB/s[0] which is slightly more than the base Apple Silicon M1 (66.67GB/s) but well under the M1 Pro.
M2 has 100GB/s of unified memory bandwidth[1].
[0] https://www.crucial.com/articles/about-memory/everything-abo...
[1] https://www.apple.com/newsroom/2022/06/apple-unveils-m2-with...
reply