3 ms·

I know I would, and I dare say most people are the same, prefer to have the option for 3% lower memory bandwidth in exchange for 4x more memory. There’s no just

I know I would, and I dare say most people are the same, prefer to have the option for 3% lower memory bandwidth in exchange for 4x more memory. There’s no justification for Apple’s memory prices.

tough 3y ago
I see it like cars, mass produced standard models are way cheaper than the ones with all the extras and whistles and what not
wtallis 3y ago
Where do you get that "3% lower memory bandwidth" figure from? LPDDR5 is running at 6400 MT/s on at least Apple and Intel platforms, and the fastest DDR5 SODIMMs I've seen so far are 5600MT/s: 12.5% slower, assuming equal bus width (only true for the base M1/2 processors).
- labcomputer 3y ago
  3%, 12% Potato, potato. It’s a small spec performance difference for a huge price difference. Meanwhile, for actual real-world workloads, not spilling to swap (because you have enough memory) will result in real-world performance that is many times (even an order of magnitude) better. Even with Apples blazing fast SSDs. Oh, and you can also spec a cheaper SSD because you don’t need one that can take a bajillion writes (because you aren’t treating your SSD as memory). AND the cheaper SSD will last longer too. In short, Apple could spec cheaper components AND get better real world performance by simply shipping computers that have enough memory. Edit: also 3% comes from shipping the absolute fasted DDR5 modules vs the second fasted DDR5.
  - sliken 3y ago
    Sure, that's 3%, but what about shipping 2x, 4x, and 8x as wide memory systems, largely enabled by not using industry standard dimms.
sliken 3y ago
3%? M2 = 100GB/sec M2 pro = 200GB/sec M2 max = 400GB/sec M1 Extreme = 800GB/sec. Most x86-64 laptops and desktops are in the 60-80GB/sec range. Some desktop or server only chip (threadripper, epyc, Xeons, a few intel workstation chips) have 150 or 300GB/sec. 400GB/sec vs 80GB/sec is quite a bit more than 3%, and is available in the mbp 14" or mbp 16".
- labcomputer 3y ago
  That’s because Apple Silicon has a stupid-wide memory bus. It has little/nothing to do with the characteristics of the memory chips shipped in the box. They just put more modules in parallel than on x86. Apple could ship cheaper memory modules with the same stupid-wide bus and achieve nearly the same performance for a fraction of the cost. And that’s what I want. Give me 4x as much DDR4 on the same 1024 bit wide bus, and I’d be happy.
  - sliken 3y ago
    Stupid wide? Pins, memory controllers, and traces (and layers) on motherboards aren't cheap. Apple just decided that the paying memory bandwidth was worth it. Is there some reason why you think the extra memory bandwidth was stupid? Sure it's easy to ask for 1024 bit wide memory bus, but it's not cheap. Done with the traditional x86-64 method you'll need two 280 watt CPUs (a few $k each), on a $1k motherboard, with 24 dimms, a 1kw power supply, ear plugs, and at least $10k. The minimum physical size is going to be substantial since you need two large CPU sockets (6096 pins each), 24 dimm sockets, and of course a substantial number of fans, dimms, and power supply. Even managing airflow over 800 watts or so of chips isn't going to be easy, cheap, or quiet. Because of the above size, cost, noise, and heat it's targeted at enterprise and cloud folks with the related cost, service contract, and sales pipelines. Or you could buy an Apple studio and hide it somewhere, you'll likely not be able to find it by listening for it.

tough 3y ago

I see it like cars, mass produced standard models are way cheaper than the ones with all the extras and whistles and what not

wtallis 3y ago

Where do you get that "3% lower memory bandwidth" figure from? LPDDR5 is running at 6400 MT/s on at least Apple and Intel platforms, and the fastest DDR5 SODIMMs I've seen so far are 5600MT/s: 12.5% slower, assuming equal bus width (only true for the base M1/2 processors).

labcomputer 3y ago
3%, 12% Potato, potato. It’s a small spec performance difference for a huge price difference. Meanwhile, for actual real-world workloads, not spilling to swap (because you have enough memory) will result in real-world performance that is many times (even an order of magnitude) better. Even with Apples blazing fast SSDs. Oh, and you can also spec a cheaper SSD because you don’t need one that can take a bajillion writes (because you aren’t treating your SSD as memory). AND the cheaper SSD will last longer too. In short, Apple could spec cheaper components AND get better real world performance by simply shipping computers that have enough memory. Edit: also 3% comes from shipping the absolute fasted DDR5 modules vs the second fasted DDR5.
- sliken 3y ago
  Sure, that's 3%, but what about shipping 2x, 4x, and 8x as wide memory systems, largely enabled by not using industry standard dimms.

sliken 3y ago

3%? M2 = 100GB/sec M2 pro = 200GB/sec M2 max = 400GB/sec M1 Extreme = 800GB/sec. Most x86-64 laptops and desktops are in the 60-80GB/sec range. Some desktop or server only chip (threadripper, epyc, Xeons, a few intel workstation chips) have 150 or 300GB/sec. 400GB/sec vs 80GB/sec is quite a bit more than 3%, and is available in the mbp 14" or mbp 16".

labcomputer 3y ago
That’s because Apple Silicon has a stupid-wide memory bus. It has little/nothing to do with the characteristics of the memory chips shipped in the box. They just put more modules in parallel than on x86. Apple could ship cheaper memory modules with the same stupid-wide bus and achieve nearly the same performance for a fraction of the cost. And that’s what I want. Give me 4x as much DDR4 on the same 1024 bit wide bus, and I’d be happy.
- sliken 3y ago
  Stupid wide? Pins, memory controllers, and traces (and layers) on motherboards aren't cheap. Apple just decided that the paying memory bandwidth was worth it. Is there some reason why you think the extra memory bandwidth was stupid? Sure it's easy to ask for 1024 bit wide memory bus, but it's not cheap. Done with the traditional x86-64 method you'll need two 280 watt CPUs (a few $k each), on a $1k motherboard, with 24 dimms, a 1kw power supply, ear plugs, and at least $10k. The minimum physical size is going to be substantial since you need two large CPU sockets (6096 pins each), 24 dimm sockets, and of course a substantial number of fans, dimms, and power supply. Even managing airflow over 800 watts or so of chips isn't going to be easy, cheap, or quiet. Because of the above size, cost, noise, and heat it's targeted at enterprise and cloud folks with the related cost, service contract, and sales pipelines. Or you could buy an Apple studio and hide it somewhere, you'll likely not be able to find it by listening for it.