6 ms·

> soldered RAM Hold on a minute. It's not "soldered". It's integrated with the SoC. The benefit is memory latency and bandwidth. If you know Framework, their

by rollcat 10mo ago

> soldered RAM Hold on a minute. It's not "soldered". It's integrated with the SoC. The benefit is memory latency and bandwidth. If you know Framework, their entire mission is to build upgradeable laptops, and they keep delivering. Now they also wanted to build an incredibly powerful, but small and quiet desktop. They went directly to AMD, asked their engineers to make the memory upgradeable. AMD worked really hard and said not possible, not unless you want all of these cores to sit idle. https://frame.work/blog/framework-desktop-deep-dive-ryzen-ai-max https://frame.work/blog/framework-desktop-deep-dive-ryzen-ai... The world has moved on. Just as you no longer have discrete cache chips or discrete FPUs, you can't do discrete memory anymore - unless you don't need that level of performance, in which case CAMM is still an excellent choice. But that's not what Apple does. M1 redefined the low-end. It will remain a great choice in 5 years, even when macOS kills it off - Asahi remains very decent.

4ggr0 10mo ago
> The world has moved on we're talking about laptops, right?
- doug-moen 10mo ago
  no, they are talking about high performance desktops, mostly. They link to the Framework desktop, which has 256 GB/s memory bandwith. For comparison, the Apple Mac Pro has 800 GB/s memory bandwidth. Neither manufacturer is able to achieve these speeds using socketed memory.
  - 4ggr0 10mo ago
    > no, they are talking about high performance desktops then i don't really get the "world has moved on"-claim. in my bubble socketed RAM is still the way to-go, be it for gaming or graphics work. of course Apple-user will use a Mac Pro, but saying that the world has moved on when it's about high-performance, deluxe edge-cases is a bit hyperbolic. but maybe my POV is very outdated or whatever, not sure.
    - everforward 10mo ago
      I think, but am not totally positive, this is primarily a concern for local LLM hardware. There are probably other niches, but I don't it's something most people need or would noticeably benefit from.
    - rollcat 10mo ago
      I agree and I do not agree. I still sometimes use a Thinkpad X230, and wait- a G4 PowerBook, and they are fine machines for many tasks. Yet even those have soldered CPUs, simply because of design constraints. You don't need to have to train models. You want to play a game like Factorio, that, of all things, is bottlenecked on memory bandwidth - you must update each entity in a huge world on every tick, at 60 UPS, and yes, the game is insanely well optimized (check the dev blog). You don't have to play Factorio, but you also technically don't need DMA.

by rollcat 10mo ago

4ggr0 10mo ago

> The world has moved on we're talking about laptops, right?

doug-moen 10mo ago
no, they are talking about high performance desktops, mostly. They link to the Framework desktop, which has 256 GB/s memory bandwith. For comparison, the Apple Mac Pro has 800 GB/s memory bandwidth. Neither manufacturer is able to achieve these speeds using socketed memory.
- 4ggr0 10mo ago
  > no, they are talking about high performance desktops then i don't really get the "world has moved on"-claim. in my bubble socketed RAM is still the way to-go, be it for gaming or graphics work. of course Apple-user will use a Mac Pro, but saying that the world has moved on when it's about high-performance, deluxe edge-cases is a bit hyperbolic. but maybe my POV is very outdated or whatever, not sure.
  - everforward 10mo ago
    I think, but am not totally positive, this is primarily a concern for local LLM hardware. There are probably other niches, but I don't it's something most people need or would noticeably benefit from.
  - rollcat 10mo ago
    I agree and I do not agree. I still sometimes use a Thinkpad X230, and wait- a G4 PowerBook, and they are fine machines for many tasks. Yet even those have soldered CPUs, simply because of design constraints. You don't need to have to train models. You want to play a game like Factorio, that, of all things, is bottlenecked on memory bandwidth - you must update each entity in a huge world on every tick, at 60 UPS, and yes, the game is insanely well optimized (check the dev blog). You don't have to play Factorio, but you also technically don't need DMA.