4 ms·

Doubling the bus width means more wiring and probably more PCB layers than offering two slots per channel. Right now, using DDR5 SODIMMs instead of LPDDR5 woul

Doubling the bus width means more wiring and probably more PCB layers than offering two slots per channel. Right now, using DDR5 SODIMMs instead of LPDDR5 would throw away half of the benefits of doubling the bus width, so that's a non-starter. Using CAMM style modules may allow for full-speed operation, but those have almost zero adoption so far and I think it's too soon to tell how widespread they'll become. Whatever option for connecting the DRAM, doubling the bus width doubles the amount of DRAM the OEMs must populate the system with, adding another cost on top of the more expensive SoC and MLB.

rbanffy 3y ago
> Doubling the bus width means more wiring and probably more PCB layers than offering two slots per channel. Wouldn't that be offset, at least on desktop processors, with more slots and memory channels?
- wtallis 3y ago
  No. Desktops make the cost issues even worse, because supporting twice the memory bus width not only means more wires, but also a larger, more expensive CPU socket, and maybe a new heatsink mounting standard.
  - rbanffy 3y ago
    Shouldn’t you get more bandwidth by having all memory slots populated?
    - AnthonyMouse 3y ago
      Current PC desktops have two memory channels and often four slots, because each channel has two slots. You get more bandwidth if you populate both channels, but four isn't better than two. If you wanted four channels the CPU socket would need more pins, because right now the second and third slots use the same pins as the first and second. What you could do is put some fast/wide DRAM on the CPU package. Then you could still upgrade the fast memory by upgrading the CPU, and the pins don't have to go through the socket. This should also be compatible with existing CPU sockets, and not incompatible with also having ordinary DDR in the existing memory slots. You basically get an L4 cache made out of HBM, which is available to the iGPU.
    - wtallis 3y ago
      You get more bandwidth by having all memory channels populated. When you have multiple slots per channel, filling all slots at best gets you a few percent better memory performance due to having more ranks of memory, but for DDR5 platforms the extra load on the memory controller due to multiple DIMMs per channel brings a big hit to the (officially supported) memory clock speeds. The fastest memory configs for a desktop currently are using one single-rank module per channel. Using one dual-rank module per channel doubles the capacity for typically a slight performance penalty.
      - rbanffy 3y ago
        Is there any literature on this you would recommend?

rbanffy 3y ago

> Doubling the bus width means more wiring and probably more PCB layers than offering two slots per channel. Wouldn't that be offset, at least on desktop processors, with more slots and memory channels?

wtallis 3y ago
No. Desktops make the cost issues even worse, because supporting twice the memory bus width not only means more wires, but also a larger, more expensive CPU socket, and maybe a new heatsink mounting standard.
- rbanffy 3y ago
  Shouldn’t you get more bandwidth by having all memory slots populated?
  - AnthonyMouse 3y ago
    Current PC desktops have two memory channels and often four slots, because each channel has two slots. You get more bandwidth if you populate both channels, but four isn't better than two. If you wanted four channels the CPU socket would need more pins, because right now the second and third slots use the same pins as the first and second. What you could do is put some fast/wide DRAM on the CPU package. Then you could still upgrade the fast memory by upgrading the CPU, and the pins don't have to go through the socket. This should also be compatible with existing CPU sockets, and not incompatible with also having ordinary DDR in the existing memory slots. You basically get an L4 cache made out of HBM, which is available to the iGPU.
  - wtallis 3y ago
    You get more bandwidth by having all memory channels populated. When you have multiple slots per channel, filling all slots at best gets you a few percent better memory performance due to having more ranks of memory, but for DDR5 platforms the extra load on the memory controller due to multiple DIMMs per channel brings a big hit to the (officially supported) memory clock speeds. The fastest memory configs for a desktop currently are using one single-rank module per channel. Using one dual-rank module per channel doubles the capacity for typically a slight performance penalty.
    - rbanffy 3y ago
      Is there any literature on this you would recommend?