3 ms·
Are they just ripping people off on memory, or is their whole “unified memory” thing more expensive than popping in a dimm? I actually don’t understand their m
by bee_rider 3y ago
Are they just ripping people off on memory, or is their whole “unified memory” thing more expensive than popping in a dimm?
I actually don’t understand their memory system at all, the typical trade off is high latency, high bandwidth for graphics memory, and low latency, relatively low bandwidth for CPU. They get those high bandwidth numbers, so it must be more like GDDR. But nobody seems to be complaining about memory latency on these chips?
Maybe they need some cheaper RAM, to get swapped into; a RAM disk formatted to swap, haha.
- nordsieck 3y agoFrom what I understand, they're using "regular ram", but it's on package, so they can have a much wider bus between the ram and the processor. Also, the traces are much shorter, so they get a small latency benefit.
- bee_rider 3y agoOh interesting, I assumed started at GDDR and done something tricky to get the latency down, but your explanation makes a lot more sense.
- Havoc 3y ago> it must be more like GDDR Its LPDDR5 so not slow but nowhere near gddr either. Best as I can tell its just more parallel than usual setups - potentially made feasible by arm. If you look at altra etc they also seem to support loads of ram