Bleeding-edge-clocked DRAM is a lot more costly per GB to produce than middle-of-the-pack-fast DRAM. (Which is weird, given that process shrinks should make things cheaper; but there's a DRAM cartel, so maybe they've been lazy about process shrinks.)
Apparently DRAM and NAND do not shrink as well because in addition to transistors in both cases you need to store some kind of charge in a way that is measurable later on - and the less material present, the less charge you are able to store, and the harder it is to measure.
> The M1's memory is LPDDR4X-4266 or LPDDR5-5500 (depending on the model, I guess?) which is about double the frequency of the memory in the Intel Macs.
That's a high frequency, but having two LPDDR chips means at most you have 64 bits being transmitted at a time, right? Intel macs (at least the one I checked), along with most x86 laptops and desktops, transfer 128 bits at a time.
> Apparently, this alone seems to account for a lot of the M1's perf wins — see e.g. the explanation under "Geekbench, Single-Core" here
That's a vague and general statement that site always says, so I wouldn't put much stock into it.
I missed that, I assumed virtualisation was dependent on Intel VT.
Then again I would have expected them to have discussed it as much as the video editing.
I am guessing that they’d need a M2 type chipset for accessing more RAM for that. Or maybe they’ve got a new way to do virtualisation since that is such a key thing these days.
Edit: thanks for pointing that out though, that’s why I mentioned it
How well this fits in with current virtualisation would be interesting to find out; I guess this will be for a later version of Big Sur, with a new beefier M2 chip.
Are they virtualizing x86 though? Having Docker running arm64 on laptops and Docker x86 on servers completely nullifies the main usecase of Docker imo.
The M1's memory is LPDDR4X-4266 or LPDDR5-5500 (depending on the model, I guess?) which is about double the frequency of the memory in the Intel Macs.
Apparently, this alone seems to account for a lot of the M1's perf wins — see e.g. the explanation under "Geekbench, Single-Core" here: https://www.cpu-monkey.com/en/cpu-apple_m1-1804
Bleeding-edge-clocked DRAM is a lot more costly per GB to produce than middle-of-the-pack-fast DRAM. (Which is weird, given that process shrinks should make things cheaper; but there's a DRAM cartel, so maybe they've been lazy about process shrinks.)