3 ms·
> ultra-fast on-die RAM The RAM is not on die. It’s just soldered on top of the SoC package. > All that is to say, where you needed 1.5TB of RAM on a Xeon, th
by masklinn 5y ago
> ultra-fast on-die RAM
The RAM is not on die. It’s just soldered on top of the SoC package.
> All that is to say, where you needed 1.5TB of RAM on a Xeon, the architectural changes on Apple Silicon likely mean you can get away with far less and still wind up performing better.
No, it does not. You might save a bit, but most of what you save is the transfers, because moving data from the CPU to the GPU is just sending a pointer over through the graphics API, instead of needing to actually copy the data over to the GPU’s memory. In the latter case, unless you still need it afterwards you can then drop the buffer from the CPU.
You do have some gains as you move buffer ownership back and forth instead of needing a copy in each physical memory, but if you needed 1.5TB physical before… you won’t really need much less after. You’ll probably save a fraction, possibly even a large one, but not “2/3rd” large, that’s just not sensible.