Unified memory changed what a chip spec means

Older computers kept the main parts apart. The processor sat in one place, the graphics chip in another, and each had its own pool of memory, with data shuttled between them across the board. A growing share of modern devices does the opposite. The processor, the graphics, and other blocks share a single pool of memory, physically close to the silicon. This is usually called unified memory, and it changed enough about how chips behave that the old way of reading a spec sheet no longer works cleanly.

What unified memory does#

In the old arrangement, when the graphics needed data the processor had, that data had to be copied from one memory pool to another. Copying takes time and energy, and it puts a ceiling on certain kinds of work. Unified memory removes the copy. Every block sees the same memory, so data does not have to move to be shared.

The gains show up in a few specific places:

• Less data movement. Work that constantly passes data between the processor and graphics avoids the copy entirely.
• Flexible allocation. Memory is not split into fixed pools, so a heavy graphics task or a large AI model can use more of it when needed.
• Better efficiency. Moving less data uses less energy, which matters on anything running off a battery.

Why it suits on-device AI#