[nre]

The 4600G supports two channels of DDR4-3200 which has a maximum memory bandwidth of around 50GB/s (actual graphics cards are in the hundreds). While this chip may be decent for SD and other compute-bound AI apps it won't be good for LLMs as inference speed is pretty much capped by memory bandwidth.

Apple Silicon has extremely high memory bandwidth which is why it performs so well with LLMs.

[AnthonyMouse]

> The 4600G supports two channels of DDR4-3200 which has a maximum memory bandwidth of around 50GB/s (actual graphics cards are in the hundreds).

DDR4-4800 exists. 76.8GB/s. You can also get a Ryzen 7000 series for around $200 that can use DDR5-8000, which is 128GB/s. By contrast, the M1 is 68GB/s and the M2 is 100GB/s. (The Pro and Max are more, but they're also solidly in "buy a GPU" price range.)

[toast0]

Doesn't change the meat of your argument, the 4000G series runs DDR4-3600 pretty well (even if the spec sheet only goes to 3200), and it's almost a crime to run an APU at anything worse than DDR4-3600 CL16. You can go higher than that too, but depending on your particular chip, when you get much above 3600, you may not be able to run the ram base clock at the same rate as the infinity fabric, which isn't ideal.

[acumenical]

Odd. The high memory bandwidth of M2 intrigues me but I have not seen many people having success with AI apps on Apple Silicon. Which LLMs run better on Apple silicon than comparably priced Nvidia cards?

[AnthonyMouse]

They don't run better on AS than on GPUs with even more memory bandwidth. They run better on AS than on consumer PC CPUs (or presumably iGPUs) with less memory bandwidth.

[seniorivn]

there are no comparably priced nvidia cards, thats the point, of comparing apple soc/apu with amd/intel, specialized hardware is and always will be better