Llama only uses dedicated memory when both shared and dedicated are available.

[jadbox]

Hey folks, the Razer Blade18 4090 has both dedicated and shared memory, but it seems most applications like Llama only are designed to access the dedicated. Is there a way to utilize both for performance video memory pipelining?

Linked reddit thread on the 4090 with both dedicated and shared: https://www.reddit.com/r/nvidia/comments/1c73zaf/my_4090_blade18_has_both_dedicated_and_shared/

[saostad]

I have the same question. Why does it never use the GPU-shared memory?

[HamzaYslmn]

same

[RazyRo]

?

[ctrysbita]

because shared gpu memory is not the real memory that is connected to gpu chip

[alirezanet]

because shared gpu memory is not the real memory that is connected to gpu chip

that is correct but using GPU-shared memory will increase the performance since GPU doesn't have to ask the CPU to access the data in the memory.

[ihor-sokoliuk]

same question here

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days since being marked as stale.