Pemalite said: PC memory works differently than consoles. |
I am aware that unified memory needs to address the demands of both the GPU and CPU.
My point was though if you have a VRAM capacity bottleneck (because you can only ever dedicate a fixed 4GB of relatively high bandwidth memory to graphics and swapping assets between the system ram and video ram slows down/interrupts the pipeline and has associated copy-function overhead anyway), having a large share of unified memory that can be apportioned flexibly is an advantage (assuming there is memory free to allocate.)
Given the rumors, we can expect the Switch 2 to be able to allocate something like 6 - 8 GB of its unified memory to graphics use-cases. That's an advantage over the 2050, given that the relative throughput of the video memory is roughly comparable (112 GB/s vs. 102 GB/s) and a large share of the laptop's system ram is eaten up by windows bloat.
So you have a situation where the 2050 has 4GB of VRAM @112 GB/s and 16GB - (Windows bloat) of system ram @ something like 42 GB/s (for DDR5 laptop.)
vs
Switch 2 with 12 to 16GB - (Switch OS bloat << Windows bloat) of unified memory at 102 GB/s.
I probably should've been more precise in what I was saying when I said "make up for lower clocks/CUDA cores." What I intended to say is that if you have a situation where the most common bottleneck in the graphics pipeline is VRAM capacity, the extra core/clock throughput is wasted anyway. And if that is the bottleneck, then yes having more effective graphics-purposed memory will lead to better performance. In those instances where it isn't the bottleneck, then no it "doesn't make up for it." I'd expect that if they loosened up the clock-rate limits in the Cyberpunk benchmark for example, they probably wouldn't get that much better performance because it is likely the VRAM that is the bottleneck, not the compute-capacity.
Last edited by sc94597 - on 15 November 2023