3 ms·
The article shows that Intel and AMD's desktop iGPUs have higher bandwidth and lower latency than the Switch's X1.
by em500 3y ago
The article shows that Intel and AMD's desktop iGPUs have higher bandwidth and lower latency than the Switch's X1.
- kevingadd 3y agoUnified Memory is not about bandwidth or latency. On a desktop PC, you have CPU and GPU side buffers with different addresses that may or may not map to the same physical memory, and the APIs require you to perform mapping/unmapping operations or perform explicit uploads and downloads. On target with unified memory, it is all one address space, the data is just where you put it. There are not split CPU and GPU memory modules. (For some consoles there are different categories of memory, usually based on traits like write combining, but it's all one unified address space). Even now with Vulkan, the experience of programming for a unified memory console is very different. A desktop PC driver can, ideally, optimize out many of your upload/download/map/unmap operations. With unified memory, they simply don't exist.
- _ph_ 3y agoAnd it hast to be noted that the Apple Silicon has also a unified memory architecture. It should allow for some interesting gaming performance if any developer tries to fully optimize for it.
- kimixa 3y agoModern GPUs allow mapping memory through PCIe, so you can treat memory as the same. Resize-able BAR effectively allows the same the other way, where all the GPU vram can be mapped into the CPU. The issue is that suddenly accessing an address can have orders of magnitude lower bandwidth and higher latency, if it is resident on the far memory bus. So for performance reasons you still really need to treat them differently.