Why no VRAMDISK?
>>59639908
Look up the SSG concept AMD put out.
They literally stuck a fucking M.2 SSD on a Fury X.
>>59639908
https://github.com/Overv/vramfs
>>59639908
This is the year that GPU memory architectures will start to get more robust, and weird.
Vega will be using at least part of its HBM like an enormous, medium latency transparent cache into host RAM and even NVMe block storage.
Previously, GPUs had to put data transfer/DMA instructions in the driver or shader programs, whereas in the future, shader programs can just try to access a big virtual memory space, and warps/wavefronts will be automatically rescheduled for deferred execution until the automatic background copy completes on a cache miss.
This is good in a way, since for most frames, a GPU only needs to touch a small portion of the data in its local RAM, so new cards can actually use slightly smaller RAM pools to save power.
The bad news is that now nothing will stop devs from shipping 100+ GB games requiring SSD installs and 16-32 GB of host RAM for secondary art asset pooling.
>>59641262
Good ol trow more resources and fuck optimization.
>>59643011
has been done lately and will continue in the future. we fucked naow bitcheseedeeddedededed