Save VRAM

With the official launch of mipmap streaming, it is possible to save a lot of VRAM in texturing.

However, another source that many may not be aware of that consumes a lot of VRAM is the shape key.

shape key on Skinmesh consume a lot of VRAM on current assets.

Some of them even have redundant vertex attributes, through which the consumed VRAM can be effectively clipped.

Even with mipmap streaming, there is still a lot of preloaded VRAM in the current VRAM, and it is not possible to visualize the real VRAM consumption.

Assuming that there are 30 avatars, the current statistics use 9GB of VRAM.

After mipmap streaming is enabled, the real VRAM consumption will be around 4.5GB.

By trimming out the unnecessary and unused shape key, it is possible to save about 1.5GB of VRAM to the overall 3GB of VRAM consumption.

Although there is still a lot of mesh memory waste, such as many polygons and vertex attributes taking up tens of MB due to multiple suits, cutting out unnecessary shape key is the most effective way.

I hope to make many people aware of this, it seems that many people still think that texturing memory is the only thing that wastes VRAM, but the arrival of mipmap streaming will make people pay attention to the problem of shape key, once both problems are solved effectively, even if the VRAM of the GPU is limited, it will still be able to perform well.

In the next year, the mainstream GPU is still expected to be around 8GB, and the 4060 will probably be number one. The 5060 is likely to emerge next year and gradually take over the lead, however GDDR7 will initially only be produced in a single die of 2GB, meaning 128bit-width will still only have 8GB of VRAM.

Trying to save as much VRAM as possible will prevent serious stuttering problems, although there are still a lot of issues to be solved.

Also, meshes usually have a much lower compression ratio than textures, which means that clipping the shape key can also have a significant enough effect to reduce the download size.

This will also save a lot of bandwidth for the current limited network bandwidth.