Direct Storage on PC is very heavy.

[C1REX]

Somebody made a cool video showing the cost of Direct Storage on PC that works differently than on consoles.
Consoles have dedicated hardware decompression and unified memory so data goes from fast NVME to VRAM directly at no performance cost.
PC needs to use CPU or GPU to decompress data what is very heavy.
Even 4090 loses 17% in 1% lows. So it’s a hard choice if we hammer GPU or CPU for data streaming.

https://youtu.be/Gz0Ay7sJETQ?si=gxhETjKzmTsVDSgd


And who knows? Maybe AMD or NVidia introduce hardware decompression to save performance.

[A&A]

RTX IO / Gdeflate sucks?

[C1REX]

Originally posted by tbialoof:

but yeah pcs are not optimally designed for gaming and a ton of power and efficiency is wasted by brute force

Yeah.
I’ve heard before that direct storage costs even 30% of a 3080 performance what was considered absurd here on forum but 10-20% hit to performance on a 4090 seems to confirm that.

It can be solved if GPUs come with a dedicated decompression chip. Or gaming NVME drives with DirectStorage decompression chip if possible.

[Masque]

I just got an NVME drive and Windows 11 (i'm back to 10 after a day of 11) so I could try out DirectStorage on my new motherboard and CPU.

Then, as an afterthought because I'm not very bright sometimes, I looked up what games have DirectStorage.

I own one of them: Horizon Zero Dawn, the remaster. Out of like 20 total games that support it.

I was under the impression that it was something I could use with all my games.

Dangit.

Today I'm trying to format the new NVME drive again, from inside BIOS using a Windows USB, since I tried to install Win11 on it and now the Windows 10 drive doesn't even know it exists.

What a pain in the butt.

[A&A]

Originally posted by C1REX:

Yeah.
I’ve heard before that direct storage costs even 30% of a 3080 performance what was considered absurd here on forum but 10-20% hit to performance on a 4090 seems to confirm that.

It can be solved if GPUs come with a dedicated decompression chip. Or gaming NVME drives with DirectStorage decompression chip if possible.

There will be a hit on the memory bandwidth too.
It has to read chunks from the GPU's memory and write the decompressed data back.

This offload technique can only be useful for something like an i3 or something.