*DX9 had to duplicate things, so system RAM performance mattered a heap. This was because most of the alterations were done by the CPU from decompression to texture swapping.
*DX10 began a de-sync of this, using the CPU less and letting the GPU do more. so it was decompressed into system RAM, then moved to VRAM.
arrays enable swapping of textures in GPU without CPU intervention.
Direct3D 9 had a limitation on the number of draw calls because of the CPU cost per draw call. On Direct3D 10, the cost of each draw call has been reduced
Then gaming websites put out articles like this saying DX10 didn't look any better
"DX10 doesn't look any better, it's not worth it" - and that hurt GPU sales
Suddenly, game devs had to implement features that made DX10 LOOK better, when it's goal was to PERFORM better.
This was rarely ever discussed, except it was a big feature of Aero in windows 7 at the time It ran Aero in DX10.1 hardware accelerated on the GPU *or* a software accelerated mode since Intel and Nvidia hadn't quite caught up at the time of release and just had 'compatible' hardwaer.
If you had true DX10.1 hardawre toto move the rendering of the OS to the GPU it made it feel more responsive and used less resources - it made those crappy intel atom netbooks feel a whole lot smoother.
DX10 and DX11 titles suffered the Crysis problem here:
Instead of using these features to improve performance, Game devs instead crammed every new lighting effect into a game -RTX style- and tanked performance instead.
Company of heroes was the first DX10 game (Added via patch and it had *major* problems vs it's DX9 rendering method, but you could run a single command line shortcut to fix it all - because they added physics enabled flowers and rocks that ran off the CPU, negating all their performance gains because
DirectX 9 vs. DirectX 10 - Real World DirectX 10 Performance: It Ain't Pretty (anandtech.com)
CoH ran at about 50% of DX9 performance - but using a single command "-nolitter" would fix it because DX10 added in LITERAL TRASH to the game world that had software based physics interactions that undid any and all performance gains. Weirdly almost all references to this command seemed to have vanished online, despite it being a huge deal when it was introduced.
Some followup titles (the anandtech link above) showed performane increases, but they were sadly the minority as we reached the crysis point because that horrible nasty performance impact of making it look a tiny bit better did one thing and one thing only:
It made the game controversial. It made the game newsworthy. It became profitable to have ultra game settings that ran like shit
, because it made people who COULD run it, brag about it and started upgrade envy, and became marketing material for new GPUs as a requirement to play these games "properly"
*Now DX12 is allowing the CPU and GPU access at the same time (to different parts) thanks to ReBAR (see link below)
Up next is DirectStorage allowing the GPU direct access to data from NVME bypassing all CPU and RAM, which the new DX12 feature would still allow the CPU to access and alter the content if needed despite being removed from the process as it has been until now - that access would be critical for anything currently being done 'in-transit' so this feature allows them to 'bugfix' easier if there are problems with this new method
This seems to be all about lowering CPU and RAM requirements without
upping GPU requirements, while direcstorage can help prevent issues with low VRAM GPU's (by swapping new data in at high speeds, with zero CPU and RAM overhead slowdown)
All they need next is for GPU's to be able to cache more data in advance to use spare VRAM - but that runs the risk of increased power consumption, temperatures and even lower performance so it's not something they'll leap at. My 3090 would use another 50W to run all the VRAM (like it did with mining) - and that 50W comes from the GPU, harming performance.