News Forspoken Showcases DirectStorage API to Boost NVMe Performance

Once DirectStorage becomes mainstream, it'll really put the squeeze on GPUs with only 4.0x4, even 4.0x8 could be a tight squeeze.
I suspect it won't matter that much, even on something like the RX 6500 XT. When a game loads from storage, which will normally be limited to x4 PCIe 4.0 for the SSD, the GPU isn't doing a ton of other stuff. It's waiting for data from the CPU/RAM/SSD to arrive, often compiling shaders for the GPU architecture as well (on first run). The GPU should be able to accept data basically as fast as the SSD can send it, even with an x4 connection on the GPU.
 

fevanson

Distinguished
Sep 20, 2014
36
11
18,535
Once DirectStorage becomes mainstream, it'll really put the squeeze on GPUs with only 4.0x4, even 4.0x8 could be a tight squeeze.

It will be limited by the nvme SSD speed (PCIE 3.0 x4 or PCIE 4.0 x4), in the current architecture the datapath is NVME->Ram->GPU Memory and the gpu performs the decompression through compute shaders.
 
It will be limited by the nvme SSD speed (PCIE 3.0 x4 or PCIE 4.0 x4), in the current architecture the datapath is NVME->Ram->GPU Memory and the gpu performs the decompression through compute shaders.
This is incorrect (unless by "current" you mean "future DirectStorage" use). The current non-DirectStorage flow is NVME -> RAM -> CPU (decompress) to RAM -> Copy to GPU VRAM:
rggYQCRpthNwHuRwqkFZ54.png


For DirectStorage, the CPU is totally removed from the equation and it's just NVME -> RAM -> GPU VRAM -> GPU decompress to VRAM:
pBQ4cmwRYPqWdVuUDF2Tf4.png
 
  • Like
Reactions: drivinfast247

fevanson

Distinguished
Sep 20, 2014
36
11
18,535
This is incorrect (unless by "current" you mean "future DirectStorage" use). The current non-DirectStorage flow is NVME -> RAM -> CPU (decompress) to RAM -> Copy to GPU VRAM:
rggYQCRpthNwHuRwqkFZ54.png


For DirectStorage, the CPU is totally removed from the equation and it's just NVME -> RAM -> GPU VRAM -> GPU decompress to VRAM:
pBQ4cmwRYPqWdVuUDF2Tf4.png

Yes I meant the current architecture of DirectStorage.
 
Yes I meant the current architecture of DirectStorage.
Either way I figured including the two diagrams would be useful for any others that might be interested.

I'm very curious to see how this actually plays out in practice, though. Seems like the biggest bottleneck might be the CPU decompression, which could be eliminated even with a SATA SSD. I'd love to see modern games where load times get down into the <10 second range from desktop to game, though!
 

InvalidError

Titan
Moderator
It will be limited by the nvme SSD speed (PCIE 3.0 x4 or PCIE 4.0 x4), in the current architecture the datapath is NVME->Ram->GPU Memory and the gpu performs the decompression through compute shaders.
Get it?

With DirectStorage normalizing the heavy movement of data from storage to RAM to GPU where system RAM can still be used as a cache for compressed (file system) data as it has traditionally been for the last 20+ years, compressed asset streaming from system memory could still benefit from 5.0x16 even if you have a SATA SSD, just going to have a bit more asset pop on first-time load.

With no caching going on, having a GPU interface that is 4X as fast as the NVMe SSD still means 1/4th as much second-hop latency relaying the data and more spare bandwidth for CPU-GPU traffic which is still needed.
 
Mar 6, 2022
1
0
10
Once DirectStorage becomes mainstream, it'll really put the squeeze on GPUs with only 4.0x4, even 4.0x8 could be a tight squeeze.
Keep in mind that geometry and texture assets through DirectStorage remain compressed until they reach the GPU. Which means you essentially get 2x the bandwidth over PCIe from RAM to VRAM. (assuming a 2:1 compression ratio)


The CPU is also still a part of the equation of course. It still needs to issue I/O requests and copy that data from Storage to RAM, and then from RAM to VRAM. We're not at the point where the CPU is out of the equation yet.. But that's the ultimate plan, as they introduce dedicated decompression/I/O chips into the pipeline.
 
Mar 9, 2022
1
0
10
I honestly am really looking forward to these truly next gen technologies making it to video games - Too bad the game was Delayed to October 11th.

I would love a demo sooner though (Like Ground Zeros with MGSV)