Ozzuneoj wrote on 2023-03-31, 06:14:
Yes, sadly, without Direct Storage or some other as-yet-unavailable technology to leverage the throughput and random access spee […]
Show full quote
Kahenraz wrote on 2023-03-31, 05:30:
The last use case I can think of, which is somewhat reasonable, is to copy games onto it simply for the fastest load times. This is somewhat more realistic, but I think it would be difficult to measure compared to a DRAM cached SSD. Even though there is raw performance available, a lot of that will be lost due to filesystem and API overhead. Actual raw memory speeds won't necessarily translate into comparable filesystem performance.
Yes, sadly, without Direct Storage or some other as-yet-unavailable technology to leverage the throughput and random access speeds of solid state drives, games are just not built to make use of high speed storage, so things just don't change that much.
I do wonder though... once Direct Storage is widely used, will we start to see some really insane stuff involving RAM drives, or even dedicated RAM based solid state storage like the i-RAM? Of course, the big issue with that will be RAM capacity. High speed loading will allow developers to make games use even larger assets and they will get even bigger than they are now. It's really going to get absurd if you combine that with AI-generated assets, so they can basically be made infinitely huge\detailed without requiring a whole team to make each asset. Get ready for 500GB game installs becoming the norm. No RAM drives for those. 🙁
... though for competitive and e-sports titles where fast loading may give someone an edge and games tend to be a lot smaller in size, it wouldn't surprise me to see RAM drives being used for millisecond-level load times in Direct Storage equipped games. It sounds kind of absurd because level load times have stayed relatively consistent with game complexity for the past 20 years or so, despite the massive leaps in bandwidth in every single aspect of a modern computer... but once software can finally make use of it properly we're going to see some crazy stuff. I mean, even now (or soon) it's possible to have interfaces to our GPUs and storage devices with up to 64GB\sec bandwidth, SSDs coming out soon that are advertised at 13GB\sec reads, memory bandwidth on the RTX 4090 has broken 1Terabyte per second, and we've got CPUs with as much high speed L3 cache as a lot of computers had RAM 20 years ago. That's 128MB of cache with up to 2.5TB\sec of bandwidth according to AMD's info on the 7950X3D.
It's really bonkers if you think about it in comparison with the hardware we tinker with here on VOGONS.
At some point you'll see nVidia and AMD just release Video cards that have both Vram and large NVME based storage built into the card, when you install the game parts of it will be installed to the GPU storage so the GPU doesn't need to go through the system to access it and can stream it directly.
We are not quite there yet but I can see it happening in the next few generations as game streaming data gets too large to be streamed over the PCIe bus fast enough, we are already there with fucking shader compilation taking forever, imagine all of that being done directly on the GPU with zero need to involve the CPU or system bus.
Same with anything requiring AI hardware, why not just remove the bottleneck and have the GPU store and process it directly, nVidia cards already have the capability of handling NVME drives directly ..so do AMD IIRC, we really do need to move away from having the CPU/chipset deal with this shit when the GPU can do it and not require any extra input other than a handshake from the CPU telling it to go ahead.
I'm also skipping the 7000 series ..the 8000 series is actually looking like something really special and truly about moving forwards with their chiplet/fabric/cache tech, DDR5 also still needs a bit more cooking, its not yet at its theoretical 12,600 MTs speed yet but they are getting close so hopefully the next year or so will let them push it there. Also waiting on AMD to really push RDNA MCM further, they really didnt leverage what their tech guys are capable of and the 7000 GPUs seemed held back on purpose and a bit rushed to market.
It could be they held it back on purpose as what they have cooking in the labs needed more time and they want to really surprise us down the road but who knows, AMD works is mysterious ways with its hardware and fine wine policy.