NVIDIA announces RTX IO, GPU to Directly Access SSD
Click here to post a comment for NVIDIA announces RTX IO, GPU to Directly Access SSD on our message forum
Maddness
Looking forward to your review of these new cards and technology. Interesting times ahead.
UnrealGaming
this one's really important. the one super impressive thing about the next gen consoles is now, in a way, available on pc.
Undying
Ps5 we have instant loading!
Nvidia hold my beer...
Noisiv
I wouldn't go that far 🙂 Need to see it in action
Caesar
Hmmmmm...
That's why there's the 24GB version.....?
8-10GB for graphics...and the rest acts as "system memory" for the RTX I/O mechanism
entr0cks
One of the more interesting parts of yesterday's event, GPU calls are quite expensive.
This removes some unnecessary CPU/RAM overhead/administration - more bandwidth available for other things. Probably using the Tensor cores for decompressing.
Eager for the reviews on Guru3D 🙂
PrMinisterGR
Wait, so now that Nvidia does it nobody will say how "fake" and "useless" it is?
Fox2232
Just do not OC your GPUs too much, or face data corruption.
@PrMinisterGR Why would it be useless, MS made DirectStorage. And nVidia's solution works on that.
MS did not make it out of thin air. They were working on it for XSX.
But problem remains. How does one boost decompression of file that's handled purely by game with poor code that can use only single CPU core and limits decompression to 200MB/s? One does not.
Will take years before games will take use of it on PC.
PrMinisterGR
There is no AAA game that has such a bad engine. The limit is memory operations and CPU power. This will handle both. It won't take years, all AAA console titles will work this way, and with the Xbox in the mix the code will be in DirectX. There will be no great gap this time between generations, people already know how to work with them. It famously takes under a month to have an engine running fairly well on the PS5, from basically nothing.
Fox2232
There are compression methods that can do 900MB/s per thread or more which are pretty close in compression ratio to those best in class.
But I have seen games that use 7zip library and encrypt lzma.
I was talking about existing games which is pretty much apparent considering I wrote about darn poor data handling methods on PC.
PrMinisterGR
But I don't even think that current games are that bad, honestly. The problem is that on PC we need to do too many memory operations. Essentially a texture travels too much to be where it needs too.
nevcairiel
RTX IO is basically an evolution of "GPUDirect Storage" that NVIDIA already revealed for AI/HPC last year
https://developer.nvidia.com/blog/gpudirect-storage/
There is rarely something "unique" that anyone comes up with, since all of these companies are working on solving the same problems.
Fediuld
Exactly. All those months when AMD was doing it, the hypocrites were saying is useless and not needed.
Now Nvidia does it the best feature ever. Ofc they forget AMD will do it even better on PC given that will go through IF2
Mufflore
Its the other way round.
This will help cards with less VRam by allowing faster replenishing of data while using less system resources too.
It uses a direct channel from SSD to VRam, bypassing system ram and the buses connecting.
System ram use will drop, more bandwidth will be available for the CPU.
It has been mentioned as a PCIE 4.0 improvement but it might hopefully work on lesser tech.
schmidtbag
Mmmm sure is nice seeing that, considering the people arguing with me saying it wasn't gonna happen any time soon.
Fox2232
Imagine option to compress any folder through DirectStorage instead of standard NTFS/compact solution in windows. That would have immediate impact.
But DirectStorage is not going to come for most of existing games. And not every new games will jump on it due to compatibility/HW requirements. (Older HW may reap no benefits. Majority of current GPU owners have that older HW.)
schmidtbag
I don't think it has to. For games that don't support DS, it could simply act as a giant prefetch file. Such games won't be anywhere near as optimized, but they'll still make a tremendous difference in load times.
Think of it like enabling SLI on an unsupported title - it's not going to be as good and it might require some extra tweaking, but it will yield a performance improvement.
Besides, most existing titles load rather quickly on high-end SSDs anyway. Games with textures ready for 8K aren't all that common yet, so, there's still time for this technology to be leveraged.
geogan
Happy to see this tech on PC. The PS5 version of this is the one thing that will make next-gen games really feel next gen. PC would have been way behind without it.
Fox2232
I used to have storage statistics via MSI Afterburner. So I kind of know at what rate games loaded data and when. And games gained almost nothing from my NVMe drives.
I got big advantage with virtual machines and loading with Unreal Engine 4 Editor. But games are generally limited by CPU to point that their loading mostly does not saturate SSD.
You can check task manager for active time of your SSD when you start loading game, and when you move past initial loading in menus into loading save game and level.
Loading of Wither 3 (while fast) does not exceed 25% active time of NVMe (S40G) which is connected via X470 (PCIe 2.0 x4).
Initial loading of Vermintide 2 (which takes a lot of time) does not exceed 20% active time on NVMe (SX8200NP) that's in 3.0 x4 and averages at 9%.
That means drives are not choked by random access. There are simply no requests to read more data in given time interval.
wavetrex
PCI to PCI device communication has been a thing since ages ago (as old as Pentium 1)
All PCI (and by extension PCIe) devices have virtual memory mapping, especially in x64 systems the hardware can be mapped to some address outside the system RAM address, and then Device 1 can just grab data from Device 2, if both support DMA (which they do).
The CPU only gives commands to the PCI devices where to look and what they are allowed to access. (But that shouldn't be an issue since both NVMe driver and GPU driver are Kernel drivers.)
Nvidia "simply" made a connection between the two, and Microsoft is probably involved as well so I'm guessing an update to the OS will be needed in order to permit the two drivers to talk to each other.
