Gamer Installs Crysis 3 On GeForce RTX 3090's VRAM - And It Runs

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
10,785
"This is really cool to see, VRAM is the fastest memory-solution in your system, exceeding the performance of system RAM, so using it as an SSD should yield some amazing loading times for video games. And theoretically, you can do this with any graphics card, as long as the game fits inside your GPU's frame buffer.

However, the real-world application for this is quite small, today's NVMe SSDs load games very quickly making this use case very niche (plus additional software features like RTX I/O will reduce load times further). But it's still a cool concept, perhaps giving us a little sneak peek into what SSDs of the future will be like for gaming."


https://www.tomshardware.com/news/gamer-installs-crysis-3-on-geforce-rtx-3090s-vram-and-it-runs
 
I didn't realize that you could make a VRAM disk. That's really cool!
YDL on the PS3 does this natively, as it takes the 256MB of VRAM and adds it to the SWAP as fast-SWAP.
This is a pretty cool use-case scenario, though, and while it would have higher latency than system RAM, it would have much faster memory bandwidth, and loading games should be much improved!
 
Isn't the future NVME drives that plug into your graphics card?
 
The future is here! One day we can use nothing more than a graphics card for a computer and run everything off the onboard ram. And then later on someone will develop an add in graphics processor to take the load off the gpu.
 
Most of the load times you see in game are actually compile times or decompression times. Fast storage is only going to get you so far with video games nowadays. Don't get me wrong it's still very very good to have but there is definitely a limit to how fast you can load in the game.
 
Most of the load times you see in game are actually compile times or decompression times. Fast storage is only going to get you so far with video games nowadays. Don't get me wrong it's still very very good to have but there is definitely a limit to how fast you can load in the game.

I think most load times are because games were designed around mechanical drives. Look at the new Ratchet and Clank game for PS5. Basically no load times, even when changing the entire game world through portals multiple times.
 
Isn't the future NVME drives that plug into your graphics card?
That's more or less what the point of the DX12u Direct Storage stuff, I mean it isn't physically plugged into the GPU but having direct I/O access to it bypassing the OS on nvme drives running PCIE4 or god forbid 5 will be like 98% as fast as a direct connection.
 
Technically SRAM is by far the fastest. Just too small to do anything like this with.

MS Flight Sim could certainly benefit from this if it would fit. It takes a while even on my nVME drive.
 
Nah, old games had load screens even with ROMs.

Not all old games did. But the ones that did, had storage hardware limitations. NES generally doesn't have a massive read speed difference between ROM and RAM, and games often had onboard RAM to compensate. Genesis generally wasn't known for load times except a few games.

Arcade games often had really slow ROM chips. Storage limitation. SNES ROMs are much slower than its RAM, so many SNES games had loading screens. Again. Storage limitation.

In other words, old games were still coded around hardware storage limitations of the day.

Developers are literally working on eliminating load times due to consoles having SSDs by default. Loading any game on PC on your SSD right now is evidence of nothing at all. Because the game was coded around crap storage speed and you installing the SSD doesn't magically re-code the source code and recompile it. We need to wait for XBS/PS5 ports to PC to really see the improvement, assuming the devs don't make crap ports assuming PC users are still using mechanicals.

"For example, when you open up a door you would have a delay between the door actually opening up, and that was streaming done in the background," says Hacura. Playing System Redux on a Series X or PS5, that door might just swing open without a noticeable pause.

"It requires a little bit of custom work," Hacura continues, "but it is usually much easier to remove hacks that were put into a game because of a lack of resources than to add those when you lack resources. It's always nicer to go this way."

https://www.usgamer.net/articles/ps5-xbox-series-x-next-gen-ssd-analysis
 
I think most load times are because games were designed around mechanical drives. Look at the new Ratchet and Clank game for PS5. Basically no load times, even when changing the entire game world through portals multiple times.
The Ratchet and Clank showing still has some disguising going on. Its essentially the same as a magician holding up a sheet so that stuff can happen behind it. But its early days. It it still very impressive and they were smart about it by keeping movement on screen and making it look and feel like your character is actually going somewhere else. Of course, Naughty Dog and Insomniac already mastered that back in PS2 days. Here the difference is its way faster. In the past, you'd be in that portal warp, or an elevator, or an airlock or an.... for 10-15 seconds. As time goes on, they will figure out ways to actually appear seamless, with no obvious curtain.
 
The Ratchet and Clank showing still has some disguising going on. Its essentially the same as a magician holding up a sheet so that stuff can happen behind it. But its early days. It it still very impressive and they were smart about it by keeping movement on screen and making it look and feel like your character is actually going somewhere else. Of course, Naughty Dog and Insomniac already mastered that back in PS2 days. Here the difference is its way faster. In the past, you'd be in that portal warp, or an elevator, or an airlock or an.... for 10-15 seconds. As time goes on, they will figure out ways to actually appear seamless, with no obvious curtain.

I feel like you're splitting hairs at this point. PS5 can fill the entire RAM in 1.8 to 3 seconds. Its than even faster than the new Samsung 980 Pro, with compression taken into account. Even high end PCs won't catch up until Direct Storage and PC games are coded to use it. Will there be pauses from time to time to load data? Yes. But they will be so short, people soon won't be using the term loading screens anymore because games will be coded with extremely high speed storage with zero access times in mind.

“What took 15 seconds now takes less than one: 0.8 seconds, to be exact,” Wired wrote of seeing Cerny demo a bit of fast travel around Spider-Man’s Manhattan-based map. This investor day video is basically a chance for you to see this for yourself.

https://www.vg247.com/2019/05/21/ps5-video-spider-man-backwards-compatibility-next-gen-hardware/
 
I feel like you're splitting hairs at this point. PS5 can fill the entire RAM in 1.8 to 3 seconds. Its than even faster than the new Samsung 980 Pro, with compression taken into account. Even high end PCs won't catch up until Direct Storage and PC games are coded to use it. Will there be pauses from time to time to load data? Yes. But they will be so short, people soon won't be using the term loading screens anymore because games will be coded with extremely high speed storage with zero access times in mind.

“What took 15 seconds now takes less than one: 0.8 seconds, to be exact,” Wired wrote of seeing Cerny demo a bit of fast travel around Spider-Man’s Manhattan-based map. This investor day video is basically a chance for you to see this for yourself.

https://www.vg247.com/2019/05/21/ps5-video-spider-man-backwards-compatibility-next-gen-hardware/
I don't think its splitting hairs, exactly. 0.8 seconds to fill RAM is blazing fast, no doubt. But, also sort of a paper spec. Is it practical to be able to display a fully featured scene in 1 second? probably not? I think the ratchet and clank portal wraps being the same exact thing every time, were as much a technical work around as they are a presentational choice. And by that I mean that if they really could, they would probably just tear the screen away to reveal a new world. as if you were revealing something covered by a sheet or pulling back a curtain.

Instead, they put you into the portal warp transition, which looks exactly the same every time, probably because its always sitting in memory, waiting to be called upon. And the amount of time it takes Ratchet to fall through this warp transition, is the exact amount of practical time it takes them to not only load the new data into ram, but then prepare the next scene to actually be looked at.
 
The future is here! One day we can use nothing more than a graphics card for a computer and run everything off the onboard ram. And then later on someone will develop an add in graphics processor to take the load off the gpu.
The Intel phis were pretty close to that madness
 
YDL on the PS3 does this natively, as it takes the 256MB of VRAM and adds it to the SWAP as fast-SWAP.
This is a pretty cool use-case scenario, though, and while it would have higher latency than system RAM, it would have much faster memory bandwidth, and loading games should be much improved!

It can be done on any Linux distro with at least kernel version 2.6.23, and it's not limited to just swap space. Once you turn the addressable memory into a block device, you can use it as a RAM disk as well.

https://wiki.archlinux.org/index.php/Swap_on_video_RAM

While it specifies video memory, it can technically be any memory that's present on the bus and mapped somewhere in the x86 address space. I've done this with PCI, AGP and PCIe video cards, as well as a few RAID controllers and some other misc peripheral cards with some significant amount of memory present and it generally works. The speed you get is variable, with AGP being the slowest because it's an asynchronous bus with most bandwidth going to the card and hardly anything going back to the system. PCI maxes out at 133 MB/s unless the slot is 66 MHz, or a PCI-X slot, like a RAID controller which can have more bandwidth.

Speed also depends how the RAM on the card is handled. Most video cards have their VRAM connected directly to the GPU, so there is added latency and slowdown from the GPU having to be a bridge between the system bus and the VRAM.

One thing that's fun to do is figure out where the active frame buffer is on the video card, you can write a bit of code in bash script to write stuff directly to the screen.
 
It can be done on any Linux distro with at least kernel version 2.6.23, and it's not limited to just swap space. Once you turn the addressable memory into a block device, you can use it as a RAM disk as well.

https://wiki.archlinux.org/index.php/Swap_on_video_RAM

While it specifies video memory, it can technically be any memory that's present on the bus and mapped somewhere in the x86 address space. I've done this with PCI, AGP and PCIe video cards, as well as a few RAID controllers and some other misc peripheral cards with some significant amount of memory present and it generally works. The speed you get is variable, with AGP being the slowest because it's an asynchronous bus with most bandwidth going to the card and hardly anything going back to the system. PCI maxes out at 133 MB/s unless the slot is 66 MHz, or a PCI-X slot, like a RAID controller which can have more bandwidth.

Speed also depends how the RAM on the card is handled. Most video cards have their VRAM connected directly to the GPU, so there is added latency and slowdown from the GPU having to be a bridge between the system bus and the VRAM.

One thing that's fun to do is figure out where the active frame buffer is on the video card, you can write a bit of code in bash script to write stuff directly to the screen.
Well hot damn, how did I not know that.
Awesome, thanks for the information, it is much appreciated! (y)
 
For fast streaming to GPU VRAM without being limited by PCIe.....

Having flash memory directly on the card wouldn't really improve performance because the flash memory is the speed bottleneck. A 16x PCIe 3.0 link is 15.75 GB/s and even the fastest M.2 SSD has trouble getting to half of that. Sure there are super fast discrete PCIe SSDs, but those are very expensive and out of most consumer's price range.
 
Back
Top