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

Alvar "Miles" Udell

Respectable
Apr 1, 2020
573
305
2,260
0
Remember the Radeon Pro SSG? Given the bandwidth of PCIe 4 and 5 which will not restrict GPUs for some time to come, I wonder if we will start seeing GPUs with their own M.2 slots on the back? Imagine how useful it would be for a mATX build which only has 1 M.2.
 
Apr 10, 2020
75
13
35
0
Remember the Radeon Pro SSG? Given the bandwidth of PCIe 4 and 5 which will not restrict GPUs for some time to come, I wonder if we will start seeing GPUs with their own M.2 slots on the back? Imagine how useful it would be for a mATX build which only has 1 M.2.
Real benefit of VRAM is in its width, 384 bits compared to standard PC of 64 bit. SSD, NVMe, is just 2 serial channels. Also, why to extend GPU memory when manufacturer could charge you much more for VRAM.
As an idea for mATX might be ok but then, why not CPU too?
 

ITX God

Honorable
Mar 31, 2015
7
0
10,510
0
Remember the Radeon Pro SSG? Given the bandwidth of PCIe 4 and 5 which will not restrict GPUs for some time to come, I wonder if we will start seeing GPUs with their own M.2 slots on the back? Imagine how useful it would be for a mATX build which only has 1 M.2.
That's basically the idea with the PS5 SSD. It has a direct connection to the VRAM via the Kraken decryption chip. Future AAA titles will be designed with this architecture in mind.
 

epobirs

Distinguished
Jul 18, 2011
189
5
18,695
1
Crysis 3 is fairly ancient, coming out in early 2013 when the PS3 and Xbox 360 were still the core consoles in the market. If we want to talk about something modern that will seemingly use all the VRAM it can get, look to the latest release of Flight Simulator, with everything turned up to eleventy. This may be an outlier compared to most recent games but the same can be said for the first major games to demand 8GB, which nobody considers remarkable today.
https://www.tweaktown.com/news/74642/flight-simulator-at-8k-requires-16gb-of-vram-geforce-rtx-3090-needed/index.html
 

funguseater

Distinguished
Sep 20, 2009
605
8
19,065
49
Crysis 3 is fairly ancient, coming out in early 2013 when the PS3 and Xbox 360 were still the core consoles in the market. If we want to talk about something modern that will seemingly use all the VRAM it can get, look to the latest release of Flight Simulator, with everything turned up to eleventy. This may be an outlier compared to most recent games but the same can be said for the first major games to demand 8GB, which nobody considers remarkable today.
https://www.tweaktown.com/news/74642/flight-simulator-at-8k-requires-16gb-of-vram-geforce-rtx-3090-needed/index.html
The game is running on the GPU's memory, not the PC's RAM, that is what this is about, not how much vRam the game normally uses. I used to use these ram harvesting TSR's back in the day to skim 250KB of ram off my EGA card to use for DOS text mode. Every KB counted back then.
 
Reactions: spentshells

InvalidError

Titan
Moderator
Real benefit of VRAM is in its width, 384 bits compared to standard PC of 64 bit. SSD, NVMe, is just 2 serial channels.
The VRAM's bandwidth is of little importance for RAM-disk purposes since it ends up bottlenecked by the 4.0x16 bus' 32GB/s raw capacity, half of dual-channel DDR4-3200's bandiwdth. Also, NVMe on PCs is usually x4, not x2. Some motherboards do have NVMe lanes shared with SATA ports where using one of those SATA ports reduces NVMe from x4 to x2.
 
Reactions: JarredWaltonGPU

russell_john

Reputable
Mar 25, 2018
54
22
4,535
0
That's basically the idea with the PS5 SSD. It has a direct connection to the VRAM via the Kraken decryption chip. Future AAA titles will be designed with this architecture in mind.
Not exactly true, both the Sony and the Xbox are using the 4 lanes of PCIe 4.0 built into the Ryzen 3700 chiplet ..... The reason Sony's is faster sans compression is because it uses all 4 lanes and is the exact same speed as my NVMe M.2 drive (5 GB/s) while Microsoft went a different route using only 2 PCIe lanes for the internal SSD (2.5 GB/s) and chose to use the other two for their external NVMe drive meaning their external drive will be just as fast as the internal drive .... 5 GB/s is just a theoretical speed that is seldom if ever reached in practice and only for large files like HD video. Games and most programs on the other hand are made up of many smaller files and won't reach those speeds .... While both company's marketing (AKA Corporate Propaganda) push this theoretical nonsense like it was fact the actual truth of the matter the compression they are using is more for making large game files smaller so you can get more on the drive and again 2 times compression is the theoretical limit and is seldom reached in practice unless the files are simple text files or something like that ..... Another use is for the Series X fast resume, likely they'll image (a form of compression) the game memory onto the SSD and then stream it back into memory when needed and that is where you will see it shine as an image is one large file which is where you start hitting the theoretical top speeds

Will they be fast? Yes especially compared to the nearly obsolete hardware they are replacing but a lot of it is just marketing BS ..... Right now it's completely impractical to do texture streaming like Sony's marketing talks about and still have backwards compatibility to the PS4 or compatibility with any system other than the PS5 ..... Microsoft's implementation will prove to be better because they are making it cross platform so it will eventual work with both the Series X and the PC ...... It's no longer Xbox vs PS ..... It's Xbox AND PC vs PS and those combined markets are bigger than PS alone .... Even Sony is seeing the light which is why they are porting exclusives over to the PC or they'll get left behind because there is no money in selling consoles, the money is in selling the most games
 

Shadowclash10

Prominent
May 3, 2020
184
46
610
0
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.
I'm a bit confused here.... I was under the impression that games don't really benefit from super fast SSDs that much right now - they aren't optimized from it yet. Especially not a game as old as Crysis 3. So the loading times should really be equivalent to any SSD, not really faster than one?
 

escksu

Respectable
Aug 8, 2019
678
248
2,260
0
The article is wrong. It doesn't work that way. VRAM may be fast, but its only fast between GPU and VRAM.

The processing is still done by CPU. So what really happens is that the data in VRAM transfers to system RAM via PCIE (31.5GB/s for PCIE 4.0).... Then CPU reads the data from RAM and process it.

Even if you somehow bypass system RAM, you still can only communicate via PCIE.
 

epobirs

Distinguished
Jul 18, 2011
189
5
18,695
1
The game is running on the GPU's memory, not the PC's RAM, that is what this is about, not how much vRam the game normally uses. I used to use these ram harvesting TSR's back in the day to skim 250KB of ram off my EGA card to use for DOS text mode. Every KB counted back then.
And this method works for Crysis 3 because it has such low VRAM needs by today's standards. A lot of people have been reacting to the new cards as if they're wildly over the top designs when they are merely the leading edge for what will eventually be common. Considering how cheaply one can now build a 64GB PC, with denser DDR5 parts on the horizon, the use of RAM drives should make a comeback to greatly accelerate games that aren't going to be revised for new things such as RTX I/O any year soon.

DirectStorage on PC will be great once it's in wide use in 2-3 years but plenty can be done to mitigate the poor use of available resources right now. Developers have their hands tied by the desire to reach as large an audience as possible, which constraints their ability to fully exploit the modern PC. Building a RAM disk function into the Steam client could many thousands of existing games to be enhanced with little or no new work by the developers.
 

epobirs

Distinguished
Jul 18, 2011
189
5
18,695
1
The article is wrong. It doesn't work that way. VRAM may be fast, but its only fast between GPU and VRAM.

The processing is still done by CPU. So what really happens is that the data in VRAM transfers to system RAM via PCIE (31.5GB/s for PCIE 4.0).... Then CPU reads the data from RAM and process it.

Even if you somehow bypass system RAM, you still can only communicate via PCIE.
The place where a RAM disk could really shine for games in a relatively low cost system is the upcoming APUs using a shared pool of DDR5 and RDNA2. Late next year should be when this happens. There will likely be a generation of ZEN 3 paired with Vega APUs first before AMD gets that major upgrade to the GPU side in place, drawing in part on their console work.
 

escksu

Respectable
Aug 8, 2019
678
248
2,260
0
The place where a RAM disk could really shine for games in a relatively low cost system is the upcoming APUs using a shared pool of DDR5 and RDNA2. Late next year should be when this happens. There will likely be a generation of ZEN 3 paired with Vega APUs first before AMD gets that major upgrade to the GPU side in place, drawing in part on their console work.
The problem with this is the low memory bandwidth. Mainstream CPUs are limited to just 2 channels (total 128bits).
 

nofanneeded

Respectable
Sep 29, 2019
1,562
252
2,090
52
Consoles are using shared GDDR6 memory now and the previous one GDDR5 .. it seems it is easy to move the whole desktop from using DDR into GDDR .. I hope at least they consider this option in future CPUs ... This would uplfit the integrated graphics speed by two folds , and add more bandwidth without the need to go 4 channels or six channels.
 

Friesiansam

Distinguished
Feb 9, 2015
144
66
18,670
2
Remember the Radeon Pro SSG? Given the bandwidth of PCIe 4 and 5 which will not restrict GPUs for some time to come, I wonder if we will start seeing GPUs with their own M.2 slots on the back? Imagine how useful it would be for a mATX build which only has 1 M.2.
Slap an SSD on the back of a hot GPU, what a great idea...
 

CyranD

Distinguished
May 6, 2013
26
0
18,530
0
The article is wrong. It doesn't work that way. VRAM may be fast, but its only fast between GPU and VRAM.

The processing is still done by CPU. So what really happens is that the data in VRAM transfers to system RAM via PCIE (31.5GB/s for PCIE 4.0).... Then CPU reads the data from RAM and process it.

Even if you somehow bypass system RAM, you still can only communicate via PCIE.
I agree except the last statement IF you using something like RTX-IO. If you look at the slide deck from the RTX-IO presentation

Traditional pipe line is

Storage - CPU - System Memory - GPU - GPU memory

In this case you limited by PCIE and would probably be best off using a ram disk from system memory then GPU memory since system memory to CPU going to be faster then going through PCIE or in other words

System Memory (Ram disk) - CPU - System Memory - GPU - GPU memory

faster then

GPU memory (ram disk) - cpu - system memory - gpu - gpu memory

On other hand the pipe line for RTX-IO is

Storage - GPU - GPU memory

in that case you can avoid pcie completely in theory since it moving data within the gpu

GPU memory (ramdisk) - GPU - GPU memory
 
I'm a bit confused here.... I was under the impression that games don't really benefit from super fast SSDs that much right now - they aren't optimized from it yet. Especially not a game as old as Crysis 3. So the loading times should really be equivalent to any SSD, not really faster than one?
Well yes and no. If you go to the channel "Testing games" they compare load times of SATA 3 HDD, SATA 3 SSD, and NVME m.2 PCIe drives. There isn't that big an improvement.

Most of it comes from the "compile" stage of the game. A lot of code (shader code and sometimes IL) and textures are not in a finalized state when they are loaded off storage. So they have to be "compiled" into the final workable form.

What Sony and Microsoft consoles are doing is unloading compressed textures to VRAM and letting the GPU uncompress them, as the GPU is significantly faster than the CPU for this. It's an approach I admittingly didn't think of when I heard MS and Sony announce this tech. But I am impressed none the less.

This tech has been around for a while when AMD created their video/rendering editing work station cards where they had direct attached storage attached to the GPU. But the textures there were already decompressed. You just had faster access to them as they were arranged in a memory access optimized format. It's all part of heterogeneous memory access they have been working on for a while.

The console texture decompression is just a logical evolution. But I do give them credit for thinking of it.
 
Reactions: Shadowclash10

neojack

Reputable
Apr 4, 2019
450
78
4,840
44
just a question.

i read somewhere that VRAM is faster than system ram, because the specifications are more relaxed on error control. thus the VRAM is more prone to error (but it's not a big deal in most cases)
that's why whe have GDDR5-6 , while system's DDR generation is still DDR4

question : by using VRAM as file storage, isn't it more prone to corruption of the files ?
i guess it's only for reading in this scenario, so not relevant
but if it's the case, that's still something to be considered if someone has the idea to use a VRamDisk for a project.
 
just a question.

i read somewhere that VRAM is faster than system ram, because the specifications are more relaxed on error control. thus the VRAM is more prone to error (but it's not a big deal in most cases)
that's why whe have GDDR5-6 , while system's DDR generation is still DDR4

question : by using VRAM as file storage, isn't it more prone to corruption of the files ?
i guess it's only for reading in this scenario, so not relevant
but if it's the case, that's still something to be considered if someone has the idea to use a VRamDisk for a project.
Where did you get this idea?

VRAM is optimized for large sequential access and has optimized paths to the GPU which can be controlled. But GDDR6/6x VRAM also runs extremely hot and is pricey by comparison.

With a standard motherboard the timings are looser to be compatible with a large stack of RAM modules. They also run a lot cooler.

Because the gpu memory bus can be overclocked to the bleeding edge, there are redundancies put in there to make sure the data is valid. Thus if the checksum doesn't match, the GPU might go back and ask for the data again. So while you could really push the memory, you increase the likelihood of actually slowing the rendering down due to a data correction. But the optimal data rate is where there are no data errors.
 
Last edited:
While interesting they are using it as a RAM Disk. Given that it has to travel over the PCIe BUS to the CPU. Given the limits of PCIe and increases in latency. I'd think in reality load times would be slower than a traditional RAM Disk.

What a cool Idea! When we were still using single core Athalon chips, I used to make ram drives and install unreal tournament to it. Nice to see this has been improved upon.
That made me remember. I had a whopping 72MB RAM back in the day (8MB was standard). I'd load programs through the RAM disk to really speed up load times. What else was I going to do with all that memory. Even when taxing Photoshop I rarely went over 30MB.

Those were the good ole days. When you could specify how much memory a program could use.
 
Reactions: spentshells

vinay2070

Distinguished
Nov 27, 2011
225
40
18,720
3
While interesting they are using it as a RAM Disk. Given that it has to travel over the PCIe BUS to the CPU. Given the limits of PCIe and increases in latency. I'd think in reality load times would be slower than a traditional RAM Disk.



That made me remember. I had a whopping 72MB RAM back in the day (8MB was standard). I'd load programs through the RAM disk to really speed up load times. What else was I going to do with all that memory. Even when taxing Photoshop I rarely went over 30MB.

Those were the good ole days. When you could specify how much memory a program could use.
What proccy did you have?
 

ASK THE COMMUNITY