A Query regarding V-RAM Behaviour

[apparentlyarhm]

Just a simple query and I apologize in advance, if my question is dumb.

Lets suppose there's a PC A with 8GB DDR3 RAM with some CPU and apparently other components. if I slam in a card, say a GT 210 with 1GB DDR3 memory, so shouldn't the card take some memory from the system, I mean, RAM's also DDR3, and essentially faster (lets suppose the RAM is 1666MHz; because i had owned a GT210 in the past, with memory clocks 500ish MHz ( MSI Afterburner reported this))

I know that a GPU "prefers" a memory right next to it, but it should theoretically apply to GDDR5 and DDR3? In my example, both RAM and V-RAM is DDR3? Please clarify my doubts and I again apologize if its a dumb question.

[ProjectBox153]

The GPUs memory and the system memory are separate. I don't know if they share because they're doing completely different things. 

[apparentlyarhm]

I am wondering.....

If the GPUs would actually share memory, would it make a difference? (stick to my PC example)

[Levent]

I believe shared memory side of things are handled by the driver, also yes it will work (as long as that card is supported for shared memory)

[partymarty]

It works as far as I know this way:

The textures get loaded onto the VRAM.

When the card runs out of VRAM, it will start using your system ram... System ram is slower to access, so you will see your performance taking an heavy hit at that point.

When the the system runs out of ram, it will start using the Pagefile on your ssd/hdd, which will further decrease the performance (because your hdd/ssd is a lot slower to access than ram).

 

The further down the line you get, the more bandwidth get limited + latency will go up.

[Sakkura]

6 minutes ago, Muffi said:

I am wondering.....

If the GPUs would actually share memory, would it make a difference? (stick to my PC example)

It could, depending on the workload. AMD have done a lot of work on this with their APUs, but they are probably too small to really get it to take off. And Intel hasn't had reason to push the concept as they've traditionally been a lot slower than AMD on the GPU side.

 

https://en.wikipedia.org/wiki/Heterogeneous_System_Architecture

 

It's much more difficult with dedicated GPUs though, especially when you can swap those out at will.

 

 

[apparentlyarhm]

So lets say i have a high performance NVMe ssd and 8GB DDR3 ram, and a gt 210 (DDR3 Vram) (IK it doesnt make sense). So essentially theres no bottleneck from other components

in that case, if the GPU starts sharing, would it be noticeable? (latency, etc)?

 

[partymarty]

1 minute ago, Muffi said:

So lets say i have a high performance NVMe ssd and 8GB DDR3 ram, and a gt 210 (DDR3 Vram) (IK it doesnt make sense). So essentially theres no bottleneck from other components

in that case, if the GPU starts sharing, would it be noticeable? (latency, etc)?

 

Oh yeah. It is definitely noticeable. Even a NVMe SSD doesn't nearly have the bandwidth that ram or vram has

[apparentlyarhm]

and lets say i have sufficient RAM, say 16GB DDR3, so there would be no need for pagefiles, what would happen in that case?

 

[partymarty]

Just now, Muffi said:

and lets say i have sufficient RAM, say 16GB DDR3, so there would be no need for pagefiles, what would happen in that case?

 

As soon as the GPU starts loading off textures to the RAM most games become nearly unplayable because of the massive slowdowns in throughput which results in stuttering and huge fps drops.

Ram is like 5 times slower than VRAM + the latency is a lot higher.

 

To make a game playable often the solution is to turn down the settings until textures won't get loaded onto the RAM anymore.

[Sakkura]

3 minutes ago, Muffi said:

So lets say i have a high performance NVMe ssd and 8GB DDR3 ram, and a gt 210 (DDR3 Vram) (IK it doesnt make sense). So essentially theres no bottleneck from other components

in that case, if the GPU starts sharing, would it be noticeable? (latency, etc)?

 

If it has to go over the PCIe connection, it will be noticeable. Maybe not so much with the GT 210 because it's already very slow, but certainly with faster cards it would be crippling.

 

But if you had a faster connection between them, it could work. You'd just need faster system memory, since GPUs tend to need a lot more memory bandwidth than CPUs.

 

This is exactly how things work in current consoles.

[Sakkura]

1 minute ago, partymarty said:

As soon as the GPU starts loading off textures to the RAM most games become nearly unplayable because of the massive slowdowns in throughput which results in stuttering and huge fps drops.

Ram is like 5 times slower than VRAM + the latency is a lot higher.

 

To make a game playable often the solution is to turn down the settings until textures won't get loaded onto the RAM anymore.

System memory types like DDR3 and DDR4 actually tend to have lower latency than VRAM memory types like GDDR5 etc. But the bandwidth can be an issue.

 

But you can use unified memory, like the 8GB GDDR5 in the PS4 or the 12GB GDDR5 in the Xbox One X.

[apparentlyarhm]

5 minutes ago, partymarty said:

As soon as the GPU starts loading off textures to the RAM most games become nearly unplayable because of the massive slowdowns in throughput which results in stuttering and huge fps drops.

Ram is like 5 times slower than VRAM + the latency is a lot higher.

 

To make a game playable often the solution is to turn down the settings until textures won't get loaded onto the RAM anymore.

even if the RAM has the same Architecture? (say both are DDR3s)(gt 210)

[partymarty]

1 minute ago, Sakkura said:

System memory types like DDR3 and DDR4 actually tend to have lower latency than VRAM memory types like GDDR5 etc. But the bandwidth can be an issue.

 

But you can use unified memory, like the 8GB GDDR5 in the PS4 or the 12GB GDDR5 in the Xbox One X.

I think the latency added by using system memory will already give you a fairly big performance penalty.

[Mira Yurizaki]

17 minutes ago, Muffi said:

and lets say i have sufficient RAM, say 16GB DDR3, so there would be no need for pagefiles, what would happen in that case?

Applications could "run out" of memory, even though the amount of physical RAM being used isn't 100%.

 

When the application asks for memory, it often asks for more than necessary for potential growth. So it asks the OS to reserve some amount of memory, even if the application doesn't actually use it. Say you disabled the page file and all of the applications reserved memory totals physical RAM. When an application asks for more memory, the OS looks to see if there's any space left it can reserve. Since there is no more space and the OS will not boot an application's reserved memory space, the OS returns with "there's no more memory left" and the application will complain the system is "out of memory." The page file exists not only to store data when RAM gets full, but to also basically act as a buffer for reserved memory spaces that aren't actually in use.

 

This is the reason why it's almost never a good idea to turn off the pagefile.

[partymarty]

6 minutes ago, Muffi said:

even if the RAM has the same Architecture? (say both are DDR3s)(gt 210)

I think it's not the problem that one of those components is slower.

I think the problem is in 'how' the gpu gets it's data. The data needs to be send to the ram + being received from the ram. So the 'travel distance' of the data is higher when the data is send/received from elsewhere in your pc than just your gpu. There's probably a driver too which starts to manage some things, which often also gives some overhead.

 

The problem is not in how fast those things might be. It's in how the connection between those 2 parts works.

 

This stuff is not my expertise, I can be completely wrong with my assumption. But I think it works this way... Would love to be corrected if I'm wrong about it  

[apparentlyarhm]

2 minutes ago, partymarty said:

I think it's not the problem that one of those components is slower.

I think the problem is in 'how' the gpu gets it's data. The data needs to be send to the ram + being received from the ram. So the 'travel distance' of the data is higher when the data is send/received from elsewhere in your pc than just your gpu. There's probably a driver too which starts to manage some things, which often also gives some overhead.

 

The problem is not in how fast those things might be. It's in how the connection between those 2 parts works.

 

This stuff is not my expertise, I can be completely wrong with my assumption. But I think it works this way... Would love to be corrected if I'm wrong about it  

well, your assumption seems legit, at least to me! :-)