Dedicated GPUs, onboard GPUs... and then there are half-and-half GPUs?

[Pawnless Endgame]

Dedicated - means that the card has its own memory.

Onboard - means it has no memory and it uses some of the system RAM.

6 years ago, I had a laptop with a GTX 960M, sold as having 4GB. That was a pretty decent gaming spec for a laptop and it lasted me well. Just before lockdown, I got a new laptop with a GTX 1660 Ti, sold as having 6GB.

I listed the GTX 960M laptop on eBay, and after it sold, my buyer said that the 960M only had 2GB. I had screenshots saved from device manager showing it as 4GB. The buyer then responded with screenshots from CPU-Z, showing it as being 2GB dedicated and 2GB shared memory. I.e. a half-and-half card. The buyer feels cheated, quite rightly so, but I felt cheated too because I have been mis-sold a GPU. It's not an OcUK laptop by the way. It was a house insurance handout that could only be spent in a certain retailer! I offered the buyer either a return for a full refund or a partial refund. He accepted a partial refund and kept the laptop and he gave me good feedback.

I had a quick panic about my new GTX 1660 Ti laptop, so I ran CPU-Z on that, and thankfully it does show the full 6GB. Both 960M and 1660 Ti systems had 16GB system RAM btw.

So just wondered: are half-and-half cards that common? What's the right word for this type of card, and how's the best way to not being mis-sold in the future?

 

[Sasahara]

I'd be very surprised if it was sharing ram, as that isn't a thing. Ram sharing only works when all the ram is from the same source.
It could be a 2GB card. And CPU-Z is showing the 2gb shared ram for the onboard graphics for when you enable that.

CPU-Z could well be just wrong, all it's doing is looking up a database, it's not actually checking the ram.

Get him to run a benchmark like Unigine Heaven 4.0 / superposition 2017 or any high end game.

Use MSI afterburner to display the GPU RAM usage, 1080p with everything max should be over 2GB.

If it shows more than 2gb usage, it's a 4 GB card.

 

[Pawnless Endgame]

Thanks

Yes I've only known them as onboard or dedicated cards, not half-and-half / hybrid types. I fear it's too late for me to ask the buyer to run tests as I have already given him good feedback, but it will be interesting to see what the tests look like on my 1660 Ti. Also, to see how it compares with desktop cards. Bear in mind it's a 1660 Ti (without the "M"), so it might well be a desktop card inside of a laptop, maybe stripped down somewhat to fit inside.

 

[Heinz57]

ATI and Nvidia used to do this years ago, branded as hypermemory and turbocache. The cards were advertised as 256mb or whatever, but it was actually only a small amount of onboard memory, the rest being supplemented by system RAM.

https://en.m.wikipedia.org/wiki/TurboCache

It's possible they still do it I suppose!

 

[Sasahara]

ATI and Nvidia used to do this years ago, branded as hypermemory and turbocache. The cards were advertised as 256mb or whatever, but it was actually only a small amount of onboard memory, the rest being supplemented by system RAM.

https://en.m.wikipedia.org/wiki/TurboCache

It's possible they still do it I suppose!

Ah yes turbocache it looks like is didn't have a long shelf life, mainly as it was pretty bad and only used for the lower end 6200 and GeForce Go 7200/7400 GPUS.

Certainly as GPUs start using modern faster GDDR Ram, using system ram would completely cripple the performance. If it's even possible to use GDDR and DDR in that way, I suspect not as they very different.

 

[Sasahara]

Thanks

Yes I've only known them as onboard or dedicated cards, not half-and-half / hybrid types. I fear it's too late for me to ask the buyer to run tests as I have already given him good feedback, but it will be interesting to see what the tests look like on my 1660 Ti. Also, to see how it compares with desktop cards. Bear in mind it's a 1660 Ti (without the "M"), so it might well be a desktop card inside of a laptop, maybe stripped down somewhat to fit inside.

From the 10 Series onwards, as far as I know all mobile GPUS are the same chip as the desktop, but maybe run at a lower clock speed. So they are only a little bit slower than their desktop counter parts. Before that the m versions could be half the performance of the desktop versions.

 

[d_brennen]

There were some 2GB 960M laptops that could share up to 2GB system RAM.

I'd be very surprised if it was sharing ram, as that isn't a thing. Ram sharing only works when all the ram is from the same source.
It could be a 2GB card. And CPU-Z is showing the 2gb shared ram for the onboard graphics for when you enable that.

CPU-Z could well be just wrong, all it's doing is looking up a database, it's not actually checking the ram.

Get him to run a benchmark like Unigine Heaven 4.0 / superposition 2017 or any high end game.

Use MSI afterburner to display the GPU RAM usage, 1080p with everything max should be over 2GB.

If it shows more than 2gb usage, it's a 4 GB card.

If it shows more than 2GB usage, it's using system RAM

 

[Sasahara]

There were some 2GB 960M laptops that could share up to 2GB system RAM.



If it shows more than 2GB usage, it's using system RAM

No modern GPUS do not use system ram as video RAM.

 

[shankly1985]

No modern GPUS do not use system ram as video RAM.

They definitely do if you run out of VRAM the GPU will use system ram.
If you run out of Ram then the system will go to the hard drive or ssd.

Failing all that the computer will either lock up or BSOD.

 

[d_brennen]

No modern GPUS do not use system ram as video RAM.

Except when they do You have the internet, you have the ability to look up the 960M 2GB that could share 2GB system RAM

 

[Pawnless Endgame]

Some very interesting posts from all of you. Thanks again

I didn't know about Turbocache, so it seems that Nvidia took a step back with the 960M, and another post on here confirmed that there was a 2GB version of the 960M too.

I ran Unigine Heaven 4.0 on my 1660 Ti and got a steady 60 FPS (110 max FPS) @ 1920x1200 on ultra settings. Breathtaking scenes of suspended fantasy towns/cities linked with bridges and airship, with some steampunk thrown in

About a desktop vs laptop 1660 Ti, I found out on Passmark that it does exist in both versions. The desktop card scores 12802 and the laptop version scores 10215.

I thought I'd then try running Passmark itself on my laptop and my card scored somewhere in between the 2 scores. Just shy of 11000. So well below 12802 yet well above 10215! I guess it depends on how the cards have been calibrated.

 

[Sasahara]

Except when they do You have the internet, you have the ability to look up the 960M 2GB that could share 2GB system RAM

Show a link that confirms a 960m has Turbocache? I can find no such card, but would be happy to be proven wrong.

What I have found there are 2GB and 4 GB versions. And they can't share system ram with their own GDDR5.
https://en.wikipedia.org/wiki/GeForce_900_series#GeForce_900M_(9xxM)_series

The onboard (probably intel) graphics can share up to 2GB or ram, that is not the 960m using system ram. I'm not sure if that 2GB has to be pre allocated, I know my HD630 laptop doesn't and will use ram as it needs it and the HD 630 has a maximum VRAM limit of 64GB, but don't know any applications that could use that amount on such a low end graphics if I had that amount of system ram installed.

So depending on the how the laptop is configured it could use the onboard GPU for 1 screen and the 960M for another. So the laptop could use 4 GB vram in total, but not on the same application

 

[Sasahara]

I thought I'd then try running Passmark itself on my laptop and my card scored somewhere in between the 2 scores. Just shy of 11000. So well below 12802 yet well above 10215! I guess it depends on how the cards have been calibrated.

The chip is the same, just the clock speed maybe different as it's harder to cool the chip inside a small laptop. This depends on the laptop design, some laptops with high end cooling can match a desktop GPU, but they can be noisy.
So the only reason for a lower performance is the max boost clock is lower than the desktop chip and in some cases the power limit. They may limit the power consumption so you don't overload the laptops PSU.

https://en.wikipedia.org/wiki/GeForce_10_series#GeForce_10_(10xx)_series_for_notebooks
https://en.wikipedia.org/wiki/GeForce_16_series#Products

 

[Sasahara]

They definitely do if you run out of VRAM the GPU will use system ram.
If you run out of Ram then the system will go to the hard drive or ssd.

Failing all that the computer will either lock up or BSOD.

When a GPU uses shared GPU memory it's using that for basic tasks like a swap file for the video card a memory cache.

It's not useable in the same way as dedicated VRAM and shouldn't be reference as that. My 1080ti isn't a 27 GB card using 16 GB of shared system ram.

Most games or high end applications will crash or slow to unusable rates if you go over the dedicated VRAM limit. When the game runs out of addressable VRAM it will crash or in some cases lower the settings on the fly.

This is a reasonable explanation.
https://linustechtips.com/main/topi...-eats-my-ram/?do=findComment&comment=12550393

One simple example is Ethereum mining, there was a time when the amount of VRAM ram required was reaching over 2GB to mine this, once it hit this limit any 2 GB card was unable to mine, you could not use the "shared" GPU memory as VRAM.

 

[chaparral]

If i remember right the 4GB 970 GTX shared some the system ram

 

[Justintime]

If i remember right the 4GB 970 GTX shared some the system ram

No, from what i remember and could be explaining it wrong the last 512mb or something was segmented differently and ran slower than the previous 3.5gb but it did have 4gb physical RAM.