VRAM size - what about the other specs?

ninjag · 21 Aug 2025 at 10:26

Was just wondering about VRAM and the focus tends to be on the amount of VRAM being priority. But can a lesser amount of VRAM be ok if the memory interface width and bandwidth is higher?

For example, 8GB VRAM with a 256-bit interface and 448.0GB/s bandwidth vs 12GB VRAM with a 192-bit interface and 360.0GB/s bandwidth. Which would perform better, does application matter etc?

Ch3m1c4L · 21 Aug 2025 at 10:31

it somewhat depends on application, but the main problem is fitting everything into the amount of VRAM. the point is that at a certain point all the current needed textures etc dont fit into 8GB, so you have to have 12GB.
The fast 8 gb vs slow 8gb would be better for when the engine can stream in higher res textures, but its still going to be reliant on your storage subsystem/system RAM for this.

Just look at AI performance. if it fits into your VRAM it tends to run image generation/text that is counted in seconds, or minutes. As soon as it can't fit into VRAM and needs to swap things in and out, that goes up to hours/days

Smeghead · 21 Aug 2025 at 12:32

Assuming you're talking about gaming, I think there are too many variables to purely look at the GPU specs on paper to evaluate a card's performance. The overall system, from CPU, RAM, hard drive/SATA SSD/nvme and the application/game optimisation and single core vs multi core performance etc. all collectively makes a GPU's raw specs a somewhat useless metric.

That being said, I think one card's slightly bumped up compensation specs with less VRAM is unlikely to give you a huge difference compared to the one with more VRAM. Probably in the single percentile differences. I'm kind of pulling that out of no where though, so happy to be corrected on this.

BigBANGtheory · 21 Aug 2025 at 12:56

basically (according to my understanding)...
- the VRAM amount effects the performance ceiling i.e. you run out of VRAM and performance drops off a cliff presumably because it has to load it in from system RAM or storage
- bandwidth and speed really comes into play the higher the resolution you go or conversely the lower the resolution the less it matters where you really notice a difference is powering the render resolutions for high end PCVR

Hotwired · 21 Aug 2025 at 13:29

There's a lot of frothing about 8GB cards existing and youtube channels have been baiting views by doing videos showing that if you try to use more than 8GB of memory on an 8GB card it won't run as well. Cutting edge research.

Since the dawn of time, gamers have been aware that graphics settings can be lowered to improve game performance. Apparently in 2025 this is now an alien concept.

However if making the workload suitable for the card is not an option then ideally you magic up some more money and get a card with enough memory to suit the workload.

SpudMaster · 22 Aug 2025 at 02:16

Ok so I feel like all the previous answers have just over complicated what is a relatively simple answer.

The card with the faster VRAM and bus etc will perform better than a card with slower VRAM/bus even if it has more vram providing the cards vram capacity is not exceeded. Once VRAM capacity is exceeded then performance will drop considerably. Faster VRAM and larger bus will not allow for more data to be stored in the same about of VRAM

ninjag · 22 Aug 2025 at 12:12

SpudMaster said:
Ok so I feel like all the previous answers have just over complicated what is a relatively simple answer.

The card with the faster VRAM and bus etc will perform better than a card with slower VRAM/bus even if it has more vram providing the cards vram capacity is not exceeded. Once VRAM capacity is exceeded then performance will drop considerably. Faster VRAM and larger bus will not allow for more data to be stored in the same about of VRAM

This is the part I was wondering about. If it's going slightly over the capacity, can the faster bus etc not clear it etc quick enough to compensate? But I'm guessing that at best that might only be the case if it's only going very slightly over, whereas several GB over would be a problem?

Hotwired · 22 Aug 2025 at 12:29

AMD tried that.

Fury X, a 4GB HBM card with a 4096 bit interface, at a time when alternatives ranged from 4 to 12GB.

It was "ok" but AMD decided HBM was too good for gaming peasants and we never saw it again.

SpudMaster · 22 Aug 2025 at 13:11

ninjag said:
This is the part I was wondering about. If it's going slightly over the capacity, can the faster bus etc not clear it etc quick enough to compensate? But I'm guessing that at best that might only be the case if it's only going very slightly over, whereas several GB over would be a problem?

No it doesn't work like that unfortunately. You run out of VRAM and performance tank. You could go from 100fps to single digits.

Hotwired said:
AMD tried that.

Fury X, a 4GB HBM card with a 4096 bit interface, at a time when alternatives ranged from 4 to 12GB.

It was "ok" but AMD decided HBM was too good for gaming peasants and we never saw it again.

Vega GPUs used HBM also.

Calin Banc · 22 Aug 2025 at 13:12

Fitting everything into vRAM also depends on how much the game WANTS to do things SMART. With SSDs/nVME drives these days, it can stream a lot more data constantly, however, the games chose not to do it and prefer to just dump everything into vRAM. Result? 8GB runs into problems, no matter how fast it is.

Disco_P · 22 Aug 2025 at 13:27

Hardware unboxed tested this, running out of vram with a PCIe 5.0 card on a system running very fast DDR5 RAM is less bad than on a PCIe 4.0 card running on a system with slower DDR4. It's still not good, you really want to avoid running out of VRAM but it's the difference between literally unplayable and just unenjoyable..

eeii · 22 Aug 2025 at 13:41

A sensible conversation about vram. Let me check I’m on the overclockers forums.

ninjag · 22 Aug 2025 at 16:14

I think it was the 3060 which prompted my interest with this and that's where I got the numbers from in my opening post.

The 3060 Ti had 8GB VRAM but it was 256-bit with 448.0 GB/s bandwidth, whereas the 3060 with 12GB of VRAM was only 192-bit and 360.0 GB/s bandwidth. But I'm sure the reviews said the 3060 Ti performed better despite the lesser VRAM, although I'm not sure if other factors were at play such as GPU clock speeds etc.

ninjag · 22 Aug 2025 at 16:14

eeii said:
A sensible conversation about vram. Let me check I’m on the overclockers forums.

oh oh, am I playing with pandora's box or something?

Quartz · 22 Aug 2025 at 16:34

ninjag said:
Was just wondering about VRAM and the focus tends to be on the amount of VRAM being priority. But can a lesser amount of VRAM be ok if the memory interface width and bandwidth is higher?

No. Assorted sites have shown the performance disparity and game spikes when you run out of VRAM. The bandwidth of the PCIe bus is vastly slower than the VRAM bandwidth of a modern GPU: 64GB/s for a v4 x16 slot vs 640 GB/s on a RX 9070 XT. That's a 10x difference; you're not going to get faster than the native VRAM.

alec · 22 Aug 2025 at 16:43

3060 Ti is still faster than 3060 across the board, except in certain games 8-12GB usage scenarios you basically have to specifically target. And even then not always slower.
Its not like either of these is max settings 4k card, but the 8GB one is still on top

Tetras · 23 Aug 2025 at 15:17

alec said:
3060 Ti is still faster than 3060 across the board, except in certain games 8-12GB usage scenarios you basically have to specifically target. And even then not always slower.
Its not like either of these is max settings 4k card, but the 8GB one is still on top

That's true, you often have to be in technically unplayable territory for the VRAM to push higher VRAM but slower cards ahead, though texture issues can rear their head earlier.

no_1_dave · 24 Aug 2025 at 11:56

ninjag said:
I think it was the 3060 which prompted my interest with this and that's where I got the numbers from in my opening post.

The 3060 Ti had 8GB VRAM but it was 256-bit with 448.0 GB/s bandwidth, whereas the 3060 with 12GB of VRAM was only 192-bit and 360.0 GB/s bandwidth. But I'm sure the reviews said the 3060 Ti performed better despite the lesser VRAM, although I'm not sure if other factors were at play such as GPU clock speeds etc.

8GB will be faster until the game needs more than 8GB of vram.

This can be mitigated in most scenarios by setting textures to a lower setting
In a lot of games, textures going from Ultra down to High is just compressing them, instead of leaving them uncompressed, and can be indistinguishable to the player during gameplay

When people take an issue with 8GB vram cards, it's not the 8GB per se they're taking issue with. It's the price point and expected performance.
A 3060TI will rarely have the power to run modern games with settings needing above 8GB anyway, but a 5060 with ray tracing etc will often need more and is priced above what you would expect to pay for 'just 8GB'.

PrincessFrosty · Friday at 20:16

ninjag said:
Was just wondering about VRAM and the focus tends to be on the amount of VRAM being priority. But can a lesser amount of VRAM be ok if the memory interface width and bandwidth is higher?

For example, 8GB VRAM with a 256-bit interface and 448.0GB/s bandwidth vs 12GB VRAM with a 192-bit interface and 360.0GB/s bandwidth. Which would perform better, does application matter etc?

It's entirely dependent on the GPU it's paired with and the applications demands of the card. Memory amount and speed is normally selected so that it does not bottleneck the GPU in most games, but is also no larger than generally useful to keep prices competitive. (More vRAM on the card, the more expensive it is to produce)

Sometime an application like some kind of video editing, CADCAM, or AI generation needs a lot more VRAM, but specialist users of these tools tend to know what they need and if necessary can buy high vRAM cards targeted at them, like RTX xx90's, or previously Titans and before that Quadros. (and all the AMD equivalents)

Keep in mind that for performance the primary unit we care about is the GPU itself, it takes GPU cycles to draw your next frame. Memory is just there for it to read data in and out of, so there's no point in supplying a slow GPU with more bandwidth than it can use, it's a waste.

The only time this choice has ever really mattered for gamers is when dual variants of the same card were available, with more or less vRAM. That typically happens when the lower value isn't quite enough in all cases, but the next biggest value is too much. The choice (and price) is passed to the consumer, this is quite rare though. The vast majority of cards have vRAM allocation that's appropriate for their GPU that covers the kind of games it will render in its useful lifetime.