The Announcement Nvidia Didn't Want to Promote, RTX 5070 Mobile Shows Nearly a Decade Without Increasing VRAM

[Inkz]

 

While Nvidia seems fairly confident announcing their desktop GPUs, they mostly focused only on the price for the mobile parts while not diving too much into their specs at CES. However, it has been revealed that their RTX 5070 mobile will only be equipped with 8GB of VRAM and a 128-bit bus. For reference, their 1070 mobile in 2016 had 8GB of VRAM. When asked in a press Q&A about the decision to give only 8GB on mobile Nvidia CEO Jensen Huang seemingly acknowledges they know 8GBB now isn't very much. 

 

Quote from Adam Murray of PCWorld directly to Jensen

Quote

Jensen Huang when asked about 8GB on 5070 Laptop: "We strive to reach a balance between the [whole system]. We don't always hit the right balance, but that's the goal."

 

My thoughts
 

 While VRAM has become a contentious topic, especially with Nvidia seemingly being rather stingy with it recently. As someone who owned a 1070 laptop, that 1070 was a huge increase in performance, that generation the mobile parts gave fairly similar performance to the desktop counterparts. However, I saw the news of the 5070 mobile specs and it's even more atrocious than the desktop counterparts in terms of VRAM. The 1070 mobile had 8GB of VRAM back in 2016 and here we are almost a decade later and the brand new and unreleased 5070 mobile is stuck with the exact same amount.

 Now there have been several other notable improvements since the 1000 series, but it boggles my mind that they had the capacity for this in 2016 but somehow with higher prices can't get past that. Unless they have a mid-gen refresh or a version later with more VRAM we will likely get stuck with this until the 6000 series, which likely will be at least 2 more years, going 11 years without upgrading the VRAM minimum.  And that's not even to mention the paltry 128-bit bus. The specs seem more like a 50-class card or a really low-end 60 to me. 8GB on a mid to high-end part seems unexcusable for 2025.

 

Sources

 https://www.pcworld.com/article/2572674/nvidia-rtx-5070-gaming-laptops-hide-a-terrible-secret.html

https://www.techpowerup.com/gpu-specs/geforce-gtx-1070-mobile.c2869

https://www.techpowerup.com/gpu-specs/geforce-rtx-5070-mobile.c4237
 

 

[thrasher_565]

ied imagine the gddr7 is hotter then the gddr5 as well as more cores and. dont no if the diy is more power or not but point is probably heat. and since they only make thin and light thats what you get. not saying there are faster laptops but there probably hitting a price point. but just a guess.

they also might have 4.0 ssd too witch also adds heat and well the battery it self.

Edited January 11 by thrasher_565

[starsmine]

They can refresh to 12GB 5070 mobile by year end.

Also, the amount of cares I have about laptop gaming is close to zero. 

[Inkz]

1 hour ago, thrasher_565 said:

ied imagine the gddr7 is hotter then the gddr5 as well as more cores and. dont no if the diy is more power or not but point is probably heat. and since they only make thin and light thats what you get. not saying there are faster laptops but there probably hitting a price point. but just a guess.

they also might have 4.0 ssd too witch also adds heat and well the battery it self.

Yeah although I can't imagine that over 4 generations they couldn't figure out more because of heat when the 90 and 80 class do have more vram

[Zodiark1593]

1 hour ago, thrasher_565 said:

ied imagine the gddr7 is hotter then the gddr5 as well as more cores and. dont no if the diy is more power or not but point is probably heat. and since they only make thin and light thats what you get. not saying there are faster laptops but there probably hitting a price point. but just a guess.

they also might have 4.0 ssd too witch also adds heat and well the battery it self.

Downclocking the VRAM for lower power limits (the OEM can decide the limit, so larger laptops get higher limits) should be doable. So given that laptop GPUs typically don't get upgraded, I've a difficult time calling that a good tradeoff (especially to cover use cases that rely on CUDA). Are 3GB chips not available yet?

[thrasher_565]

25 minutes ago, Zodiark1593 said:

Downclocking the VRAM for lower power limits (the OEM can decide the limit, so larger laptops get higher limits) should be doable. So given that laptop GPUs typically don't get upgraded, I've a difficult time calling that a good tradeoff (especially to cover use cases that rely on CUDA). Are 3GB chips not available yet?

i dont no im just guessing. laptops are odd balls lets thow the newest stuff in it, 1 stick or ram and have it thermal throttle...

 

but amd had fast cool lower power cpus so im guessing there going to so nice laptops from them that will give more head room for gpu cooling. just have to see i guess.

 

ya like i said there's faster laptops but cost like $6k+ xx70 is what upper mid tire so what $2000 $2500?

 

i no there were laptops that had 2 psus but i dont no if they still make em? i just hate laptops...

[Zodiark1593]

1 hour ago, thrasher_565 said:

i dont no im just guessing. laptops are odd balls lets thow the newest stuff in it, 1 stick or ram and have it thermal throttle...

 

but amd had fast cool lower power cpus so im guessing there going to so nice laptops from them that will give more head room for gpu cooling. just have to see i guess.

 

ya like i said there's faster laptops but cost like $6k+ xx70 is what upper mid tire so what $2000 $2500?

 

i no there were laptops that had 2 psus but i dont no if they still make em? i just hate laptops...

xx70 tier tends to be between $1100-$1400 for gamer oriented machines (not counting

 

Good thing I went to check prices before posting. Prices for gaming laptops jumped pretty greatly since having picked up my Legion in July. Picked mine up with Ryzen 5 7640HS and RTX 4060 for $800, while now it's difficult to find an RTX 4060 laptop for under $1000, and there's a couple RTX 4070 models under $1500. 

 

Laptops requiring 2 PSUs are pretty rare. While the average gaming laptop is pretty chonky compared to a normal laptop, the sort of machine that requires 2 PSUs easily dwarfs standard gaming laptops. I actually borrowed a machine like that for college. It was an HP ZD8000 with a Pentium 4 3.2 GHz CPU, and ATI Mobility Radeon x600. It only required one PSU, but the thing was pretty insane to carry around.

[Techstorm970]

I think there's no way in hell the RTX 5070 Mobile is faster than a desktop 4060 Ti, which famously did not benefit from having 16GB over 8GB at 1080p or 1440p.  They better upgrade the VRAM next generation, but otherwise I don't see this as a huge problem.  I say this as a laptop gamer (RTX 3070 80w).

[starsmine]

1 hour ago, Techstorm970 said:

I think there's no way in hell the RTX 5070 Mobile is faster than a desktop 4060 Ti, which famously did not benefit from having 16GB over 8GB at 1080p or 1440p.  They better upgrade the VRAM next generation, but otherwise I don't see this as a huge problem.  I say this as a laptop gamer (RTX 3070 80w).

The speed of the core and VRAM capacity are not correlated like that. 

There are literally dozens of games that do not function correctly on 8GB the work fine on 16 GB. But when you are averaging a bunch of games, this fact gets hidden. Especially when one of the ways a game does not function correctly, is not drawing the textures, so you dont see the impact in the framerate necessarily. 

The issue with 8GB of VRAM, is overflowing the capacity. 

Fun fact, if a game only fills 7GB of VRAM, if you could make a 4090 only have 7GB without effecting its memory bandwidth, it would perform the EXACT same as the 24GB 4090. So you could argue that the 4090 does not benifit from that extra 17 GB of VRAM. 

[porina]

Newer generations of ram tend towards lower voltage thus lower power. DDR3 default was 1.35v, DDR4 was 1.2v, and DDR5 is 1.1v. GDDR is related to DDR although the generational numbers don't line up. Note that is to illustrate the trend, low voltage versions as well has higher voltage overclocked enthusiast offerings are also available.

 

As for the tiresome VRAM capacity debate, remember it isn't like the GPU can't run at all on lower VRAM. The user just has to ensure settings are appropriate for what's available. The apparent scaling with GPU perf is indirect, that a higher perf GPU is more likely to used with settings that use more VRAM.

[starsmine]

1 hour ago, porina said:

Newer generations of ram tend towards lower voltage thus lower power. DDR3 default was 1.35v, DDR4 was 1.2v, and DDR5 is 1.1v. GDDR is related to DDR although the generational numbers don't line up. Note that is to illustrate the trend, low voltage versions as well has higher voltage overclocked enthusiast offerings are also available.

power is Cl*f*v^2, so yes each bit is lower powered, and its less Jules per clock, but they are clocked higher to get the higher bandwidth. I do not know enough about how they are making this dram to speak on if Capacitance load increased, that is a ratio of transistors and node features. I can ask micron when I tour their fabs in late Feb, but who knows what answers they will actually be able to give me, or if I will have enough context to figure out how much the difference is. 

The main point is that lowering the voltage doesnt mean lower power when you also increase performance at the same time. 

1 hour ago, porina said:

As for the tiresome VRAM capacity debate, remember it isn't like the GPU can't run at all on lower VRAM. The user just has to ensure settings are appropriate for what's available. The apparent scaling with GPU perf is indirect, that a higher perf GPU is more likely to used with settings that use more VRAM.

 

[Winterlight]

8Gb VRAM shouldn't exist at all anymore. Maybe only for like xx50 series it would be acceptable but xx60 should get 12 VRAM, xx70 - 16Gb VRAM and xx80 - 20Gb VRAM at least. It good example 4070 Ti (not super) than GPU have power but it massive bottleneck by its VRAM. Or if you look in AMD side their GPU usually have more than enough VRAM but bottlenecked by GPU core.

[LAwLz]

9 hours ago, starsmine said:

There are literally dozens of games that do not function correctly on 8GB the work fine on 16 GB.

Got any examples?

 

1 hour ago, Winterlight said:

It good example 4070 Ti (not super) than GPU have power but it massive bottleneck by its VRAM.

Source?

 

 

As someone who happily uses a graphics card with 8GB of VRAM I am very surprised by all these outcries about how 8GB shouldn't exist, how it's a massive bottleneck and so on.

[Zodiark1593]

1 hour ago, LAwLz said:

Got any examples?

 

Source?

 

 

As someone who happily uses a graphics card with 8GB of VRAM I am very surprised by all these outcries about how 8GB shouldn't exist, how it's a massive bottleneck and so on.

I’m kind of in the middle. 8 GB hasn’t been a meaningful obstacle that can’t be addressed by adjusting settings. Though on the other hand, there are games where VRAM is a limiting factor with settings where the GPU would otherwise have the performance for. 
 

The 128-bit bus used on these GPUs is particularly thorny. This certainly reduces the costs for implementing the dGPU (fewer traces, fewer parts, etc, probably why so many laptops use Nvidia), and the cards themselves don’t appear starved for bandwidth; though if chips are still limited to 2 GB/module, it’s understandable (albeit I still don’t like it) that 8 GB would be adhered to. 
 

My bet is that once 3 GB GDDR7 modules are available, we’ll probably leap past 8 GB pretty quick. 

[porina]

6 hours ago, starsmine said:

The main point is that lowering the voltage doesnt mean lower power when you also increase performance at the same time. 

I could have worded that better. There is generally generational power efficiency improvement, especially comparing equivalent speeds. I agree that depending on exactly where the operating point is, there may or may not be a power reduction going to a newer gen but faster implementation.

 

56 minutes ago, Zodiark1593 said:

The 128-bit bus used on these GPUs is particularly thorny. This certainly reduces the costs for implementing the dGPU (fewer traces, fewer parts, etc, probably why so many laptops use Nvidia), and the cards themselves don’t appear starved for bandwidth

There is a trend in recent gens to increase the cache sizes within the GPU helping to improve the effective bandwidth available. AMD documented it on RDNA2 and I assume similar principles apply to their newer GPUs as well as Nvidia. Just we don't have the numbers for those. AMD's calculation was:

Effective bandwidth = [cache hit rate%] * [cache bandwidth] + [VRAM bandwidth]

Hit rate did vary with the cache size and the resolution you game at, since higher res means more data in flow and lower hit rate.

 

If I can find the slide again I'll edit something in later. Found it! Main take away is that effective bandwidth can be increasing even if external bandwidth alone doesn't look it.

Spoiler

 

 

 

 

[Inkz]

2 hours ago, LAwLz said:

Got any examples?

 

Source?

 

 

As someone who happily uses a graphics card with 8GB of VRAM I am very surprised by all these outcries about how 8GB shouldn't exist, how it's a massive bottleneck and so on.

It is though. As someone who also uses an 8GB card if you run out you can get severe stuttering, missing or glitching textures, or not have textures loading at all. If you don't think it matters check Hardware Unboxed's video directly comparing the 4060 ti 8GB vs 16 GB and even on a mid range part the difference is night and day. While if you don't exceed the 8GB it makes basically no difference. And while the averages might be similar you can get a lot more stuttering and be a much worse experience even if on a graph of averages doesn't show it.  As someone with a 1070 mobile on my old laptop it's shocking that they refuse to give more capacity even as the games have started using so much more and my 1070 is from 2016.

[starsmine]

7 hours ago, porina said:

There is a trend in recent gens to increase the cache sizes within the GPU helping to improve the effective bandwidth available. AMD documented it on RDNA2 and I assume similar principles apply to their newer GPUs as well as Nvidia. Just we don't have the numbers for those. AMD's calculation was:

Effective bandwidth = [cache hit rate%] * [cache bandwidth] + [VRAM bandwidth]

Hit rate did vary with the cache size and the resolution you game at, since higher res means more data in flow and lower hit rate.

 

If I can find the slide again I'll edit something in later. Found it! Main take away is that effective bandwidth can be increasing even if external bandwidth alone doesn't look it.

  Hide contents

 

 

 

 

 

 

I want to build off of this, I found it kinda crazy that both NVIDA and AMD went down this approach after three nodes of near zero SRAM scaling, N3 has pretty much zero (which Is why I think nvidia did not even bother). This is WHY amd did chiplets for RDNA 3, if the size was going to be the same, why do it on the expensive node? 

But forward looking, GAAFET will be the first node in a long time to have SRAM scaling, and a lot of it, because essentially you can fold the structure in half like a sandwhich, so it you will get like the same amount of SRAM in 2/3rds the die space and the latency of SRAM is also improved (I wanted to a find a source for my Gaafet sram scaling claim, but all I could find was vertical Gaafet in my 15 min of googling, if I find the source for this ill edit it in)

https://ieeexplore.ieee.org/document/10161740

Quote

At the highest associativity of 8 with the CuBKM BEOL option, the 3 nm GAAFET presents a 26.46% decrease in read latency and a 23.93% decrease in write latency over the 7 nm FinFET and a 3.04% decrease in read latency and a 5.69% decrease in write latency over the 3 nm FinFET

Vertical Gaafet in another 10 years looks to do the whole SRAM scaling thing again. by taking that sandwich and putting it on its side. 

Yes that means moors law is 2x every 10 years 

What I do mean is that Nvidia will be able have a lot more die space to play around with and use how they want with the next architecture if they are skipping N3 and going straight to N2 or 18A or whatever samsung is on, allowing them to go kinda ham on it. The choice to pretty much 10x the amount of SRAM from ampere to ada (and therefore blackwell cant have less) when there was something like a 10% SRAM scaling from Samsung's 8nm to TSMC N5, really locked up ALOT of the die space from what could have been used for more cuda cores, a smaller cheaper die, or more memory buses. AMD made the same choice as Nvidia, but mitigated the cost by just moving all that cache off the expensive die. 

here is a picture of a 6T sram diagram and its planer/finfet structure 

https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7890390

best I got for you is
 

https://www.tomshardware.com/tech-industry/sram-scaling-isnt-dead-after-all-tsmcs-2nm-process-tech-claims-major-improvements

Anyone know what happened to wikichip? Usually some of this information is there. 

Samsung 8nm looks to be 0.0262um, that jump to TSMC 5nm was bigger then I remembered. 

6 hours ago, Inkz said:

As someone with a 1070 mobile on my old laptop it's shocking that they refuse to give more capacity even as the games have started using so much more and my 1070 is from 2016.

The architecture of this chip was solidified at least 3 years ago, 3GB GDDR7 was anticipated to be out by now. They are not refusing to give more capacity, they just are not going to wait half a year for it when that's the easiest refresh opportunity of their life to get people to buy this generation twice. 

Edit: seems I was mistaken about Normal Gaafet SRAM scaling, I was remembering the vertical Gaafet SRAM scaling, which is why all the sources I could find were about Vertical Gaafet. 
Intel 18A SRAM is also .021, so same as TSMC's 5nm and 3nm. Boys, we are cooked. 
 

Quote

We find comparable PFFET and GAAFET cell areas for all designs. In fact, with the same hSi for the PFFETs and the GAAFETs in the HD SRAM cell, we find exactly equal cell areas (0.008µm2).

 

[Inkz]

4 hours ago, starsmine said:

 

The architecture of this chip was solidified at least 3 years ago, 3GB GDDR7 was anticipated to be out by now. They are not refusing to give more capacity, they just are not going to wait half a year for it when that's the easiest refresh opportunity of their life to get people to buy this generation twice. 

 

I mean I do get that and I have heard this before. But both Intel and AMD seemed like they were able to design around it on their mid-range cards going to 16 so it seems surprising Nvidia as sometimes the biggest company in the world can't figure that out. I will admit I can't speak into the super deep engineering side but I hope they do a sooner than later mid-gen refresh with more vram.

[starsmine]

23 minutes ago, Inkz said:

I mean I do get that and I have heard this before. But both Intel and AMD seemed like they were able to design around it on their mid-range cards going to 16 so it seems surprising Nvidia as sometimes the biggest company in the world can't figure that out. I will admit I can't speak into the super deep engineering side but I hope they do a sooner than later mid-gen refresh with more vram.

Intel did not. Intel is not making money on battlemage. intel has to be "charitable" and lose money to gain market share at this stage. No one else is willing to straight up lose money to do something similar. Nvidia burned a lot of bridges with the RTX 3060 where they forced AIBs to lose money. (who could have foreseen GDDR6 prices tripling months after the release of that card... I actually woulda made the same bet nvidia did though)

AMD did chiplets which has consequences for the 256bit 16GB cards, but did any of those even make it to mobile? I dont think they did. The 7600xt was a 128bit gpu sure, the 16GB varient was clamshelled, and I also dont think was ever put in mobile. 

yes the 5070 mobile could clamshell as well, but that's detrimental to battery life, and if its going to be refreshed anyways to 12GB by year end, why even bother with it.  DRAM isnt SRAM, you cant turn it off. The reason you dont see mobile AMD GPUs is because of battery life, laptop makers get the blame if battery life is bad, not the Nvidia or AMD. 

But also this is a mobile chip, why are you trying to play new games anyways and hoping it performs like a desktop or even a console. 

I guess AMD did launch two mobile GPU with chiplets.

Probably sold a whole dozen of them. 

[LAwLz]

21 hours ago, Inkz said:

It is though. As someone who also uses an 8GB card if you run out you can get severe stuttering, missing or glitching textures, or not have textures loading at all. If you don't think it matters check Hardware Unboxed's video directly comparing the 4060 ti 8GB vs 16 GB and even on a mid range part the difference is night and day. While if you don't exceed the 8GB it makes basically no difference. And while the averages might be similar you can get a lot more stuttering and be a much worse experience even if on a graph of averages doesn't show it.  As someone with a 1070 mobile on my old laptop it's shocking that they refuse to give more capacity even as the games have started using so much more and my 1070 is from 2016.

Quote from the video:

Quote

Here's a quick look at the 15-game average data. The 16GB model is just 4% faster on average which makes sense as most of the games saw no performance change. We're also looking at just a 3% improvement at 1440p because, again, most of the games tested saw no real performance change and a lot of the more demanding games were tested using dialed down quality settings which is done to allow for more payable frame rateson the more mainstream models.

 

Then, oddly enough, he concludes with:

Quote

So there you have it. The 16GB VRAM buffer certainly works and it does make the RTX 4060 Ti a much better product in my view. And I am confident that in years to come that this version will hold up reasonable well.

 

 

I'm not sure why he considers the difference "much better". The small performance gap, even when including 1% lows, doesn't align with that claim. I suspect it's because of his focus on "future-proofing", which he has emphasized in other videos, hence the comment about "years to come".

 

 

 

There are a handful of modern games that push VRAM limits, but HUB seems to prioritize VRAM in their videos/articles more for its potential future importance than its relevance today.

 

 

Personally, I find the "future-proofing" argument flawed. By the time higher resource demands are standard, it's usually better to either:

1. Slightly lower game settings (seriously, "Ultra" settings are usually awful in terms of optimization. Lowering the settings to "high" offers almost the same quality but at much higher performance).
2. Upgrade the GPU to a newer one that offers far better performance in all aspects compared to today's "future-proofed" options.

 

Since you brought up "VRAM usage" and it is relevant to the conversation, I want to address something.

When monitoring memory usage (RAM or VRAM), the number reported (for example ~10GB used) doesn't necessarily mean the game or program needs that much.

 

Just like system RAM, VRAM is opportunistic. If there's free space, it will hold onto resources it no longer needs to save time in case it needs them later. This doesn't mean the game would crash or perform worse with less memory. It simply means the system is optimizing its performance. As VRAM fills, it starts clearing out unused assets more aggressively, which might not affect performance at all.

 

For example, my Windows PC currently shows ~12GB of system RAM in use, even though I only have two browser tabs open. The "extra" RAM usage isn't critical, it's simply the OS keeping old resources cached because there's no risk of running out and it MIGHT need it later. VRAM works the same way.

This is why VRAM "usage" graphs should be interpreted carefully. At most they can indicate the maximum memory needed, but not the minimum necessary to run smoothly.

[Inkz]

2 hours ago, LAwLz said:

Quote from the video:

 

Then, oddly enough, he concludes with:

 

 

I'm not sure why he considers the difference "much better". The small performance gap, even when including 1% lows, doesn't align with that claim. I suspect it's because of his focus on "future-proofing", which he has emphasized in other videos, hence the comment about "years to come".

 

 

 

There are a handful of modern games that push VRAM limits, but HUB seems to prioritize VRAM in their videos/articles more for its potential future importance than its relevance today.

 

 

 

Yeah like I said in many games if you don't exceed that buffer or are close it doesn't matter that much. But some things can't be shown on averages, such as looking at that example for extremely small spikes, which don't really affect averages but can be felt. Certain games don't let you select the higher-quality options, textures won't load, or glitch out. You can't just always plot this stuff on a graph but it changes the feeling of the game.  There is a reason why higher-end cards still come with more VRAM than the lower ones besides productivity.  I could find other sources besides HUB, I just like their channel a lot so it was the first one I went too.

 

3 hours ago, LAwLz said:

 

 

Since you brought up "VRAM usage" and it is relevant to the conversation, I want to address something.

When monitoring memory usage (RAM or VRAM), the number reported (for example ~10GB used) doesn't necessarily mean the game or program needs that much.

 

Just like system RAM, VRAM is opportunistic. If there's free space, it will hold onto resources it no longer needs to save time in case it needs them later. This doesn't mean the game would crash or perform worse with less memory. It simply means the system is optimizing its performance. As VRAM fills, it starts clearing out unused assets more aggressively, which might not affect performance at all.

 

For example, my Windows PC currently shows ~12GB of system RAM in use, even though I only have two browser tabs open. The "extra" RAM usage isn't critical, it's simply the OS keeping old resources cached because there's no risk of running out and it MIGHT need it later. VRAM works the same way.

This is why VRAM "usage" graphs should be interpreted carefully. At most they can indicate the maximum memory needed, but not the minimum necessary to run smoothly.

It's true if you have a ton spare you system can use more than you actually need, but some people's needs can be higher than others. I have 32GB but I still ran out playing a very demanding game. And similar to VRAM when I exceeded it my game kept freezing over and over until I closed basically everything except that program. Same thing with VRAM, the capacity doesn't matter that much if you don't use all of it but things really suck if you go over that limit.  Would you argue that my program freezing and being unusable is not critical? 

[LAwLz]

12 hours ago, Inkz said:

But some things can't be shown on averages, such as looking at that example for extremely small spikes, which don't really affect averages but can be felt.

That should show up as the "low 1%". That's what that graph should be measuring and showing (in theory), and even that shows very little difference between the cards.

 

 

12 hours ago, Inkz said:

Certain games don't let you select the higher-quality options, textures won't load, or glitch out.

Do you have any examples?

This conversation started because someone else said that, I asked for some examples and I have still not been shown any. I don't doubt that they exist, but the lack of examples does make me skeptical of the claim that it is widespread.

I also, as I said earlier, strongly feel like this could be a case of where people just select the highest settings without thinking and then get mad when they can't select it, without reflecting over if it is even necessary. How much detail do you really lose by not being able to run the game at the highest settings? Is it a big deal even if we find these examples?

 

Of course, more would be better (with the exception of higher power consumption at a higher cost) but what I am trying to figure out is how much better would it actually be.

 

 

 

 

13 hours ago, Inkz said:

There is a reason why higher-end cards still come with more VRAM than the lower ones besides productivity.

Yes, but more VRAM doesn't automatically makes a card better, because at some point you run into other bottlenecks as well.

The card needs to also have the necessary compute resources to actually be able to handle the additional data.

If we go back to CPUs to illustrate the point a bit easier, a single-core CPU at 1Ghz probably won't benefit from having 128GB of RAM vs 4GB of RAM, because the CPU isn't fast enough to process all that data anyway.

The same applies to graphics cards as well. More VRAM is only a benefit if the card is able to handle the additional data in other areas too.

 

 

13 hours ago, Inkz said:

I have 32GB but I still ran out playing a very demanding game.

32GB of RAM isn't enough for some of the games you play?

I find that very hard to believe. Are you sure you're not just assuming that it is a RAM issue and there is a possibility that there is something else that causes the hiccups? CPU related things, GPU related things or just an issue with the game itself? All of those are things that could show up as small hiccups in the game.

The reason why I ask is because I have noticed a tendency on this forum (and other forums) where people experience an issue and then pin it on a certain thing without actually ruling out other possibilities. Pinning an issue on a specific component like that requires quite a bit of work that people generally don't do. That a game (or maybe a program?) needs more than 32GB of RAM seems bonkers to me, and I think it is a far more likely explanation that it is related to something else. Are you sure it's not a memory leak inside the program? If that's the case then it doesn't matter how much RAM you have, you will still run into issues. A hardware upgrade could delay the issue slightly, but it wouldn't actually solve it.

[Inkz]

10 hours ago, LAwLz said:

That should show up as the "low 1%". That's what that graph should be measuring and showing (in theory), and even that shows very little difference between the cards.

 

That's what the spikes are as shown in the screenshot.

 

 

11 hours ago, LAwLz said:

 

Do you have any examples?

This conversation started because someone else said that, I asked for some examples and I have still not been shown any. I don't doubt that they exist, but the lack of examples does make me skeptical of the claim that it is widespread.

I also, as I said earlier, strongly feel like this could be a case of where people just select the highest settings without thinking and then get mad when they can't select it, without reflecting over if it is even necessary. How much detail do you really lose by not being able to run the game at the highest settings? Is it a big deal even if we find these examples?

 

Sure, do. There are plenty of examples online, not very hard to find. This is an example from Hogwarts Legacy using the same settings.  And also here is the whole
video if you wanted to inspect the whole thing. 

 
Yes, I mean if you think using 99-100% of RAM couldn't be RAM related I am not sure what to say. I did state to be fair that they were more demanding games so I was not making the argument that are gonna play something basic like CS and use that much RAM. Some games simply use that much. Flight Sim 2024 recommends 64GB now even.  I played a larger VRChat lobby which models use a ton of RAM in and was using like 22 GB of RAM just for that application alone.  If you want I can personally walk you through a demonstration if you truly find it so mind-boggling to use a lot of RAM in demanding games. And this isn't a new phenomenon, even Linus has talked about it back like almost 10 years ago in the NCIX days.  If you want I can keep looking for more examples of textures issues, it isn't that hard to find.

11 hours ago, LAwLz said:

 

 

 

32GB of RAM isn't enough for some of the games you play?

I find that very hard to believe. Are you sure you're not just assuming that it is a RAM issue and there is a possibility that there is something else that causes the hiccups? CPU related things, GPU related things or just an issue with the game itself? All of those are things that could show up as small hiccups in the game.

The reason why I ask is because I have noticed a tendency on this forum (and other forums) where people experience an issue and then pin it on a certain thing without actually ruling out other possibilities. Pinning an issue on a specific component like that requires quite a bit of work that people generally don't do. That a game (or maybe a program?) needs more than 32GB of RAM seems bonkers to me, and I think it is a far more likely explanation that it is related to something else. Are you sure it's not a memory leak inside the program? If that's the case then it doesn't matter how much RAM you have, you will still run into issues. A hardware upgrade could delay the issue slightly, but it wouldn't actually solve it.

This is a Strawman fallacy. People weren't saying something unreasonable. We are talking about a 70 class new GPU that is stagnating in capacity despite the requirements of programs requiring more, not an old one that it wouldn't matter that much with.  That's like someone making a new system with a 9800 X3D but only giving it 8GB of RAM, it just doesn't make sense. A 5070 should not be a weak GPU that can't hope it's VRAM when some games have been pushing that infamous 8GB frame buffer from 2 generations ago. 

11 hours ago, LAwLz said:

 

 

Yes, but more VRAM doesn't automatically makes a card better, because at some point you run into other bottlenecks as well.

The card needs to also have the necessary compute resources to actually be able to handle the additional data.

If we go back to CPUs to illustrate the point a bit easier, a single-core CPU at 1Ghz probably won't benefit from having 128GB of RAM vs 4GB of RAM, because the CPU isn't fast enough to process all that data anyway.

The same applies to graphics cards as well. More VRAM is only a benefit if the card is able to handle the additional data in other areas too.

 

 

 

[starsmine]

27 minutes ago, Inkz said:

Sure, do. There are plenty of examples online, not very hard to find. This is an example from Hogwarts Legacy using the same settings.  And also here is the whole
video if you wanted to inspect the whole thing. 

 

Signal boost that those are the same settings, which means the FPS comparison between them is a lie, because in practice, they are not. A benchmark chart will literally say they are all running at 50FPS and not tell you WHY that is happening. The textures failing to load is a problem even with 8GB cards, this has NOTHING to do with the compute capabilities of the GPU core. 

 

11 hours ago, LAwLz said:

If we go back to CPUs to illustrate the point a bit easier, a single-core CPU at 1Ghz probably won't benefit from having 128GB of RAM vs 4GB of RAM, because the CPU isn't fast enough to process all that data anyway.

The same applies to graphics cards as well. More VRAM is only a benefit if the card is able to handle the additional data in other areas too.

That isnt how ram works, Thats like saying a computer does not benifite from a large drive because it cant process that data. 
A single core absolutly would benifite from 128GB of RAM if the work load task can use more then 4GB. It has NOTHING to do with how fast the CPU is outside of extremes (your example, is not extreme enough)

Every time the core goes to cache beyond L1, it STALLS. L2 stall is short, L3 stall is longer, Ram stall is longer still, and if you have to go beyond that because you overflowed ram in your workload, that stall for ALL real time applications is noticeable to people. It has nothing to do with how fast the compute core is. You can mask this, like when you context shift on PC it can swap what was in the page file over to ram in milliseconds, and a person is mentally trained to accept the 100s of milliseconds it takes for it to load back up, but if those stalls are constantly happening as you are actively using the application you notice. 

As soon as a workload overflows ram, regardless of computer power, you have stalls. 

bandwidth comes in when you are trying to fill the L2 cache, which is why bandwidth is needed for higher compute cards. Not the VRAM ammount.
 

[LAwLz]

12 hours ago, Inkz said:

That's what the spikes are as shown in the screenshot.

What is your point?

As I said, even the low 1% shows very little difference in the chart. You picking one example where it might be a slightly bigger difference and going "look, there is a difference!" is not really a response to what I said. 

 

 

12 hours ago, Inkz said:

Sure, do. There are plenty of examples online, not very hard to find. This is an example from Hogwarts Legacy using the same settings.  And also here is the whole
video if you wanted to inspect the whole thing. 

Again, what is your point? This is your response to me saying games that require more than 8GB might exist, but I would like to see some examples.

It's a comparison between an 8GB and a 4GB card. That is not the kind of example I was asking for. I also said that in those cases where 8GB of VRAM might not be enough when playing on the highest settings, simply lowering the settings a bit could probably help a lot. The video you linked shows the same settings, so it is not really a response to that either.

 

I genuinely do not understand why you posted this video and what you think it proves in this context.

 

 

 

12 hours ago, Inkz said:

Some games simply use that much. Flight Sim 2024 recommends 64GB now even.

Well, if we are being pedantic they recommend 32GB, not 64GB.

They claim the ideal spec is 64GB. I quickly added together all the "ideal specs" for Flight Simulator 2024 and the PC cost about 2700 dollars. So maybe that is not the ideal game to use as a baseline for what PC equipment people who want to game should buy. It seems more like a very big outliar, not the norm. 

 

 

 

  

12 hours ago, Inkz said:

Yes, I mean if you think using 99-100% of RAM couldn't be RAM related I am not sure what to say.

Are you talking about the texture not loading in the picture with the 3GB card? Okay... What is your point? We are talking about 8GB cards, not 3GB cards.

That video also shows GPU usage being at 100%, so how do you know it's not a compute issue or something else? A bug in the software perhaps?

Again, I think you are jumping to conclusions and are also talking about quite different situations from what I was talking about.

 

 

 

12 hours ago, Inkz said:

This is a Strawman fallacy.

In what way was it a strawman argument?

 

 

12 hours ago, Inkz said:

A 5070 should not be a weak GPU that can't hope it's VRAM when some games have been pushing that infamous 8GB frame buffer from 2 generations ago. 

Again, do you have any examples of where this might be an issue?

And no, Hogwards Legacy not loading some textures on a graphics card with 3GB of VRAM is not evidence of a 5070-mobile class GPU with 8GB of VRAM being limited by the VRAM.

 

 

 

12 hours ago, starsmine said:

That isnt how ram works, Thats like saying a computer does not benifite from a large drive because it cant process that data. 

You know what I mean...

The point is that the bottleneck means adding resources in other places doesn't make sense, because it will not give a meaningful speed improvement. If you're trying to encode a video on a single-core CPU then adding more RAM won't make things much faster, because the issue is the CPU itself.

 

 

 

 

12 hours ago, starsmine said:

As soon as a workload overflows ram, regardless of computer power, you have stalls. 

I know.

That is why I am asking for examples of when this might happen on a graphics card with 8GB of VRAM. I have still not gotten any examples of it. I think it is crazy that I keep asking for something that so many people say is widespread, and yet nobody is posting any examples.

 

 

This is what I am asking for:

1) Mainstream cases where a card with 8GB of VRAM runs out of memory.

2) A subgroup of those cases where there is also evidence that lowering the graphics fidelity slightly doesn't alleviate the issue, or has a major negative impact on the visual quality.