NVIDIA Explains GeForce RTX 40 Series VRAM Functionality

[Nostras]

When AMD does it, it is to give your cards an edge.
When Nvidia does it, it is to skimp on VRAM/bus width.
Got it.

What? If Nvidia released a 4060Ti with 12GB of VRAM Nvidia could've used it in a way to give an "edge" competition/last-gen.
It's kind of the same thing as AMD releasing a 6700 with 6GB of VRAM and then saying "but we increased cache!!!"

But they did not.

AMD does some stupid shit but they have no blame here.
Maybe we can dogpile on them when we know more about pricing of the 7600.

 

[kawice]

The cache with good hit-rate might compensate for lower bus bandwidth, but it won't compensate for not enough of VRAM.
8 GB is barely enough for 1440p now, not to mention future or badly optimized games. And you don't buy new GPU to use if for year or two.
In the past Nvidia released cards with odd (non-mainstream) VRAM sizes like 6 GB, 11 GB or 10 GB. It's purely the design decision they chose clearly knowing what they're doing.
Is there any reason GPC cluster need to be multiply of two on any SKU? Can we have odd number of GPC cluster like 7 or 9 with 7 or 9 32-bit memory controllers?

That would allow below memory busses

160-bit bus (5 memory controllers) * 2GB VRAM = 10 GB VRAM

224-bit bus (7 memory controllers) * 2GB VRAM = 14 GB VRAM

288-bit bus (9 memory controllers) * 2GB VRAM = 18 GB VRAM


Don't tell me adding 1 GPC cluster with 1 additional memory controller and extra 2 GB of VRAM would make those cards much more expensive.
And seeing how 4080 is behind in performance to 4090 is clearly showing bad placing of all other cards. Not to mention 12 GB version of 4080 fiasco.
They should make a 40xx refresh with SUPER branding asap and fix this nonsense.

edit:
It looks like RTX 2080Ti had odd number of GPC clusters and that card was top model for 20xx series. It was basically cutdown version to RTX Titan back then.
2080Ti had 11 memory controllers with 1GB memory chips, while RTX Titan had 12 memory controllers with 2GB memory chips.

Nvidia no one will believe your lies anymore.

RTX-2080 Ti     4352     650 watt     GDDR6     352 bit     616 GB/s     1350 MHz     1545 MHz     Standard with 11 GB of Memory
Titan RTX       4608     650 watt     GDDR6     384 bit     672 GB/s     1350 MHz     1770 MHz     Standard with 24 GB of Memory

 

[Vayra86]

Well then I guess it's best to get a 7900XT/XTX since they have both large amount of VRAM and large amount of cache. Thanks Nvidia!
Props to PR team.

Duh

and CPU starts to matter as framerates increase. A Ryzen5 or i5 can handle 100fps without dying.

This. Anything over 100 FPS is in the realm of high end systems, not just GPU. But yeah that target is very sensible for an x60 now.

 

[chrcoluk]

So Nvidia have noticed the negative talk and gone PR on it (regarding VRAM).

It doesnt change anything though, either release a tech that reduces VRAM usage significantly to stop your cards been a VRAM bottleneck, or add more VRAM to your cards.

 

[fevgatos]

8GB are not enough. When developers are lazy

 

[chrcoluk]

Nvidia themselves has said the 4060ti 8GB is targeting 1080p gaming.

Perhaps they also need to add in their marketing it also targets only 1024x1024 texture resolution as well, as the two are independent of each other.

So e.g. 3080 1440p/4k combined with 1k texture resolution.
3090 4k combined with 4k texture resolution.
4070ti 1400p combined with 1k texture resolution.

This is been generous, ff7 remake, my 10 gig 3080 struggles with 512x512 textures.

Plus a disclaimer that although the cards have dedicated RT chips, the VRAM capacity may not allow you to enable RT in games. As well as textures going *poof*, blurry and so forth are a normal experience and not to be reported as a fault.

8GB are not enough. When developers are lazy

Remember dev's rule the roost, they make the software we use.

The modern way of making software for multiple platforms isnt to redesign it for each platform, but to port it over, which is why we have consistent UI methodology across devices, and typically the biggest platform wins when it comes to optimisation for the platform.

In short if I am making hardware that maybe sells 1 million units a year, and someone else is making hardware that sells 10 million a year, I better make mine like their's as the devs arent going to write their code for my hardware that sells 10% of the other platform.

I will test all cards at the same settings for the foreseeable future, which are maximum

Please test for dynamic quality drops, texture pop ins etc. as well, as VRAM issues wont necessarily slow down frame rates.

 

[Vayra86]

Perhaps they also need to add in their marketing it also targets only 1024x1024 texture resolution as well, as the two are independent of each other.

So e.g. 3080 1440p/4k combined with 1k texture resolution.
3090 4k combined with 4k texture resolution.
4070ti 1400p combined with 1k texture resolution.

This is been generous, ff7 remake, my 10 gig 3080 struggles with 512x512 textures.

Plus a disclaimer that although the cards have dedicated RT chips, the VRAM capacity may not allow you to enable RT in games. As well as textures going *poof*, blurry and so forth are a normal experience and not to be reported as a fault.


Remember dev's rule the roost, they make the software we use.

The modern way of making software for multiple platforms isnt to redesign it for each platform, but to port it over, which is why we have consistent UI methodology across devices, and typically the biggest platform wins when it comes to optimisation for the platform.

In short if I am making hardware that maybe sells 1 million units a year, and someone else is making hardware that sells 10 million a year, I better make mine like their's as the devs arent going to write their code for my hardware that sells 10% of the other platform.


Please test for dynamic quality drops, texture pop ins etc. as well, as VRAM issues wont necessarily slow down frame rates.

Can't wait to see all those Nvidia pilgrims continue their arduous journey in the future of high VRAM releases where it works exactly like you say. Already the excuses posited are getting pretty hilarious. Especially when a game post optimization still eats over 10GB. Its quite interesting if you run a higher VRAM GPU to see the real potential allocation. 11GB is commonplace, even in older titles.

I know, I'm a sadist like that, I know.

The rage vs devs is misplaced though. Rage on Nvidia, or yourself, for being an idiot buying yet another 8GB card and therefore preserving a status quo you dislike with a passion - I mean that is the actual rationale here if you get an 8GB x60(ti) in 2023. And... I know its a nasty truth, but it was also the rationale when Ampere launched. 8-10GB x70-x80 was similarly ridiculous bullshit - and it shows today, back then I was the bad guy for hammering on that fact

See the thing is, we've had >8GB for 6 years now, it should damn well be entry level at this point. Fun fact, it was Nvidia first releasing 8GB midrange in 2016, while AMD got stuck at 4! Just because Nvidia is stalling ever since, doesn't mean the world caters to that. Just like the world isn't catering to their ridiculous RT push, that they've now managed to damage by their own product stack. RT in the midrange just got pushed back for Nvidia by a full gen. Its like... help me understand the madness here!

 

[fevgatos]

Can't wait to see all those Nvidia pilgrims continue their arduous journey in the future of high VRAM releases where it works exactly like you say. Already the excuses posited are getting pretty hilarious. Especially when a game post optimization still eats over 10GB. Its quite interesting if you run a higher VRAM GPU to see the real potential allocation. 11GB is commonplace, even in older titles.

I know, I'm a sadist like that, I know.

The rage vs devs is misplaced though. Rage on Nvidia, or yourself, for being an idiot buying yet another 8GB card and therefore preserving a status quo you dislike with a passion - I mean that is the actual rationale here if you get an 8GB x60(ti) in 2023. And... I know its a nasty truth, but it was also the rationale when Ampere launched. 8-10GB x70-x80 was similarly ridiculous bullshit - and it shows today, back then I was the bad guy for hammering on that fact

See the thing is, we've had >8GB for 6 years now, it should damn well be entry level at this point. Fun fact, it was Nvidia first releasing 8GB midrange in 2016, while AMD got stuck at 4! Just because Nvidia is stalling ever since, doesn't mean the world caters to that. Just like the world isn't catering to their ridiculous RT push, that they've now managed to damage by their own product stack. RT in the midrange just got pushed back for Nvidia by a full gen. Its like... help me understand the madness here!

Ιf a patch is enough to drop vram usage by a pretty huge chunk while also making the textures look much - much - much better I feel like the vram of the card is kinda irrelevant. Current Medium textures on TLOU look like the old High while consuming less than half the amount of vram. Just stop for a second and think about it.. That's a COLOSSAL difference. Launch day it required twice the vram for the same quality!! That's insane. I hope you remember that back then I was saying Plague Tale looks much better than TLOU while using 1/3 of the vram. Well, here we are..

Im not suggestingg anyone to buy an 8gb card in 2023 for 4060 money, but damn even if the card had 16 gb it wouldn't matter seeing how devs just release games that actually hog up twice or three times the amount that is actually needed.

 

[Vayra86]

Ιf a patch is enough to drop vram usage by a pretty huge chunk while also making the textures look much - much - much better I feel like the vram of the card is kinda irrelevant. Current Medium textures on TLOU look like the old High while consuming less than half the amount of vram. Just stop for a second and think about it.. That's a COLOSSAL difference. Launch day it required twice the vram for the same quality!! That's insane. I hope you remember that back then I was saying Plague Tale looks much better than TLOU while using 1/3 of the vram. Well, here we are..

Im not suggestingg anyone to buy an 8gb card in 2023 for 4060 money, but damn even if the card had 16 gb it wouldn't matter seeing how devs just release games that actually hog up twice or three times the amount that is actually needed.

Both things are true. Yes, optimization can do a lot. Yes, games will exceed 8GB required to get the intended experience. I mean let's consider that screenshot with Medium tex. The wall is still a 720p-upscaled-looking mess. There is a lot to be gained here from higher quality texturing - and that's just textures, which is far from the only thing residing in VRAM.

But they won't hog more than 16GB. The consoles carry 16. You downplay this way too much and your feeling of VRAM kinda irrelevant is nonsense.

 

[tfdsaf]

$400 to be able to barely play games at medium settings at 1080p in 2023, what an amazing progress huh?

 

[chrcoluk]

Both things are true. Yes, optimization can do a lot. Yes, games will exceed 8GB required to get the intended experience. I mean let's consider that screenshot with Medium tex. The wall is still a 720p-upscaled-looking mess. There is a lot to be gained here from higher quality texturing - and that's just textures, which is far from the only thing residing in VRAM.

But they won't hog more than 16GB. The consoles carry 16. You downplay this way too much and your feeling of VRAM kinda irrelevant is nonsense.

It is interesting the same argument doesnt come out for games needing DLSS to be playable from people defending low VRAM, as the same applies there, devs if they wanted could make a game much less demanding, but they wont, what they can do isnt relevant, what they are doing is relevant. Been people plugging their ears for too long.

 

[Gica]

TL;DR :

NVIDIA tries to justify VRAM stagnation in their overpriced cards by using caching as a cop-out and trying to doubt the accuracy of OSD readings on VRAM usage.

Correct. Only AMD fans are allowed to challenge software measurements when they are unfavorable.
I have a 3070Ti with 8GB VRAM and I see that it performs excellently in new games as well. Far from the disaster offered by HU, as, 5 years ago, the AMD hordes were using the Assassin's series because in the other 999,999,999,999,999 games Intel was doing better.

 

[N3utro]

Well there is no other option, you either don't play the game or activate FG. I tried overclocking my 12900k to 5.6ghz all core at 1.64 volts, it was melting at 114c but hogwarts was not budging, certain areas dropped me below 60. That was on a fully tuned 12900k with manual ram. It's just one of those games...

Hogwarts runs on unreal engine 4, this engine is notorious for fps drops and stuttering. Look for a custom engine.ini file, it does wonders. Then you find yourself wondering why custom made config files created by strangers on the net end up beeing better than the ones created by the game designers

 

[gffermari]

8GB are not enough. When developers are lazy

That's a day and night difference in both how it looks and how much vram is needed now.

But still for a x60 class gpu, the segment that should play everything with tweaked settings, 8GB is not enough.
It may be for now but not for long when the consoles have way more VRAM to use.
They should go the odd way of 160bit 5x2GB GDDR6 and call it a day. No 2 versions of Tis etc. A normal 4060Ti 10GB and a cut down 4060 10GB and that's all.

The 16GB 4060Ti may be useful for some content creation software but with 128bit, I don't think it will be up to the task for gaming. Yes the capacity is there but the bandwidth is not.
(we'll see in the reviews though)

 

[yannus1]

...which, they do. NVIDIA offers wonderful, concise, well-supported features, and AMD often does not, or they are not good or popular enough to set the industry standard every time. There's no grand conspiracy here. In my opinion, Hardware Unboxed is a trustworthy source and they are generally unbiased, willing to point out strengths and weaknesses regardless of brand or product they are reviewing. Like they said on their 8 vs. 16 GB comparison video, AMD adding more VRAM to their cards isn't done out of kindness of their hearts, but because they had to pitch something to offer gamers.

It is true that their workstation segment is practically moribund (Radeon Pro is and has always been a bit of a mess, their support for most production applications is poor to non-existent especially if an app is designed with CUDA in mind - OpenCL sucks) and their high VRAM models offer 32 GB to those who need to work with extra large data sets, so giving an RX 6800/XT 16 GB isn't as big of deal to them as it is to Nvidia, who wants to ensure that their overpriced enterprise RTX A-series sell. This ensures that "hobbyist-level" creative professionals purchase at a minimum RTX 3090/3090 Ti or 4090 hardware, or supported professional models such as the RTX A4000 instead of a 3070/4070 and calling it a day.

If they were trustworthy, they wouldn't have started a censorship campaign on their forum like they just did, keeping about 20 % of the posts. Most people were complaining about the lack of VRAM and high prices and all their posts magically disappeared. Nvidia paid for RTX boxes on reviewers stages and continues paying.

 

[Chomiq]

Duh


This. Anything over 100 FPS is in the realm of high end systems, not just GPU. But yeah that target is very sensible for an x60 now.

Unless you're running Siege:

 

[Vayra86]

Unless you're running Siege:
View attachment 297036

Yeah... or Warframe, or Minesweeper, Unreal Tournament '99... etc etc etc etc etc

But you didn't buy your new GPU for that shit

 

[Dr. Dro]

If they were trustworthy, they wouldn't have started a censorship campaign on their forum like they just did, keeping about 20 % of the posts. Most people were complaining about the lack of VRAM and high prices and all their posts magically disappeared. Nvidia paid for RTX boxes on reviewers stages and continues paying.

Ich liebe kapitalismus... What do you expect? It's business and the largest the share a business holds the louder greed talks. I recall reading that Nvidia's average margins actually surpass Apple's significantly.

Every company, when in a market leader position, will use underhanded marketing tactics. See AMD when they launched Zen 3; the fabrication and widespread lies about BIOS ROM sizes/refusal to update AGESA to upsell motherboards, etc.

 

[Nhonho]

I wanted to know why only now they decided to release GPUs with larger amounts of cache.
Why didn't they do something so obvious sooner?
(I haven't read all the comments yet)

 

[Dr. Dro]

I wanted to know why only now decided to release GPUs with larger amounts of cache?
Why didn't they do something so obvious sooner?
(I haven't read all the comments yet)

Memory cache is extremely costly in regards of die area and amount of transistors required for it to function, which is why previously only data center grade processors had a large cache.

Approaching 5/4 nm class lithography, and developing advanced 3D stacking techniques, it is now feasible to spare this valuable die area without having the GPU be over 1000 mm² and costing thousands of dollars.

Expect future generations to further increase cache sizes as fabrication process node advances are achieved.

 

[londiste]

Memory cache is extremely costly in regards of die area and amount of transistors required for it to function, which is why previously only data center grade processors had a large cache.
Approaching 5/4 nm class lithography, and developing advanced 3D stacking techniques, it is now feasible to spare this valuable die area without having the GPU be over 1000 mm² and costing thousands of dollars.
Expect future generations to further increase cache sizes as fabrication process node advances are achieved.

Adding to that - memory speed used to scale up very quickly. This is no longer the case and dedicating relatively large amounts of cache to augment bandwidth is the next best thing.
Of course, when looking for optimum points both bigger manufacturers ended up cutting down the memory bus widths quite heavily. For now.

Not sure about the further increases in cache size. At least when talking about helping with the memory/bandwidth issues. It is costly as you said and looking at cache size going down from RDNA3 to RDNA2 (and sizes on Ada GPUs) it looks like there is an optimum size both AMD and Nvidia have found (and it is similar to boot).

 

[Wirko]

Adding to that - memory speed used to scale up very quickly. This is no longer the case and dedicating relatively large amounts of cache to augment bandwidth is the next best thing.
Of course, when looking for optimum points both bigger manufacturers ended up cutting down the memory bus widths quite heavily. For now.

Not sure about the further increases in cache size. At least when talking about helping with the memory/bandwidth issues. It is costly as you said and looking at cache size going down from RDNA3 to RDNA2 (and sizes on Ada GPUs) it looks like there is an optimum size both AMD and Nvidia have found (and it is similar to boot).

Adding more to that - with each new manufacturing node, I mean a full node like 7nm-5nm-3nm, the logic density increases by a factor of ~1.7x, but static RAM density only increases by ~1.2x. I don't understand why, maybe someone can enlighten me, but that's the way it is. So cache is getting relatively more expensive over time compared to logic, although it's still denser.

 

[Nhonho]

I'm not convinced yet.

I think AMD and Nvidia just recently discovered (the obvious) that more cache memory makes the GPU faster, due to the high latency of GDDR memory..

While an RTX 3090 Ti only has 6MB of L2, an RTX 4060 Ti will have 32MB!

I know that, in a new lithography, the cache memory area decreases very little. Sometimes, in a new lithograph, the cache area decrease doesn't even happen:

"TSMC's N3 features an SRAM bitcell size of 0.0199µm^², which is only ~5% smaller compared to N5's 0.021 µm^²SRAM bitcell. It gets worse with the revamped N3E as it comes with a 0.021 µm^² SRAM bitcell (which roughly translates to 31.8 Mib/mm^²), which means no scaling compared to N5 at all."

TSMC's 3nm Node: No SRAM Scaling Implies More Expensive CPUs and GPUs

Big problems from tiny memory cells.

www.tomshardware.com

 

[Pooch]

Not beyond fast, but beyond overpriced.

We increased DLSS performance yeeeeey. Which we have already made few other advertisments while we hype the dlss 3
3.0 against 2.0 and we limit dlss 3.0 so that it can be ran on 4xxx. Hype the fake resolution baby.

Im gonna quote the first guy because really he was enough for this whole thread, and add that they did skip the hogwarts, they are scared , and they want to convince you that you should still buy the 8GB model just so you can find out what you thought in the first place; you need the 16 GB model. Thankyou. Less damage control and more last ditch attempt to milk the public for a product that is handicapped.

Im gonna quote the first guy because really he was enough for this whole thread, and add that they did skip the hogwarts, they are scared , and they want to convince you that you should still buy the 8GB model just so you can find out what you thought in the first place; you need the 16 GB model. Thankyou. Less damage control and more last ditch attempt to milk the public for a product that is handicapped.

Yea that fake resolution crap is really just another crutch. So they don't have to give you the benchmarks you deserve, those screens were a joke, the first one literally says DLSS and the second not, i know it says both on the bottom, but clearly it was only on for one test so the results would be closer together to make the story look better. Like how they mentioned the performance for certain games would be almost identical, seems like the screens are a little too close nah mean?

 

[OneMoar]

I am not defending nvidia
but game devlopers need todo there part as well
a game using 10+GB of vram due to shoddy programing is unacceptable