The 3080 Ti probably won't even be that price this time, it sold poorly in everywhere except Germany & places using it for Ai, unlike the other Ti's that sold like hotcakes.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
Rumor: NVIDIA GeForce RTX 3080, RTX 3070 leaked specs: up to 20GB GDDR6 RAM
- Thread starter Earvin Infinity
- Start date
It will be '$699' with a real world +15-25% delta over a 2080Ti. This is a node shrink generation with what are likely to be legitimately competitive AMD cards. They can't get away with gouging or renaming/reshuffling the performance stack again. 2080Ti will be a 2 year old card at time of release, a super equilatent for $1000+ is a ridiculous musing.
If the 3080ti can 5k/60FPS FFXV I'll be happy.
Have to upgrade regardless, 1080ti is getting long in the tooth. :(
Have to upgrade regardless, 1080ti is getting long in the tooth. :(
Next gen is arriving and devs will have access to 16GB of unified ram to play with. Expect min and max reqs to go up.
Yeah but that's like having 11-12GB VRAM now. It's not a realistic requirement for games, and 20GB would be overkill for outperforming consoles that have 16GB shared with the OS.
If there's any truth to this rumour at all (which I'm not convinced there is), I would have to assume the more conservative values of 10GB for the 3080 and 8GB for the 3070. If it turns out that the consoles are higher than this figure, then these values would seem more reasonable.
I'm thinking I'm skipping out on Ampere and waiting on Hopper so I can just do a complete system rebuild utilizing DDR5 which should be consumer ready by the time NVDA's MCM GPU is released. I have a feeling we'll see 8800GTX levels of performance increase. I've never bought new GPUs right before a console release. I've always gotten the first new architecture after a new console release. I've don't that for the last 2 console generations. I'm gonna need to talk myself down when the 3000 series is released. It's very tempting.
With MCM, I also think it will be the nail in the coffin for consumer mGPU.
Speaking of older GPUs...I was looking at MSRP for the 8800GTX and 8800GTX Ultra and in today's money they would be $750 and $1030 when taking inflation into account. Flagship cards have always been destroying bank accounts. I couldn't afford those cards back in the day since I was in my early 20s and broke.
With MCM, I also think it will be the nail in the coffin for consumer mGPU.
Speaking of older GPUs...I was looking at MSRP for the 8800GTX and 8800GTX Ultra and in today's money they would be $750 and $1030 when taking inflation into account. Flagship cards have always been destroying bank accounts. I couldn't afford those cards back in the day since I was in my early 20s and broke.
Of course, as with any generation. 20GB is enormous though. Wouldnt complain if they do end up using this ballpark.
Modern cards have enough VRAM to render 4K. 20GB even with a new generation of games should be more than sufficient.
I think it's impossible to estimate how much vram we'll need in the future without knowing how much shared memory the new consoles are going to have.
A 24gb PS5 would probably mean ~22gb(?) shared between the CPU and GPU. Which could become a problem for 8 to theoretical 10gb cards. Especially if you want to run even higher Shadow resolutions, details or above console texture settings.
That said, I really don't think we'll see 20gb on gaming cards anytime soon. Maybe a 3080ti with 16gb? But I don't expect that card to be below 1400€.
A 24gb PS5 would probably mean ~22gb(?) shared between the CPU and GPU. Which could become a problem for 8 to theoretical 10gb cards. Especially if you want to run even higher Shadow resolutions, details or above console texture settings.
That said, I really don't think we'll see 20gb on gaming cards anytime soon. Maybe a 3080ti with 16gb? But I don't expect that card to be below 1400€.
Wasn't 16 GB and 13 GB usable RAM the latest and most reliable information? That would mean roughly 5 GB VRAM if the consoles use 8 GB as RAM
The 20xx series are huge chips, which makes them expensive and also means that they have limits to what they can cram onto them. It's not an accident that the Titan RTX is not significantly faster for gaming than the 2080 Ti. For a first gen raytracing product Nvidia also had no reason to go all in on RT cores since even now only a handful of games offer support. When the majority of games shift to use raytracing more then we will start seeing GPUs that will increasingly have more hardware towards accelerating those processes.
Just throwing in more RT cores isn't necessarily going to give us raytracing with a low performance hit. As I understand it, the problem isn't that the BVH tree traversal acceleration (which the RT cores do) is too slow but the shading that follows. So having more efficient CUDA cores with some extra functionality to handle raytraced data faster might be the solution rather than just doubling the amount of RT cores.
At the same time developers and API vendors are working on techniques to reduce the rendering load and do raytracing more efficiently.
1. Raytracing will have a large performance hit so I expect that we won't see it in the same capacity on consoles as we do on PC. I expect that it will be instead used for more controlled situations like realtime cutscenes, interior sections etc. Wherever it will make a significant impact visually.
2. Depends largely on what GDDR6 or whatever costs on the market. 20 GB seems excessive and we are unlikely to see that on anything but the Titan parts. You have to remember that consoles require to cram in the same memory pool everything that is handled by system RAM on computers which means less left for games to utilize.
3. Probably not. More than last gen probably especially if they have several tiers of device.
Pre-DLSS, FFXV with all of the settings turned on/max chewed through 8GB cards and made 11GB cards sweat. Now imagine the whole game updated to support RT/PT like the Back Stage demo. 10GB would probably be a minimum recommended spec, with 16GB+ being the ideal.
Its funny because you could choose how much total memory you wanted to use for the game before "flushing", tweaking the TRAM setting in the .ini file for the VRAM and using the Special K mod for system ram.
Its funny because you could choose how much total memory you wanted to use for the game before "flushing", tweaking the TRAM setting in the .ini file for the VRAM and using the Special K mod for system ram.
Nowhere did I say that such a card wouldn't sell. You're not reading what I write. It doesn't matter whether there is a market, it would be a big mistake for Nvidia to do it, so it wont happen.
If this were 2017 and we were speculating about Turing release, you'd have a great point. But Nvidia has already hitched their wagon to RTX and they cannot about face. Ampere or whatever comes next will be plenty fast at rasterisation and you seem to be under the impression that ray tracing hardware support is the sole reason why prices are where they are. This isn't the case.
Fair to assume that BIG Navi (N23) will sit in between a 3070 and 3080, and the full/uncut GA102 chip will be 20-25% faster?
20GB figure is there simply because that's what you may get on a 320 bit bus. Same goes for the 16GB one.
I think it's possible that this time around NV will allow the AIBs to produce custom models with twice the VRAM so a 3070 with 16 and a 3080 with 20 will be available - but the prices will likely be considerably higher than those of reference 8/10GBs models.
We can't assume anything on their comparative performance just yet.
16 GB of unified memory isn't that much when OS take 2-3 GB just for itself and then you divide the rest as general memory and as VRAM, start to look like ~ 6 / 6 split all of sudden.
the RTX 2080 had fewer than the 1080 ti as did the 1080 from the 980 ti. It's a pretty typical trend tbh.
Even with the news of all this, I'm still planning getting a 2060 or 5600 XT soon(maybe I'll push for a 5700, maybe not). It's likely the highest end cards will be introduced first and with the 3070 being the cheapest but still costing at least $500. Secondly I was cheap and bought a 1050 TI back in late 2016 and I can't stand how weak it is anymore, I'm ready to upgrade now. Even if I only stick with it for a year lol.
If I could go back in time I'd tell myself to at least put $40 more on the RX 470, then I think I could have held out until the 3060 or whatever the 1660 TI successor is called.
If I could go back in time I'd tell myself to at least put $40 more on the RX 470, then I think I could have held out until the 3060 or whatever the 1660 TI successor is called.
Honestly, if Big Navi isn't at least on par with the RTX 3080 while being $50 cheaper, all those inner nicknames of "Nvidia Killer" would look really dumb in retrospect
16gb sounds realistic currently imo. Maybe they'll add slower ddr3/4 to just hold the OS, just like they did with ps4pro. 8-10gb for video memory could happen that way
Maybe they'll add a bit more shared ram "last second" like they did with ps4... We'll see.
2060 for 300 is the best bet for a "budget" card, VRS will be super important next gen & the AMD GPU's are going to age poorly fast because of it.
I get it, but they're absolutely going to show their Ampere Quadro lineup at GTC 2020, even if this gets postponed to Siggraph or Gamescom (like the 20** RTX SKU reveal).
I was just basing the numbers on those "Big Navi 17% faster than 2080Ti rumors" earlier, and just assumed the full/uncut GA102 core will be 40% faster than the TU102, which will power their high end Quadro lineup, and the $2500 (or probably even more this time) consumer variant with Deep Learning performance cut down, Titan RTX Ampere.
I would be very (pleasantly) surprised if the fastest Radeon consumer GPU, that's going to release in 2020, will be on par with a 3080 or even trades blows. Even an RVII situation this time will be surprising here, not because of AMD renaming their upcoming workstation Navi products to their discrete graphics segment (they did that with the MI50 because they needed to show their presence in the high-end discrete graphics market and clear the volume of MI50's), as Lisa Su already confirmed that we'll be seeing them getting back to high end, but the leap of Ampere over Turing will be massive. A little more than Maxwell to Pascal.
It would be a big mistake to sell a ton of cards? Getting rid of all the extra silicon wouldn't reduce the price significantly?
I guess we'll agree on disagreeing.
If Nvidia continues with these crazy prices they'll end up destroying the high-end PC market. I'm sure the 2080Ti is the worst selling 'Ti' of all time, by far. And my guess, Turing cards in general.
Many people will abandon PC and get a next-gen console. They look quite beefy this time and will have a reasonable price, unlike Nvidia.
I think another generation of crazy prices, to force people into raytracing, will be the real mistake.
Nvidia can keep selling RTX cards for the people that are willing to pay for them. But don't force it on everybody.
For many people it's still not worth it: going back to 1080p or playing at 30 fps on a 1.000 dollar card. They would rather wait a couple of generations for the tech to evolve. And developers are going to continue to support baked lighting for many years anyway.
I think a high-end GPU at a reasonable price ($699 for a Ti) with no RTX would be very healthy for the PC market.
Thanks for the insight. I was leaning more towards AMD because they provide better bang for buck according to all these benchmarks I've been looking at recently. But I constantly hear about how the RTX cards come with extra features like DLSS. And now you mention VRS and the importance of it so I guess I'll be sticking with team green.
The 2060 should be a good card for at least 2 years. By then I may be able to put together a mid range PC that should beat either next gen consoles.
I don't buy a £1000 GPU just for this year's RTX implementation, I want it proofed for at least a few years in the future and it seems to be all about RTX in the future. Are you telling me that the 3080 isn't going to have way more RT cores? How can it not be a good thing - having more cores means more paths and bounces can be calculated, and the better the effect would look. I want to believe you that it wouldn't be faster because the shading is the bottleneck but 46 and 68 RT cores sounds PUNY!
Eitherway, performance needs to go up considerably now that we know RTX is what's coming.
Last edited:
Don't worry, just because 1st gen of RTX didn't sell insane numbers it didn't start death of PC as gaming platform.
Or are you hoping it did?
I don't see Nvidia putting anywhere near that much ram on even the higher end cards. More ram is one of the carrots they use to move people off the cards they already own. The only way that much ram would make sense is if they have some crazy new AI features that chew up vram like it's nothing.
Could just be future-proofing. Higher resolution textures that are drawn further could probably increase how much VRAM you need quite quickly.
That's my point. They don't really have any incentive to future-proof their cards. They want customers to buy the next product, and more ram is one of the incentives they give users to upgrade.
None of us have any idea how Nvidia will improve raytracing performance for next gen. What I'm saying that simply increasing the number of RT cores won't necessarily result in massive performance improvements for raytracing because it's the shading part after BVH and ray intersection checking that is expensive in terms of processing time.
I do agree that we need better performance for it but you can't simplify it to "more RT cores = 100 fps with raytracing". That will only work to a certain point before making the BVH traversal and intersection checking faster yields no further benefits if it is waiting for the rest of the pipeline. If you then add more rays to cast then you again run into the same issue that the shading will take more horsepower. That's how I understand it.
Likewise you can't compare RT cores between generations. For all we know Nvidia may have significantly improved their performance for 30xx series so maybe 60 RT cores on 3080 is better than 68 RT cores on 2080 Ti.
So to improve raytracing performance you need to improve several aspects on both software and hardware level. Software is handled by API vendors (MS, Vulkan group etc), Nvidia and game developers. Hardware side is up to Nvidia and we will probably see what they have done on that front at GTC in March.
It's fascinating to me how my 1080Ti is still rocking anything I'm throwing at it. I can still max out everything and get 60fps as long as I scale back the resolution. Many believe that the next step below 4K is 1440p and that it's too much of a down grade but this couldn't be further from the truth. Custom resolutions do wonders. If 4K60 is a no go, you can always try 2088p, 2016p, 1944p, 1872p, 1800p, 1728p, 1652p, 1620p, etc. There are so many games that I'm running at all these custom resolutions. Monster Hunter World for example is struggling in the 40s fps at 4K, but drop that resolution down to 1872p and boom, I get that locked 60fps. From where I sit, it's basically undistinguishable. Assuming I'd be willing live with 60fps with some small hiccups/drops here and there, I could do 1944p or even 2016p.
Then there are the frivolous ultra settings that honestly aren't all that better than some high or very high settings who also have the benefit of not taxing the ressource by an additional 20+%. Seriously, volumetric lighting, ambiant occlusion, shadows and the likes can be scaled back a notch without making any apparent visual difference. I once infliged myself the perogative to run everything maxed out, but not anymore.
So live on, 1080Ti, you amazing beast :)
I understand that it isn't a guarantee, we can't know for sure but we need way, way higher performance and if they've already hit a wall then maybe they shouldn't have bothered.
I still want way more RT cores. I don't care if the CEO of Nvidia told me otherwise, 68 sounds puny for light rays. Give me the cores, I'll turn down the other settings. And give me like 4-10x the cores, you can keep the tensor ones.
Raytracing isn't a new concept itself. Nvidia bought out Mental Ray 13 years ago, back in 2007! I don't think they're going to suddenly improve the calculations for it by the landslide that we need. I'm hoping to god it's the cores and they were just being tight with them to get guaranteed upgrades going into next-gen. Raytracing might be all we have to show for next-gen for years.
Last edited:
GPU designs take years. Nvidia are likely to have at least two teams currently working in parallel on future designs. There is no Future Proofing as the next design is likely 2.5 years away. Ampere, or whatever it's named, will be designed for whatever software developers have been working on since 2018 or so.
Developers have been making their games using Turing Quadro GPUs with up to 48GB VRAM for the last couple of years. Nvidia has a very good idea about how much VRAM this upcoming GPU will likely need to be suitable for the next 2-3 years of releases.
Developers have been making their games using Turing Quadro GPUs with up to 48GB VRAM for the last couple of years. Nvidia has a very good idea about how much VRAM this upcoming GPU will likely need to be suitable for the next 2-3 years of releases.
Pheww, I can't wait to build a brand new PC using these new cards.
My desktop is currently using a GTX 780...XD
My desktop is currently using a GTX 780...XD
They won't show anything this far from launch. They need to sell their current products. GTC is the most likely event for HPC GA100 announcement, with likely wide availability at the end of 2020. New GeForces and Quadros will be there only if they'll go on sale in a month or less from that day.
We don't know anything solid about either big Navi or any Ampere GPU right now.
This "17% faster" rumor is based on just one result in an obscure benchmark, with clocks and drivers being in an unknown state.
We also know nothing about architectural changes which Ampere will have over Turing and thus can't really predict its performance even from SP numbers.
Because without Nvidia we have had so many high end options to choose from!
Also do note that "high end" is almost solely defined by pricing. When you say that high price will destroy high end market you show that you don't really understand how market segmentation work.
Yes, because Nvidia doesn't sell anything at $400 or $500. Oh, wait.