New nvidia GPUs coming soon?

Admin

Administrator
Moderator
Messages
2,915
#1
Code:
Name          GPU    SM count  Cuda cores  Memory       Mem Speed  Bus     Bandwidth  Cache   TGP  MSRP
RTX 4060      AD107  24        3072        4/8 GiB G6   16Gbps     128bit  256GiB/s   32MiB
RTX 4060ti    AD106  28        3584        8 GiB G6     16Gbps     128bit  256GiB/s   32MiB
RTX 4070      AD106  36        4608        8 GiB G6     18Gbps     128bit  288GiB/s   32MiB   230W
RTX 4070ti    AD104  56        7168        10 GiB G6    18Gbps     192bit  432GiB/s   48MiB   270W
RTX 4080 12GB AD104  60        7680        12 GiB G6    18Gbps     192bit  432GiB/s   48MiB   285W  899$
RTX 4080 16GB AD103  76        9728        16 GiB G6x+  21Gbps     256bit  672GiB/s   64MiB   320W  1199$
RTX 4080ti    AD103  84        10752       16 GiB G6x+  21Gbps     256bit  672GiB/s   64MiB
RTX 4090      AD102  128       16384       24 GiB G6x+  21Gbps     384bit  1008GiB/s  96MiB   450W  1599$
RTX 4090ti    AD102  142       18176       48 GiB G6x+  21Gbps     384bit  1152GiB/s  96MiB   800W
Initially there were only one rumoured version of the AD103 i would surprise if the AD103 die would get used only for one GPU. Most likely it will be used for at least 2 (to minimize the amount of die having to be thrown away). Large dies are more likely to be used for multiple models due to large price difference between for example 3080 and 3080ti, for this reason we can expect a 4090ti or a titan card.

Models with MSRP listed have been launched and had their specs confirmed by nvidia (so far the leakers have been correct).
 

Admin

Administrator
Moderator
Messages
2,915
#3
The site techpowerup also has alleged specs listed for these, these do not agree with the specs leaked by various youtubers, we will eventually find out who (if anyone) was correct here.

AD102 https://www.techpowerup.com/gpu-specs/nvidia-ad102.g1005
AD103 https://www.techpowerup.com/gpu-specs/nvidia-ad103.g1012
AD104 https://www.techpowerup.com/gpu-specs/nvidia-ad104.g1013
AD106 https://www.techpowerup.com/gpu-specs/nvidia-ad106.g1014
AD107 https://www.techpowerup.com/gpu-specs/nvidia-ad107.g1015
 

Admin

Administrator
Moderator
Messages
2,915
#4
Update from kopite7kimi
Bad 'news' with regard to the 4070 and to a less extent the 4080. Nothing with regard to the potential ti models.

1656014427185.png


Still no good news with regard to power consumption.
1656014483557.png


Code:
RTX 4070  (old rumored specs)  AD104  60        7680        12 GiB G6    18Gbps     192bit  432GiB/s   48MiB   300W
RTX 4070  (new rumored specs)  AD104  56        7168        10 GiB G6    18Gbps     192bit  432GiB/s   48MiB   300W
 

Admin

Administrator
Moderator
Messages
2,915
#5
Should you buy a GPU now or wait for lovelace/RDNA3?
With computer hardware there will always be reasons to wait but when you wait you miss out on enjoying nice hardware such as RTX 3090. One very important factor is how good your current hardware is. If you are satisfied with your current GPU then you may aswell wait for something better but if you have significant performance issues right now it's probably better just to upgrade now when you can for a decent price (new/used).

There is a lot of uncertainty with regard to upcoming GPUs, we don't know how expensive they are going to be. Judging by the leaks it seems like the power consumption for the 4090 and 4080ti will be really bad and if the trend continue the MSRP of the 4090 will be around 1999$ while you need to pay even more (such as 2699$) for a 4090ti. That would not be something particularly great to wait for in terms of price/performance.

Rumours indicate that the lower end cards (such as the 4080) will come later and these cards while likely significantly cheaper will not offer that big of a performance uplift over current graphics cards (such as 3080 12 GiB).

GPUs used to be really overpriced but latenly nvidia has gradually cut the FE prices to the point where AIBs where forced to sell at big losses to get rid of their inventory.

Code:
            Current  earlier
3090ti      1100$    1499$
3090        1000$    1299$
3080ti      900$     1099$
3080 12GiB           799$
3080 10GiB  699$
3070ti      599$
3070        499$
3060ti      399$
3060        329$
3050        249$
That would leave the 3060, 3060ti, 3080 12GiB, 3080ti, 3090, 3090ti as the nvidia options to seriously consider.

When you look at real prices however you will find that sometimes 3080ti will be priced well below the current MSRP (well below 1000$) potentially making it a decent option in terms of price/performance.

3070ti and 3070 are still overpriced, ignore these on the new GPU market. The 3080 12 GiB seems to be the sweet spot for gaming.

nvidia hinted that the price cuts would only be temporary. The reason for the cut in official MSRP was likely that most people were buying the 3080 rather than the more expensive GA102 models when they could (at MSRP) and this would have forced nvidia to lock down dies that could have been used for 3090/3090ti to be used for cheaper models to sell GPUs at all, luckily it seems like nvidia instead opted for adjusting their priccing to sell more 3090 and 3090ti versions. Nvidia rather sell you a 1299$ GPU than a 699$ GPU.

The AMD prices has also gone down significantly to the point where you can get the 6900xt well below its MSRP of 999$.

The prices on the used market has also fallen significantly to the point where you might get a pretty good deal by not buying new.
 

Admin

Administrator
Moderator
Messages
2,915
#6
Waiting now can backfire badly
The risk you take by waiting is that nvidia may discontinue production of 3000 series GPUs leaving only 4000 series (such as 4090 for 2600$) as an option, this is pretty much what happened with the ampere launch (a lucky few got a 3080 for 799$ but most ended up overpaying for a 3080 or had to buy the 3090 instead for 1499$+

It's unlikely that the shortage will be that bad with lovelace but it's not something that cannot happen again. Since the crypto market isn't very rational etherium could end up going much higher again (while not actually switching to proof of stake) or some other coin that is profitable to mine with a GPU might go up massively.


But we can also see nvidia having to sell the 4090 for a decent price (such as 1499$) due to them having ordered too much production at TSMC and incoming competition from AMD. Nvidia might also have to lower the price further to get rid of 3000 series stock.


Nvidia probably isn't going to do general price cuts unless they are forced to by the competition. Otherwise they will set the lovelace prices so high people keep buying 3000 series. Another potential reason for high prices is the manufacturing being very expensive, we don't really have any details on this but it might result in 3000 and 4000 series GPUs co-existing for a while until the manufacturing cost drops.
 

Admin

Administrator
Moderator
Messages
2,915
#7
Raytracing is cool but even the 3090ti is too weak for it
When it comes to framerates you definitely want to be consistently below 0.01 seconds between each frame. The issue is that even with DLSS at 1440p the performance will not really be there.

https://youtu.be/1rYWxAtE3V8?t=420

https://www.youtube.com/watch?v=09ZcasVAziw

This is the biggest reason for waiting for the 4000 series. By buying a 4090 you will likely get very nice framerate even with raytracing max and you will not have to significantly drop the resolution.

Hopefully the 4000 series will also come with displayport 2.0 support enabling 4K and 360hz.
 

adolf512

Staff member
Moderator
Messages
367
#8
Price prediction
Rather than delaying the launch nvidia is likely to launch 4090 (and maybe also 4080ti and 4090ti) at a such high price that people will keep buying the 3000 series.

here is my guess (prices when 4000 series arrives) if their is no significant AMD competition

4090ti: 2699$
4090: 1999$
4080ti: 1649$
4080: 1399$
3090ti: 1150$
3090: 1000$
3080ti: 850$
3080 12GB: 700$
3080 10GB: 650$
3070ti: 550$
3070: 470$
3060ti: 400$
3060: 330$
3050: 250$

Nvidia isn't going to cut prices much more unless they face serious competition from AMD. Nvidia would stop instead hike the 4000 series prices and stop making the more affordable 3000 equivalent if they could get away with it.

It is worth nothing that 3080ti, 3090 and 3090ti are all significantly yield-constrained which might cause them to remained priced too high to offer great value (except for very temporary deals)

Based on current price development (as of september 5) it seems like the 3000 series will be kept down as nvidia is pushing out more GPUs.

4090ti: 1999$
4090: 1599$
4080ti: 1349$
4080 16GB: 1199$
3090ti: 1000$
4080 12GB: 899$
3090: 860$
3080ti: 700$
3080 12GB: 620$
3080 10GB: 600$
3070ti: 500$
3070: 430$
3060ti: 350$
3060: 280$
3050: 200$

This is assuming the 4080ti will be based on AD103.

Update (september 20): the predicted prices for the 4080 16GB and the 4090 were correct (launched models in bold).

The 4080 12GB did come as a surprise, most people did not expect nvidia to launch that to compete directly with 3090 and 3090ti but nvidia did that anyway. Expect limited supply of the 4080 12GB since it uses the full die and nvidia has 3000 series GPUs to sell you (they are still making ga102 chips).
 

Admin

Administrator
Moderator
Messages
2,915
#9
We might finally see a real successor to the RTX titan

1658957456962.png

That's 2 SMs disabled.

They will probably reserve the full die for quadros.

Code:
new RTX 4090ti AD102  142       18176       48 GiB G6x+  21Gbps     384bit  1152GiB/s  96MiB   800W
old RTX 4090ti AD102  144       18432       24 GiB G6x+  21Gbps     384bit  1152GiB/s  96MiB   600W
 

Admin

Administrator
Moderator
Messages
2,915
#10
4070 may have been upgraded
This is likely a case of rebranding (4070ti renamed to just 4070).

1659714014993.png


Because of this we also need to update the 4060ti (to what used to be the 4070). The old specs mentioned here were just my speculation.
Code:
new RTX 4060ti AD104  56        7168        10 GiB G6    18Gbps     192bit  432GiB/s   48MiB   300W
old RTX 4060ti AD104  46        5888        10 GiB G6    18Gbps     192bit  432GiB/s   48MiB?
 

Admin

Administrator
Moderator
Messages
2,915
#11
RTX 4070ti or 4080 specs may have leaked
This is not the full AD103 die so it might instead be the 4070ti (76 of 84 SMs enabled), that would be 4 off my earlier guess of 72 (9216 cuda cores).

kopite7kimi wrote:
1659723097468.png


This being the 4070ti does fit far better with earlier leaks the old 4070ti being renamed to 4070 indicates that they were making place for a cut down version of the AD103.

Otherwise nvidia would have had to release the full AD103 as 3080ti (or only sell as quadro cards) which would mean also mean the rumor of a 20GiB 4080ti being false, instead we would likely see a 16GiB 3080ti from AD103 in addition to some defective AD102 dies used for it similar to how some defective ga104 dies are used for 3060

Nvidia might also just release a 20GB 4090 version that will have the same number of SMs as the 4090 but 4GiB less VRAM and worse memory bus.

In a later 'leak' the TGP was lower while the memory bandwidth was highes

1661373820355.png
 

Admin

Administrator
Moderator
Messages
2,915
#13
Did nvidia cancel the AD102 4080ti ?
Initially this was posted from an unreliable source
Code:
RTX 4080ti AD102  108       13824       20 GiB G6x+  21Gbps     320bit  840GiB/s   96MiB?  450W
Lately however we never really got any leak confirming this and nvidia probably doesn't want to release a GPU specifically to only use 108 or 142 SMs. Increasingly nvidia has become unwilling to needlessly bin down healthy dies for segmentation purposes, instead they typically try taking a good usage of each die even if it means supplying more quantity of a card there is oversupply of (and cut prices).

Nvidia did end up releasing a 12 GB 3080 that will accompany the 10GB version, they might be doing something similar with AD102 in cases where 1 or 2 memory controllers are defective (such as "4090 20GB").

There isn't really any evidence that nvidia actually planned a 20GiB 4080ti in the first place. Nvidia may have initially planned to needlessly lock down AD102 dies for segmentation purposes or they may have had it as a potential option in the case of very bad yield only to later find out that the dies didn't have enough defect to warrant an official 108SM version.

If nvidia does release a a 20GiB AS102 GPU with 108 SMs it will probably be for their quadro lineup (in small quantities), it would probably not be released for the wider consumer market (in part due to limited supply of suitable dies).

Nvidia was actually close to releasing a 20GiB 3080ti for segmentation purposes in 2021 but they changed their mind and sold them as 3090 instead.

https://www.tekfiz.com/2021/03/nvidia-caught-using-old-rtx-3080-ti-gpu.html

Eventually nvidia just dropped the price of the 3090 (after the 3090ti was launched) finally offering people more than 12GiB of VRAM for a decent price. The 4000 series will be better in the sense that the 4080/4080ti will come with 16GB+ of VRAM. In gaming 12GiB of VRAM is fine for the most part (even at 4K) but that might change in the future and there are also some productivity tasks that benefit from having more than 12GiB of video memory.
 

Admin

Administrator
Moderator
Messages
2,915
#14
The 4080 with 12GiB memory may have been canceled
Of course nvidia might launch these models later but it would be a bit confusing for the consumers.
Code:
RTX 4080   AD103  76        9728        12 GiB G6x+  21Gbps     192bit  504GiB/s   64MiB
RTX 4090   AD102  128       16384       20 GiB G6x+  21Gbps     320bit  840GiB/s   96MiB?
 

Admin

Administrator
Moderator
Messages
2,915
#15
If nvidia wanted to compete hard
If yields were good it would make maybe have made sense to upgrade the specs a bit

Code:
RTX 4050ti AD107  24        3072        4/8 GiB G6   16Gbps     128bit  256GiB/s   32MiB
RTX 4060   AD106  28        3584        8 GiB G6     16Gbps     128bit  256GiB/s   32MiB
RTX 4060ti AD106  36        4608        8 GiB G6     18Gbps     128bit  288GiB/s   32MiB   230W
RTX 4060   AD104  56        7168        10 GiB G6    18Gbps     192bit  432GiB/s   48MiB   270W
RTX 4060ti AD104  60        7680        12 GiB G6    18Gbps     192bit  432GiB/s   48MiB   285W
RTX 4070   AD103  76        9728        16 GiB G6x+  21Gbps     256bit  672GiB/s   64MiB   300W
RTX 4070t  AD103  84        10752       16 GiB G6x+  21Gbps     256bit  672GiB/s   64MiB   330W
RTX 4080   AD102  128       16384       20 GiB G6x+  21Gbps     320bit  840GiB/s   96MiB   360W
RTX 4080ti AD102  136       17664       24 GiB G6x+  21Gbps     384bit  1008GiB/s  96MiB   375W
RTX 4090   AD102  142       18176       48 GiB G6x+  24Gbps     384bit  1152GiB/s  96MiB   420W
RTX 4090ti AD102  144       18432       48 GiB G6x+  26Gbps     384bit  1248GiB/s  96MiB   850W
This of course is was very unlikely to happen in part due to the competition from AMD probably not being very strong.
 

Admin

Administrator
Moderator
Messages
2,915
#17
My 4090 price prediction was exactly correct (1599$) and i also got the price of the 4080 16GiB exactly right (1199$). I am however surprised that they released the 4080 12GiB for 899$ since that is going to cannibalize on the 3000 series (uses the full AD104 die).

https://www.youtube.com/watch?v=Uo8rs5YfIYY

1663688662260.png


Nvidia did end up with the confusing scheme of having 2 completely different 4080 versions. The 16GiB version will be much better (not just more
memory). The leaked specs for the 4080 16GiB ended up fully correct. Nvidia may have renamed the 4070ti to 4080 12GiB to obscure the fact that they are massively raising their prices relative for each performance tier.

This does hint at further price cuts of the 3000 series

1663688817641.png
 

Admin

Administrator
Moderator
Messages
2,915
#19
Performance uplift estimate using the digital foundry data
They didn't compare the 4090 directly to the 3090ti but we can easily do some maths

4090 with DLSS 3: +134%
DLSS 3 over DLSS 2: +55%

Uplift without DLSS 3: +50%

 

Admin

Administrator
Moderator
Messages
2,915
#20
DLSS 3 might make the experience worse
Looking at the digital foundry data DLSS3 might only boost the framerate by about 55%, if half of the frames are fakes that means 22.5% worse base FPS which is a rather high price to pay for fake frames. DLSS3 will also add half of a frame of latency since it will need to wait for the next frame before adding the interpolated frame.

The worse the base framerate is the harder it will be for DLSS3 to do the interpolation well. The issue is that if you reach a good base framerate you are probably close to maxing out your monitor, most screens still don't go above 144hz and nvidia didn't add support for displayport 2.0 (limiting the framerate to 240 at 4K).


We might see something like DLSS3 being beneficial once the 5000 series has come out with displayport 2 and we have (4K, 360hz) screens to properly utilize it. Let's say for example that you have a 144hz screen, if you then get 120fps without DLSS 3 the base fps will need to be dropped to just 72fps in order to add the interpolated frames, in these cases you will actually be better off not using interpolation at all.

Low framerate: DLSS 3 adds way too much latency and isn't accurate.
High framerate: DLSS 3 destroy base FPS due to the limitations of your monitor.

DLSS 3 might be beneficial if you have a monitor of very high refresh-rate (such as 360hz) but in these cases the benefit in terms of better motion is not going to be particularly great since it will be just fine without it (such as 160 fps).
 

Admin

Administrator
Moderator
Messages
2,915
#21
Why the 4090 will be the only model worth even considering
Until other models (such as 4090ti) comes out or nvidia cuts the prices of their inferior card it will be (4090, nothing) that are the by far the best options to consider. This is not like 30 series where you can get a cheaper model and still get similar performance (as long as 12GiB vram is enough).

The 4080 12GB will have only have half the bus width as the 3080 12GB so it's actually a 60-class card being renamed as an 80 class card to fleese poor people who are unable to afford the 4090 and push people to buy 30 series cards instead (which are much cheaper to make and they have a lot of current stock to sell).

4090ti/titan: 144 or 142 SMs
4090: 128 of 144 SMs (88.9%)
4080 16GB: 76 SMs (52.8%)
4080 12GB: 60 SMs (41.7%)

3090ti: 84 SMs
3090: 82 SMs
3080ti 80 SMs
3080 12GB: 70 SMs
3080 10GB: 68 SMs

3070ti: 48 SMs
3070: 46 SMs (54.8%)
3060ti: 38 SMs (45.2%)
3060: 28 SMs (33.3%)

We are probably going to see even more expensive AD102 cards in the future such as 4090ti for 1999$ or "titan lovelace" for 3999$. It would however be very surprising if these cards turn out to actually offer decent value. The real successor to the 3090/3090ti hasn't been revealed yet.

You would expect the price per SM to go up as you go up in the stack (due to larger dies being harder to manufacture) but we instead see the exact opposite of that. You get more raw power for your money buy buing a card like the 3080ti and 4090 rather than cheaping out with a 3060 or "4080 12GB".
 
Top