New nvidia GPUs coming soon?

Admin

Administrator
Moderator
Messages
3,326
#36
Why chiplets and dual-gpu cards doesn't make sense anymore
The dual-gpu approach (such as raden HD 5970) made sense when GPUs were low in power-draw and reviewers rarely did any proper tests for frame consistency. Hardly any new games support more than one GPU anymore and the chiplet approach would just add latency. The only way forward is to make GPUs more and more power-hungry, eventually over 1000W will be needed to properly utilize the top GPU.

But nvidia has managed to make very powerful monolithic dies. A single 4090 will draw up to 450W by default and with unlocked power-budget it can draw as high as 600W. The 3090 die was so large that the card would throttle under furmark since even the miniimum voltage was too high, the reason why this isn't a problem in normal usage is that most loads will only load a part of the GPU maybe 40% allowing the GPU to boost higher.

The 4090 isn't even the full die. The full die does have 12.5% more SMs but the actual performance uplift in games would of course be a lot smaller unless there is a significant increase of the power-budget. Power-consumption is the limiting factor, not yield.

Many games will not even load all gpu cores, this is especially true at lower resolutions, this becomes more and more of an issue as you make the die larger.
 

Admin

Administrator
Moderator
Messages
3,326
#38
7900XTX failed to beat the 4080
It's embarrassing how AMD failed to beat what would in the past been a 70 class card. Sure the 4080 is 200$ more expensive (if we go by MSRP) but then you get nvidia drivers and features in addition to better efficiency and also an overall better card.

If nvidia cut the price of the 4080 by 200$ it will be game over for AMD but they probably do not even have to do that, that's how bad RDNA3 turned out to be. It's in raytracing when you need performance the most.

 

Admin

Administrator
Moderator
Messages
3,326
#39
The 4070ti is decent value at 800$
Thanks to nvidia dropping the price to 800$ it's actually just fine for the price. Especially compared to the nvidia alternatives. It's just around 200$ more than the 3070ti while being 44% faster and having 50% more GDDR5x.



It delivers similar performance as the 3090ti while drawing significantly less energy. The 3090ti had an MSRP for 1999$ just a year ago and was later discounted down to around 1100$ before they got rid of that inventory, a lot of people bought that and for the time it was a fine purchase.

Main drawback with the 4070ti is only getting 12 GiB of vram compared to the 24 GiB you got earlier with a 3090 and 3090ti. Currently to get 24 GiB you need to pay up to get the 4090 which is an overkill CPU for gaming.

Sure RDNA2 is a decent budget option just for gaming but if you actually want to do work 4090 is the way to go (unless you are on a tight budget in which case 4070ti, 4080 or used 3090/3090ti are decent options).
 

Admin

Administrator
Moderator
Messages
3,326
#40
30-series GPUs are now awful for the price when bought new
Naively i was expecting nvidia to also drop the prices of the cheaper 30 series cards a bit but that just hasn't happened. A lot of people are still overpaying for cards like the 3070ti which doesn't even have enough vram.

1679903657355.png


3060: 365$ #0 390$ #2 350$ #5 390$ #10
3060ti: 490$ #1
3070: 548$ #3

Interestingly the 3070ti couldn't be found anywhere in that top50

There were however some people buying the 3070ti from newegg for 620$ (after 30$ off with code) placing it #17 in their top20

We see that the 3080 to 3090ti inventory has already been largely sold out but nvidia is still managing to sell entry and level GPUs like the 3070 for very high prices which is why they have not dropped these prices or launching any card below the 4070ti.

 

Admin

Administrator
Moderator
Messages
3,326
#41
Should nvidia cut the 4080 price down to 1100$ ?
Right now almost everyone is just going for the 4090 or 4070ti instead which does make sense given that these cards offer significantly better value for gaming. This is a shame given that the 4080 has the potential to be a great "go to card" unlike the 4090 that is needlessly powerful and costs 1600$.

Earlier finding a 4090 for a decent price was rather difficult pushing many people to pay 1200$ for a 3080 instead but now 4090 prices are going down so you can get a decent model for around 1700$ fairly easily.

What nvidia should maybe do is to launch a 4080ti based of the AD103 using the full die and sell that for 1200$ and drop the current 4080 down to 1100$. That would turn the 4080 into a great option for 1440p gaming (where the 4090 is a massive overkill). Getting 16GiB vram currently seems like a good comprimise for 1440p gaming with 24GiB being a bit overkill and 12GiB barely being enough.

But it might instead be better for nvidia to have the 4080 be unappealing so people go for the 4090 instead, currently the 4090 is selling well and nvidia might not want to change that by making the 4080 a decent buy, hence they might keep the 1200$ price tag.

The 4080 is still a better card than anything AMD has to offer, sure it's 200$ more than the 7900XTX but a lot of people are willing to pay that extra price to get a better nvidia GPU (that draws less power, has better features, better productivity, etc).
 

Admin

Administrator
Moderator
Messages
3,326
#42
4070 incoming for 599$
This further indicates that the 4070 will have 46 SMs and not the 56 rumored earlier.

https://videocardz.com/newz/nvidia-geforce-rtx-4070-to-retail-at-599

before it was a fake leak showing a MSRP of 750$ which was obvious nonsense, given how cut down the card is with 14 of 60 SMs disabled.

Obviously this is a much better buy than a 3070ti for 599$ which is why nvidia hasn't launched it yet, they are still milking people with their old 30-series cards.
 

Admin

Administrator
Moderator
Messages
3,326
#43
The 40-series GPUs has been an improvement despite the high prices
The 30-series GPUs was grossly overpriced for over a year with them not being anywhere near to the MSRP and the MSRPs weren't even that great in the first place.

3080ti had an MSRP of 1199$, same as 4080. The 4080 despite being priced 100$ too much is still a much better GPU than the 3080ti. I chatted briefly with a guy who canceled his 3090ti order to get a 4080 for less money (the 3090ti could no longer be had for 1100$ when the 4080 launched).

The 3080ti did had an inflated MSRP due to the severe shortage but even if you look at the price after the price cuts (900$) a 4070ti is still cheaper and better at 799$.

The 3090ti was dropped down to around 1100$ but even compared to that a 4080 for 1200$ is more bang for the buck, you get similar performance from the 800$ 4070ti now.

The 3080 was released as a good value CPU if you go by the 699$ MSRP but it was basically impossible to get it for that price and it also only had 10GiB of vram. Now it seems like the 4070 will come soon for 599$ at least matching the 3080 12GB while drawing a lot less power.

Hopefully intel will be able to deliver some great GPUs giving us better value but even as of now things have already changed for the better.

RDNA2 cards is also an option for getting a new GPU at decent price/performance for gaming (might not be great for productivity though).
 

Admin

Administrator
Moderator
Messages
3,326
#44
4070 vs 3080 12GB
Arguably the 4070 is overall a better GPU than the 3080 12GB largely due to being a lot more efficient. It's also a bit cheaper than the 699$ msrp of the 3080 10GB (which was practically impossible to get for that price until mid 2022).

1681306796030.png

https://www.sweclockers.com/test/36143-nvidia-geforce-rtx-4070-energieffektiv-och-kompakt/

1681306997388.png

https://www.techpowerup.com/review/asus-geforce-rtx-4070-dual

In the digitalfoundry review it instead lost overall to the 3080 10GB

1681312541279.png

https://www.youtube.com/watch?v=DNX6fSeYYT

1681322204153.png

https://www.youtube.com/watch?v=Zdkr13ah-iQ

1681322572092.png

https://www.youtube.com/watch?v=Zi3hdm_JZ8U
 

Admin

Administrator
Moderator
Messages
3,326
#45
Did nvidia shoot themselves in the foot gimping the 4070?
The 4070 don't seem to be selling particularly well even at launch.

#17 #18 #23 on the amazon top50 (starting at 0 excluding accessories) 2023-04-16

Later it did to better though https://archive.is/Ou4rp but still not in the top5.

Unclear why it doesn't even show up on the newegg top100, maybe it's calculated for a longer period or it hasn't been updated yet?

It's too slow to be an interesting GPU, doesn't really outperform the older 3080 12GB, it just saves electricity and money (100$ cheaper to buy than the 3080 10GB MSRP to buy).

They could have gotten better performance out of it by enabling more cores and pushing for higher frequencies, but nvidia instead decided to save money which resulted in lame performance.

It seems like ordinary consumers just don't care that much for power efficiency. Maybe 4070 will mostly find it's way into prebuilds instead?

The AD104 die used for the 4070 has 14 of 60 SMs disabled meaning around 23% of the die is disabled, this can be compared to the 3070 which only had 2 of 48 SMs disabled.

With AD104 dies with with 2 SMs disabled will instead be reserved for laptops which is why no such GPU came to the desktop, desktop 4070 instead get even worse dies who couldn't even make it into laptops (the end result is still ok since the bad SMs will be disabled).

Still even with 46 SMs (5888 cores) it would have been possible to get better performance simply by pushing for higher clock speeds but unfortunately most 4070 cards are currently powerlimited to 220W of less which does contribute to worse overclocking potential. This however might simply be about saving money since it's more expensive to engineer a card to safely support a higher power limit.

igorslab.de/en/geforce-rtx-4070-brought-to-the-price-point-when-the-compulsion-to-uvp-defeats-reason/

A lot of people have already upgraded to 3080 12GB or better by now. People who wanted that level of performance have to a large extent already upgraded.

Of course people will still be getting the 4070, it's a much better buy than the 3070ti for 600$. You don't really have the option of getting a 3080 for 700$ anymore, the few that remain on the new market tends to be overpriced and nobody who knows what they are going are going to buy any of those.

6900XT and 6950XT are maybe better options if you are just going to use your card for gaming but most people don't buy AMD anyway so it maybe doesn't matter much.

https://www.techspot.com/review/2663-nvidia-geforce-rtx-4070/
 

Admin

Administrator
Moderator
Messages
3,326
#46
4060ti 8GiB incoming for 399$+
A lot of morons are going to buy this card if nvidia releases it for 399$ even though it's clearly obsolete even before launch thanks to only having 8GiB of vram.

The issue is that the AD106 die only has a 128 bit memory bus, had they had a 192 bit bus like the 4070 they could have provided it with 12GiB of GDDR6 and also get better memory bandwidth. Of course they could make a 16GiB version eventually but that would be a bad match for such a weak card and it would be weird to have the cheaper GPU provide more vram.



btw: it's not actually the full die, the full AD106 die will be used for laptops, defective AD106 dies have been saved up to be showed down the throats of ignorant gamers. Of course actually using the full die would still result in awful performance due to the bad memory bus.

The 4060ti is expected to have 34 of 36 SMs enabled and GDDR6 memory (cheaper and worse than GDDR6X).
 

Admin

Administrator
Moderator
Messages
3,326
#47
Why this refusal from nvidia to provide good memory bus on models below the 4080 ?
If we look at the die shots we see that the memory interface doesn't actually take up much die space so why just not add more of that?

4090-die.jpg


As we see here nvidia could have easily provided the 4080 with a 384 bit bus and 24 GiB memory but they decided against that maybe in order to push people towards the more expensive 4090.

What we see here is a mismatch between the computational power and the memory interface where all 40-series GPUs including the 4090 are crippled by inproper memory interface.

The 4090 should have had 32 GiB of GDDR6x (512 bit bus)

The 4080ti should have had 24 GiB of GDDR6x (384 bit bus)

The 4080 should have had 20 GiB of GDDR6x (320 bit bus).

The 4070ti should have had 16GiB of vram (256 bit bus)

The 4060ti should have had 10GiB of vram (160 bit bus)

It's almost like nvidia is intentionally making all models below the 4090 bad in order to push people towards the 1599$ 4090, unfortunately the 4090 doesn't have displayport 2.0 so it's also severely bottlenecked, just in a different manner.

Is is worth noting that the 40-series card does have a lot of cache which does partly mitigate the lack of having bad memory bandwidth.
 

Admin

Administrator
Moderator
Messages
3,326
#48
Top