Monday, August 1st 2022

NVIDIA GeForce RTX 40 Series "AD104" Could Match RTX 3090 Ti Performance

NVIDIA's upcoming GeForce RTX 40 series Ada Lovelace graphics card lineup is slowly shaping up to be a significant performance uplift compared to the previous generation. Today, according to a well-known hardware leaker kopite7kimi, we are speculating that a mid-range AD104 SKU could match the performance of the last-generation flagship GeForce RTX 3090 Ti graphics card. The full AD104 SKU is set to feature 7680 FP32 CUDA cores, paired with 12 GB of 21 Gbps GDDR6X memory running on a 192-bit bus. Coming with a large TGP of 400 Watts, it should have a performance of the GA102-350-A1 SKU found in GeForce RTX 3090 Ti.

Regarding naming this complete AD104 SKU, it should end up as a GeForce RTX 4070 Ti model. Of course, we must wait and see what NVIDIA decides to do with the lineup and what the final models will look like.
Sources: @kopite7kimi (Twitter), via VideoCardz
Add your own comment

121 Comments on NVIDIA GeForce RTX 40 Series "AD104" Could Match RTX 3090 Ti Performance

#26
ppn
I was hoping 4070Ti is the cut GA103 256 bit. 192 bit is a joke, my unrealistic expectations are shattered. But perhaps a SUper.
DavenI could be wrong but isn’t CUDA core count calculated by adding FP32 cores (7680) to FP64 cores (3840) for a total of 11,520 cores. That’s 10% higher cores than the 3090Ti at 12% lower power. I mean it’s not great but it’s not bad either.

I remember all the rumors not knowing how to count CUDA cores before the 3000 series launch. Looks like that might be the case again.
No FP64 (double) performance is executed at 1:64 speed, it is not additive, as far as I understand.

The confusion with GA102 Ampere is that 5376 Cuda can only do fp32 and 5376 Cuda that can execute both fp32 or int32, but not at the same time, so a total of 10752, but when INT32 is running the FP32 is greatly reduced.

This could change in Ada, 7680 fp32 AND 3840 int32 separately not shared =11520
Posted on Reply
#27
lexluthermiester
ppnBut cut to 192 bit bus and using same memory speed G6X.
That better not be true. 256bit mem bus or NVidia can suck a d... duck, yes, suck a duck. Seriously NVidia? Did we not learn from past mistakes? The GTX460 comes to mind, as does the GTX970...
Posted on Reply
#28
Daven
ppnI was hoping 4070Ti is the cut GA103 256 bit. 192 bit is a joke, my unrealistic expectations are shattered. But perhaps a SUper.



No FP64 (double) performance is executed at 1:64 speed, it is not additive, as far as I understand.

The confusion with GA102 Ampere is that 5376 Cuda can only do fp32 and 5376 Cuda that can execute both fp32 or int32, but not at the same time, so a total of 10752, but when INT32 is running the FP32 is greatly reduced.

This could change in Ada, 7680 fp32 AND 3840 int32 separately not shared =11520
Thank you for the clarification. I guess I mixed up FP and int terminology.
Posted on Reply
#29
IceShroom
john_So, the new 400W model matches the old 450W model? Considering they moved from the "bad" 8nm Samsung node to the "great" 5nm TSMC node, it's not exactly a breathtaking result.
I believe either performance is higher or wattage much lower. 50W less for the same performance looks too little to me.
Samsung's 8nm is quite good node. Much better then Nvidia fansboys like to tell every one. Ampere has high TDP because Ampere is inefficient.
Posted on Reply
#30
lexluthermiester
IceShroomAmpere has high TDP because Ampere is inefficient.
Not quite. Ampere has a high TDP because it is doing hardware real-time raytracing, which is a VERY complex and compute heavy type of task. When not RTRT is not being performed, Ampere GPUs are good on power. Turing is/was no different. AMD's RTRT functionality is no different, turn on raytracing and power usage takes a big bump.
Posted on Reply
#31
Daven
lexluthermiesterNot quite. Ampere has a high TDP because it is doing hardware real-time raytracing, which is a VERY complex and compute heavy type of task. When not RTRT is not being performed, Ampere GPUs are good on power. Turing is/was no different. AMD's RTRT functionality is no different, turn on raytracing and power usage takes a big bump.
AFAIK, TPU power graphs and perf/W graphs are done without RT enabled. I think theses commenters are referring to these graphs when commenting on efficiency.
Posted on Reply
#32
napata
IceShroomSamsung's 8nm is quite good node. Much better then Nvidia fansboys like to tell every one. Ampere has high TDP because Ampere is inefficient.
It's not that simple. Just look at A4000. A big part of efficiency just depends on the performance you target in terms of clock speed and the yields you target. Higher clockspeeds means less efficiency. It's possible bad yields resulted in the 3000 series being relatively inefficient.

Just look how efficient A4000 is. Less power than a 6600xt while performance is equal to a 3060Ti. Just a result of a bigger chip with lower clocks and possibily binned for low voltages.

www.techspot.com/review/2342-nvidia-rtx-a4000/#:~:text=Because%20of%20the%20lower%20core,less%20than%20the%203070%20Ti.
Posted on Reply
#33
lexluthermiester
DavenAFAIK, TPU power graphs and perf/W graphs are done without RT enabled.
You need to take a closer look.
DavenI think theses commenters are referring to these graphs when commenting on efficiency.
That's an assumption on your part. The devil is in the details.
napataHigher clockspeeds means less efficiency.
This too!
Posted on Reply
#34
TheinsanegamerN
holyprofWith electricity costs going up in the whole world, how come the 4070Ti draws more than double of power compared to a 1080 (180W)? New generations should bring more performance at the same power. We are getting expensive space heaters instead.
Add to that 400W heating in your room the costs of air conditioning and it's becoming a quite expensive hobby.
On a second thought, the RTX 3070Ti has awful performance per watt, especially compared to the RTX 3070 and RX6800, so there's hope my intended upgrade RTX 4070 will perform close to the Ti version but at much lower power draw.
Once again it needs to be stated that if an extra 200-300 watts makes you pucker at electric costs you cant afford a PC anyway. Look into how much power your HVAC takes, or those new "green" electric cars.

Power draw is going up because capability is going up and there is demand for mroe powerful cards. Economey 101. If you dont want high power draw you can still buy the mid range cards and benefit from the major perf/watt improvements, or undervolt to dramatically reduce power use while only losing <5% performance.
lexluthermiesterThis too!
Everyone always assumes samsung's process sucked just because nvidia OCed their parts too high to dominate said graphs out of the box. Sub 2 GHz the 8nm node is great.

What I wouldnt give to find one of those A2000s to upgrade my SFF box.
ColddeckedNvidia was sandbagging with maxwell. So much overclocking potential.
Honestly though that was a good thing, it resulted in maxwell gaining the reputation for being WAY more efficient then AMD, a badge it still wears today, and also allowed us enthusiasts to have tons of fun OCing them to the moon.
Posted on Reply
#35
Lei
@AleksandarK
Rumors before 30 launch were under estimating CUDA count by a large number.
It's likely they're overestimating performance this time to reach an equilibrium.

The only impressive leak was the cooler which wasn't from kopite7kimi anyway.

look at his numbers:

he's predicted half the CUDA cores for 3090. leaked 5248 vs real 10496
his 4352 for 3080 -which is same cores as 2080Ti had- is less than half the actual 8960
his leak is exactly half for 3070 and 3070Ti


And here is his crazy 20gb rumors about 3080Ti

Posted on Reply
#36
lexluthermiester
TheinsanegamerNWhat I wouldnt give to find one of those A2000s to upgrade my SFF box.
Right there with you. I'd like to find a 12GB version but a 6GB would do.
Posted on Reply
#37
TheinsanegamerN
lexluthermiesterRight there with you. I'd like to find a 12GB version but a 6GB would do.
I'd be fine with either. My RX560 is getting long int he tooth and 4GB VRAM just doesnt do it anymore for certian games. The 6400 and 6500 were major dissapointments and the 3050 is too power hungry. The intel card, if it ever get shere, likely will be slower then what I have.

I've debated trying to build something into a bk 623 and get rid of a SFF case altogether.
Posted on Reply
#38
Beertintedgoggles
lexluthermiesterYou need to take a closer look.

That's an assumption on your part. The devil is in the details.
I did take a closer look and Techpowerup measures gaming power consumption with ray tracing disabled:

Gaming: Cyberpunk 2077 is running at 2560x1440 with Ultra settings and ray tracing disabled. We ensure the card is heated up properly, which ensures a steady-state result instead of short-term numbers that won't hold up in long-term usage.
Posted on Reply
#39
lexluthermiester
BeertintedgogglesI did take a closer look and Techpowerup measures gaming power consumption with ray tracing disabled:

Gaming: Cyberpunk 2077 is running at 2560x1440 with Ultra settings and ray tracing disabled. We ensure the card is heated up properly, which ensures a steady-state result instead of short-term numbers that won't hold up in long-term usage.
And what did you see when you looked at the RTRT enabled numbers? Hmm?

I'm not going to hand-hold anyone. Pick a card review and be diligent when reading it.
Posted on Reply
#40
Beertintedgoggles
lexluthermiesterAnd what did you see when you looked at the RTRT enabled numbers? Hmm?

I'm not going to hand-hold anyone. Pick a card review and be diligent when reading it.
Don't know because the graphs show with ray tracing disabled. That was the point wasn't it?

Pick any of the card reviews you want on this site, go to the power consumption page then click on the "Power Consumption Details" portion. That's as much hand holding as I'm willing to give.
Posted on Reply
#41
Unregistered
Wonder what they mean by beat it? Maybe in Ray Tracing using the latest DLSS.
#42
Lionheart
Bomby569There is clearly a barrier neither Nvidia or AMD can pass through, they cannot increase performance in a meaningful way in the 2 year cycle without going crazy on power draw, no point in beating the dead horse. Skip this generation if you don't agree with the way things are going or deal with it.
People want big performance leaps and 4k 200Hz with the 2 year cycle with the same power draw and they are being as unrealistic as Nvidia and AMD

Set a wattage limit, stay in the limit no matter what they release, or shut up about it.
Don't tell us to shutup about it if you can't handle proper criticism.
Posted on Reply
#43
lexluthermiester
BeertintedgogglesDon't know because the graphs show with ray tracing disabled. That was the point wasn't it?

Pick any of the card reviews you want on this site, go to the power consumption page then click on the "Power Consumption Details" portion. That's as much hand holding as I'm willing to give.
Did you cherrypick this review or did you look at others?

Then there's this review where they tested everything.

Owning a 2080 myself, I can tell you the power usage goes up when RTRT is enabled VS disabled.
Posted on Reply
#44
Beertintedgoggles
lexluthermiesterDid you cherrypick this review or did you look at others?
It appears to me you are cherry picking here. Go to the home page, click the Reviews link and set the category for Graphics Cards. Pick a random one and go to the Power Consumption page and click on the drop link for "Power Consumption Testing Details". I had no idea before whether those measurements were taken with or without ray tracing so I followed your advice earlier and decided to read closer. Found out that power consumption measurements on TPU are done without ray tracing.
Posted on Reply
#45
Daven
lexluthermiesterAnd what did you see when you looked at the RTRT enabled numbers? Hmm?

I'm not going to hand-hold anyone. Pick a card review and be diligent when reading it.
What are you talking about? A commentor said Nvidia is inefficient. You said something about RT. I said the inefficiency referred to was general rasterization. You told me to look closer. Another commentor said RT was disabled for power tests. Please just stop!
Posted on Reply
#46
lexluthermiester
BeertintedgogglesIt appears to me you are cherry picking here. Go to the home page, click the Reviews link and set the category for Graphics Cards. Pick a random one and go to the Power Consumption page and click on the drop link for "Power Consumption Testing Details". I had no idea before whether those measurements were taken with or without ray tracing so I followed your advice earlier and decided to read closer. Found out that power consumption measurements on TPU are done without ray tracing.
DavenPlease just stop!
If you are not going to READ and understand context, take your own advice.
Posted on Reply
#47
Beertintedgoggles
lexluthermiesterIf you are not going to READ and understand context, take your own advice.
You got called out on an inaccurate statement. Reviews here on TPU show the power measurement numbers with ray tracing disabled. In a conversation / argument, what you are doing is called misdirection. And by the way, it was your advice I followed.
Posted on Reply
#48
Lei
Here are Kopite7kimi's leaks about 3070Ti
just see how wrong he is


7424 cores instead of real 6144
320bit instead of real 256bit
10gb instead of real 8

If he doesn't tweet, many youtubers will have no way to buy bread.
Posted on Reply
#49
Unregistered
i heard we pay to many attention to this attention whore
LionheartDon't tell us to shutup about it if you can't handle proper criticism.
criticize me away, i don't mind, it will still not change facts and tech limitations
#50
PapaTaipei
400 Watts and only 7K CUDA cores... also ppl arguing that the 980ti could OC needs to realize that 99.99% of ppl don't OC their GPU.
Posted on Reply
Add your own comment
Nov 19th, 2026 04:59 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts