NVIDIA GeForce RTX 4070: up to 30% faster than RTX 3090 in gaming

[Blade-Runner]

https://twitter.com/kopite7kimi/status/1519164336035745792

https://www.extremetech.com/gaming/335390-rtx-4090-to-offer-24gb-gddr6x-vram-rtx-4080-4070-detailed

In a series of new tweets, @kopite7kimi details the specs for these upcoming GPUs. As has been reported previously, the flagship die will be AD102. Interestingly, the top-shelf “Lovelace” GPU will be outfitted with 24GB of GDDR6X running at 21Gb/s. This is the exact same configuration as the current RTX 3090 Ti. It should be noted that the previous generation Turing-based Titan RTX also had 24GB of GDDR6 memory. It seems that for now, this is as far as Nvidia wants to take memory allocation on its high-end SKUs. There were previous rumors that AD102 might come equipped with GDDR7 memory running at 24Gb/s, but those might be incorrect. What is seemingly firmer at this stage is the GPU’s Total Board Power (TBP). It’s once again listed as 600W, which seems like a given since Nvidia wants to push this card to the absolute maximum level of performance. It might have to in order to fend off AMD for this round.

 

[Comixbooks]

https://videocardz.com/newz/nvidia-...0w-tdp-is-allegedly-twice-as-fast-as-rtx-3090

JULY 2022~!!!!!!!!!!!!!!!!

RUMORED NVIDIA GeForce RTX 40 Series Specs
	RTX 4090	RTX 4080	RTX 4070
Architecture	Ada (TSMC N4)	Ada (TSMC N4)	Ada (TSMC N4)
GPU	AD102-300	AD103	AD104-400
Board Number	PG137/139-SKU330	TBC	PG141-SKU341
Streaming Multiprocessors (SMs)	126	TBC	TBC
CUDA Cores	16128	TBC	TBC
Memory	24 GB G6X	16 GB G6X	12 GB G6
Memory Speed	21 Gbps	TBC	18 Gbps
TDP	450W	TBC	300W
Launch Date	July 2022	July 2022	July 2022

Source: @kopite7kimi

 

[DWolvin]

Wow, that makes me even less excited, unless they come in much cheaper than expected...

 

[rhkcommander959]

Much doubt, unless it's a paper launch!

 

[GoodBoy]

They sell those papers on Ebay. "RTX 3090 - Picture only" for like $1000 and people bid on them. Must be a collectible or something.

 

[Armenius]

https://videocardz.com/newz/nvidia-...0w-tdp-is-allegedly-twice-as-fast-as-rtx-3090

JULY 2022~!!!!!!!!!!!!!!!!

	RTX 4090	RTX 4080	RTX 4070
RUMORED NVIDIA GeForce RTX 40 Series Specs
Architecture	Ada (TSMC N4)	Ada (TSMC N4)	Ada (TSMC N4)
GPU	AD102-300	AD103	AD104-400
Board Number	PG137/139-SKU330	TBC	PG141-SKU341
Streaming Multiprocessors (SMs)	126	TBC	TBC
CUDA Cores	16128	TBC	TBC
Memory	24 GB G6X	16 GB G6X	12 GB G6
Memory Speed	21 Gbps	TBC	18 Gbps
TDP	450W	TBC	300W
Launch Date	July 2022	July 2022	July 2022

Source: @kopite7kimi

I may actually wait for the Ti version this time around. I don't feel like my 3090 is really struggling with anything at this point. We'll see what happens with the games coming out in the next year.

 

[Domingo]

My plan is to snag the 4080 and sell my 3090 to offset some of the cost. I'm a sucker for bigger/better/faster/more, but I can't really justify the 4090 unless one falls in my lap. 16GB is enough RAM for my needs this time around, which is nice.

 

[exlink]

I may actually wait for the Ti version this time around. I don't feel like my 3090 is really struggling with anything at this point. We'll see what happens with the games coming out in the next year.

My 3090 isn’t having any issues but I would definitely buy at launch if there is another large uptick in RT performance. From a rasterization perspective my 3090 is having absolutely no issue with any game I throw at it and I don’t see it having any issues in pure rasterization performance for the next 2 years.

 

[Thunderdolt]

Spec rumors bode well for rendering & simulation performance. Quite possibly a 60-70% bump from the CUDA cores alone, and that's before any tensor/RT improvements. I will probably dive in to some 4090s assuming they aren't massively crippled to protect the much higher margin enterprise/workstation cards.

 

[Blade-Runner]

https://www.techpowerup.com/294944/...-spied-retains-dual-axial-flow-through-design

 

[Blade-Runner]

 

[Blade-Runner]

 

[polonyc2]

I'll stick to my every other generation upgrade rule...I have a 3080 and at 1440p 144hz G-Sync it gives me everything I need...doesn't the 4000 series require a new ATX spec power supply as well?...I'll give all this new hardware a generation to mature

 

[vegeta535]

I perfectly happy with my 3080ti. I will wait til there is a significant architecture improvement. Not a brute force approach to performance.

 

[TaintedSquirrel]

I perfectly happy with my 3080ti. I will wait til there is a significant architecture improvement. Not a brute force approach to performance.

Hoping for both this time. AMD is putting the pressure on again.

 

[Lakados]

Hoping for both this time. AMD is putting the pressure on again.

I have no doubt AMD will deliver an awesome card. I doubt their ability to supply it to any large degree.

 

[Thatguybil]

As somebody who is running a system with a pair of 1600w PSU's to run a stack of GPUs yeah that is a thing, they have to crunch a shitload of data and you can only improve efficiency so much before you just have to throw more juice at it.

Must be on a 240 volt 20 or 30 amp circuit….

 

[Lakados]

Must be on a 240 volt 20 or 30 amp circuit….

240/30 on a separate panel from the AC and the rest of the building.

 

[NattyKathy]

Those memory specs on the 4070 tho... finally giving the x70 more than 8GB is nice but 12GB means either 384-bit bus which seems astronomically unlikely for a -104 card, or 256-bit with 660Ti/970 type malarkey which I assume Nvidia wouldn't do after that whole PR fiasco, or more likely a 192-bit bus which... ouch. I know these cards are (allegedly) going to have much larger caches but as we saw with RDNA2 that can only do so much at higher resolutions.

Something tells me NV still thinks people with 4K monitors should have to shell out for the x80, a position that's becoming increasingly untenable without restricting memory and RT cores. Imagine, an AD104 card with >=3090 shader perf shackled to a 192-bit bus and the same RT Core count as GA104 and TU104 before it... I hope that's not the case.

 

[funkydmunky]

I’m not sure it ever was just an Ampere refresh, it takes years to design a GPU obviously they make iterative changes as they go. Especially if you are completely changing up manufacturing and processing. But seeing a few leaks and tossing your designs wouldn’t be feasible.

Multiple design teams working closely. Power houses like Nvidia and Intel have dragged down product lines before because, no competition. Doesn't mean they don't have their uber product in the bag. You never want to launch too good of a tech when there is no market need. Instead you will go efficiency and embarrass your competition that way.
The final release will probably end up being a combo of both to hit all market segments. What is clear now is an Ampere refresh won't cut it and they are going full on with the big team.

 

[LukeTbk]

or more likely a 192-bit bus which... ouch

It is the main rumors it seem like, that would be around a 432 GB/s bandwith, which is not much slower than the 448 GB/s of the 3070, but nothing close to the older 3080 760GB/s

https://twitter.com/kopite7kimi/sta...s1_&ref_url=https://tech4gamers.com/rtx-4070/
AD102 12*6 384bit ~600mm²
AD103 7*6 256bit
AD104 5*6 192bit
AD106 3*6 128bit
AD107 3*4 128bit

Like you said 12 time more L2 cache too, which could make VRAM access of the 4070 in pratice faster than on a 3070 despite lower bandwith

 

[Flogger23m]

Something tells me NV still thinks people with 4K monitors should have to shell out for the x80, a position that's becoming increasingly untenable without restricting memory and RT cores.

Well in general, **70s just aren't capable of good frames at 4K anyways. **60 = 1920x1080, **70 = 2560x1440, **80 = 3440x1440, **80ti/90 = 4K.

Of course it varies a bit. A **70 can often deliver good frame rates at ultra wide, but sometimes it may be hovering around the 50 frame rate range at 2560x1440. Depends on the game.

Now I do think 8GB is probably going to be too little going forward. So hopefully it does have 12GB, and it isn't like the 970.

 

[xDiVolatilX]

Anything over 250W turns my room into an oven.
I wonder how much performance is lost by undervolting a 600W card down to 250W? Guess we'll find out later this year.

That's why I installed a 10" industrial exhaust fan in the wall in my PC room with a voltage controller to slow it down to make it much more quiet. my 3080Ti goes over 400 watts easy unless I undervolt it.

 

[Thunderdolt]

Multiple design teams working closely. Power houses like Nvidia and Intel have dragged down product lines before because, no competition.

Nvidia must always compete with itself. Something Turing showed us is that if the performance jump over the previous generation isn't enough, then the installed base won't upgrade. If the installed base doesn't upgrade, then Nvidia gets nothing and the stock price suffers.

AMD doesn't really factor in to this because they're a niche player with under 20% marketshare of new sales. They would have to produce cards that offer both a big enough jump over Nvidia's previous gen to get people interested in upgrading at all AND a big enough jump over Nvidia's current/next gen for those buyers to jump ship. That is a tremendous task. If the 4000-series somehow got delayed until 2024 and RDNA 3 launched tomorrow but turned out to be 20% faster than the 30-series, it wouldn't register on Nvidia's radar because 20% isn't enough to motivate purchases. This would have to happen over multiple consecutive generations to really move the needle. Nvidia is sort of in the same position when it comes to getting AMD owners to switch: Nvidia would have to launch multiple consecutive generations of cards that are 40+% faster than AMD for a meaningful number of owners to switch. To be a leader for that long and by that margin would be a monumental achievement for any company.

 

[TaintedSquirrel]

Computex tomorrow at 11 pm est.

 

[Krenum]

This is one reason I haven't been able to convince myself to buy a GPU lately. Might have to take off work for launch day.

There won't be any available on launch day.

 

[Thunderdolt]

The VRAM is expected to be the same as the 3090, where it pulls an estimated 60-100W. I'd expect the FE to be 400W while something like a Strix 4090 to be 450W.

You may want to bookmark this post so you can come back to it and mock me in a few months.

Looks like I came in 10% low on my estimate for the 4090 FE. Not bad.

MLID was over 30% high and promised the fanbois that he had gotten that number from a legitimate source. LMAO. What an assclown.

 

[FrgMstr]

Looks like I came in 10% low on my estimate for the 4090 FE. Not bad.

MLID was over 30% high and promised the fanbois that he had gotten that number from a legitimate source. LMAO. What an assclown.

The first 4090 build sheet submitted months ago in China was for a 600W card. That card does still exist in the internal NVIDIA ecosystem. I can list a bunch of reasons for not releasing that now.

 

[Thunderdolt]

The first 4090 build sheet submitted months ago in China was for a 600W card. That card does still exist in the internal NVIDIA ecosystem. I can list a bunch of reasons for not releasing that now.

A card with hardware capable of supporting 600W is not the same as a card running at 600W. That's a weird reframe in order to fit a specific bias.

An Nvidia official stated that they have at least one of the production cards running at over 3Ghz in their lab. That would fit with the production cards having hardware capable of supporting 600W.

MLID simply found a nugget of info which allowed his bias to turn it into a statement that fit the bias of his loyal band of frothy fanbois. Such an obvious trap, but people keep falling for it.

 

[FrgMstr]

A card with hardware capable of supporting 600W is not the same as a card running at 600W. That's a weird reframe in order to fit a specific bias.

An Nvidia official stated that they have at least one of the production cards running at over 3Ghz in their lab. That would fit with the production cards having hardware capable of supporting 600W.

MLID simply found a nugget of info which allowed his bias to turn it into a statement that fit the bias of his loyal band of frothy fanbois. Such an obvious trap, but people keep falling for it.

Okie dokie.

 

[next-Jin]

https://videocardz.com/newz/nvidia-...0w-tdp-is-allegedly-twice-as-fast-as-rtx-3090

JULY 2022~!!!!!!!!!!!!!!!!

	RTX 4090	RTX 4080	RTX 4070
RUMORED NVIDIA GeForce RTX 40 Series Specs
Architecture	Ada (TSMC N4)	Ada (TSMC N4)	Ada (TSMC N4)
GPU	AD102-300	AD103	AD104-400
Board Number	PG137/139-SKU330	TBC	PG141-SKU341
Streaming Multiprocessors (SMs)	126	TBC	TBC
CUDA Cores	16128	TBC	TBC
Memory	24 GB G6X	16 GB G6X	12 GB G6
Memory Speed	21 Gbps	TBC	18 Gbps
TDP	450W	TBC	300W
Launch Date	July 2022	July 2022	July 2022

Source: @kopite7kimi

Look at that 4070 12GB card, pricing should be around 500 USD for 3090 Ti performance! /s

 

[Twisted Kidney]

Holy shit those dies are massive.

 

[Armenius]

Holy shit those dies are massive.

The 4090 is 616 mm². The 3090/3080 Ti/3080 is 628 mm².

The 4070 4080 12GB is 300 mm², while the 3070/3070 Ti is 392 mm².

 

[Twisted Kidney]

The 4090 is 616 mm². The 3090/3080 Ti/3080 is 628 mm².

The 4070 4080 12GB is 300 mm², while the 3070/3070 Ti is 392 mm².

Yeah, I said the same thing last generation as well. They're still huge.

 

[Lakados]

Yeah, I said the same thing last generation as well. They're still huge.

And dense, The Samsung 8N process is their name for what is closer to a 10nm process in reality and is capable of 44.56M transistors per square mm, TSMC 4N is an advancement on their 5nm process which is 196.6M transistors, and yeah I know there is more to it than just transistor density but that should alone show the vast difference in product density. TSMC hasn't made public the density numbers for the 4N process but they say it is an improvement in both density, power, and thermals over the 5, so yeah. These chips are big and crammed full of all the goodies.

It's going to be fun to see what the AMD chips look like now that they are both on the same process nodes. Lets us see a more Apples to Apples sort of thing so fun time, but I think my wallet is going to keep me out this round for sure.

I mean last-gen Samsung 8N vs TSMC 7, TSMC was almost 3x the density and vastly better power and thermals than the Samsung node.

 

[Brackle]

And dense, The Samsung 8N process is their name for what is closer to a 10nm process in reality and is capable of 44.56M transistors per square mm, TSMC 4N is an advancement on their 5nm process which is 196.6M transistors, and yeah I know there is more to it than just transistor density but that should alone show the vast difference in product density. TSMC hasn't made public the density numbers for the 4N process but they say it is an improvement in both density, power, and thermals over the 5, so yeah. These chips are big and crammed full of all the goodies.

It's going to be fun to see what the AMD chips look like now that they are both on the same process nodes. Lets us see a more Apples to Apples sort of thing so fun time, but I think my wallet is going to keep me out this round for sure.

I mean last-gen Samsung 8N vs TSMC 7, TSMC was almost 3x the density and vastly better power and thermals than the Samsung node.

AMD chips will look vastly different because of the chiplet design. This is why AMD will be able to make more of a profit from GPU's this generation than Nvidia.

It's the same thing with their CPU's....all about efficiency, and having less yield problems. chiplets helps in that regard.

 

[Lakados]

AMD chips will look vastly different because of the chiplet design. This is why AMD will be able to make more of a profit from GPU's this generation than Nvidia.

It's the same thing with their CPU's....all about efficiency, and having less yield problems. chiplets helps in that regard.

They aren't chiplet though, I mean they are but they aren't. They have 2 chips but what they have done is move the IO off the main chip but the actual GPU itself is still a single monolithic die.
This is going to give them some small power advantages and cost savings but not significant, I expect AMD this time around to be aiming at advertising their power efficiencies and cooler temperatures, I don't expect them to be stepping up to Nvidia's performance levels I will be happy AF if they do, but I don't expect it.
I am thinking they are going to go with smaller dies, the Nvidia ones are flipping massive, and instead look at pushing market share in the middle ground and really work on punishing Nvidia for their overstock on the 3000's by making them compete there instead.

 

[equinox654]

They aren't chiplet though, I mean they are but they aren't. They have 2 chips but what they have done this time around is moved the IO off the main chip but the actual GPU itself is still a single monolithic die this time around.

Does anyone know what percentage of the die was io?

 

[drutman]

$1300-$2500 just for the GPU? lol

And then look what happened. (And people still buy them)

1300/365 = $3.56 a day to use card. Hell I know people who spend $10 a day on fast food or Starbucks coffee with no complaints.

 

[Brackle]

They aren't chiplet though, I mean they are but they aren't. They have 2 chips but what they have done is move the IO off the main chip but the actual GPU itself is still a single monolithic die.
This is going to give them some small power advantages and cost savings but not significant, I expect AMD this time around to be aiming at advertising their power efficiencies and cooler temperatures, I don't expect them to be stepping up to Nvidia's performance levels I will be happy AF if they do, but I don't expect it.
I am thinking they are going to go with smaller dies, the Nvidia ones are flipping massive, and instead look at pushing market share in the middle ground and really work on punishing Nvidia for their overstock on the 3000's by making them compete there instead.

It is not monolithic. You can try to spin it anyway you can, but it is still a chiplet design. This is how AMD is able to have less defects at TSMC and make more money. Yes there will still be an IO die. But if you call RDNA3 monolithic, then you are saying the Ryzen CPU's are also monolithic.