ada lovelace NVIDIA Project Beyond GTC Keynote with CEO Jensen Huang: RTX 4090 + RTX 4080 Revealed

[Sauron]

17 hours ago, starsmine said:

Is it wrong that both the 2.4 and 5.0 mustang are called mustangs?
I mean one has a 2.4L 4 banger, the other has a 5L v8?

By that logic why not call the entire lineup RTX 4080? The last two digits in the number plus the -ti label are supposed to subdivide the RTX 40 series (the equivalent to "Mustang" in this example) into performance tiers, same as most cars have different identification labels next to the series name depending on accessories and motor.

[ewitte]

22 minutes ago, starsmine said:

are we not choosing between 22 and 98 fps here?

Also note, those latencies are not 1/22 or 1/98, they are using whole pipeline latency.

Hopefully from both camps we can get at least 4k100-120 with normal DLSS/FSR2 and current levels of ray tracing so it is not that big of a deal.  There is no point paying 60% more on a product when I'm capped at 4k120 and anything over 80-90 fps is acceptable to me.

[ewitte]

1 minute ago, Sauron said:

By that logic why not call the entire lineup RTX 4080? The last two digits in the number plus the -ti label are supposed to subdivide the RTX 40 series (the equivalent to "Mustang" in this example) into performance tiers, same as most cars have different identification labels next to the series name depending on accessories and motor.

Looks like some of the comments I saw in one of the youtube videos.

 

Next comes 4080 10GB, 4080 8GB, 4080 6GB,, 4080 4GB, etc.  Lets try to make it as complicated to figure out as USB.

[Mister Woof]

3 minutes ago, Sauron said:

By that logic why not call the entire lineup RTX 4080? The last two digits in the number plus the -ti label are supposed to subdivide the RTX 40 series (the equivalent to "Mustang" in this example) into performance tiers, same as most cars have different identification labels next to the series name depending on accessories and motor.

TFW The Dark Lord himself agrees with u

[xg32]

18 minutes ago, ewitte said:

Hopefully from both camps we can get at least 4k100-120 with normal DLSS/FSR2 and current levels of ray tracing so it is not that big of a deal.  There is no point paying 60% more on a product when I'm capped at 4k120 and anything over 80 fps is acceptable to me.

i was definitely gonna get a 4090 til the evga thing, then i got to try out farcry 6 with free ubi+ (boring game, couldn't finish) on the 3080ti, looks like im hovering around 90-100fps high, and since the game doesn't support dlss and fsr 1.0 is just blurry i ran it on native, to date the only game i played with RTX on is metro exodus.

 

Then there's the new psu requirements and melting psu wires (even if they are harsh conditions im not buying them til a sure fix), if amd can come out with a 350w 7900xt that's within 5-10% (rumor is 3%) of the 4090 in rast...i'm just saying...

[starsmine]

17 minutes ago, Sauron said:

By that logic why not call the entire lineup RTX 4080? The last two digits in the number plus the -ti label are supposed to subdivide the RTX 40 series (the equivalent to "Mustang" in this example) into performance tiers, same as most cars have different identification labels next to the series name depending on accessories and motor.

you have the identification labels right there.

Its 4080 12GB and 4080 16GB
You have your suffix right there that subdivides the 4080s. 

[Sauron]

Just now, starsmine said:

you have the identification labels right there.

Its 4080 12GB and 4080 16GB

Yeah but 12 vs 16gb are not the only, nor the most relevant, difference between the two. The 4090 also has more vram than these models, so why is it not called the 4080 24GB? The number/ti name is conventionally meant to reflect the rendering/compute performance, the GB number (which is always next to all of these names by the way) is only typically used to indicate vram capacity. Using different naming schemes inconsistently is confusing and misleading.

[ewitte]

4 minutes ago, xg32 said:

i was definitely gonna get a 4090 til the evga thing, then i got to try out farcry 6 with free ubi+ (boring game, couldn't finish) on the 3080ti, looks like im hovering around 90-100fps high, and since the game doesn't support dlss and fsr 1.0 is just blurry i ran it on native, to date the only game i played with RTX on is metro exodus.

 

Then there's the new psu requirements and melting psu wires (even if they are harsh conditions im not buying them til a sure fix), if amd can come out with a 350w 7900xt that's within 5-10% (rumor is 3%) of the 4090 in rast...i'm just saying...

I upgraded to 1000W a few months back and its one that can take transients up to about 1500-1600W.  I really would like to see what the 7900xt brings to the table.   If it gets within 10% of the 4090 but costs the same or less than the 4080 16GB it is a win.  I'd expect raytracing isn't as good but hear it is probably still double a 3090ti which is sufficient.

[Sir Beregond]

2 minutes ago, Sauron said:

Yeah but 12 vs 16gb are not the only, nor the most relevant, difference between the two. The 4090 also has more vram than these models, so why is it not called the 4080 24GB? The number/ti name is conventionally meant to reflect the rendering/compute performance, the GB number (which is always next to all of these names by the way) is only typically used to indicate vram capacity. Using different naming schemes inconsistently is confusing and misleading.

Well let's be real, the 4080 12GB should have been a mid-range branded card - most likely the 4070, but I still call that 192-bit bus 60-class. 

 

If I had to guess, we'll get a 4080 Ti next year that is 20GB and 320-bit bus based on a cut down AD102.

[Sauron]

Just now, Sir Beregond said:

If I had to guess, we'll get a 4080 Ti next year that is 20GB and 320-bit bus based on a cut down AD102.

Likely considering how much bigger then 4090 is

[IPD]

I'm not paying over $1000 for an xx80 ti in the current gen.  Ever.  nVidia can get fucked. 

[Quackers101]

1 hour ago, ZetZet said:

But DLSS 2 decreased latency, DLSS 3 introduces it. If you get to choose between 60 and 80-90 fps, but 60 has less latency wouldn't you pick 60 most times? I know from my own experience almost all games with VSync are unplayable to me.

intel did a talk that shows a lot of this, of course this is for ARC but some of the idea is more or less the same.

Not sure what the brands do about the similar pipeline they was going for.

Also the video can explain some of the budget you have, also what gets "recycled" for other frames. to some similar systems that nvidia showed.

2 hours ago, Sauron said:

UP TO 4X PERFORMANCE

just combine 2 3080's and RIP lower end... its all consoles now.

[starsmine]

29 minutes ago, HenrySalayne said:

N4 has 125.5 MT/mm², N7 has 89 MT/mm², resulting in "only" 860 mm².

Yip, that's accurate with 45 MT/mm² like the GA102.

Isn't TSMC N5 171.3 MTr/mm^2 with n4 being close to 181?
and TSMC n7 91.2 MTr/mm^²
and samsung n8 at 61?

I just took the ratio between the two because that number is not telling the whole truth anyways, that density is best case, logic and cache have two different densities, and some parts of the chip just dont shrink at all like wires between nodes.

Now that said.
https://fuse.wikichip.org/news/6439/tsmc-extends-its-5nm-family-with-a-new-enhanced-performance-n4p-node/
for n4 numbers, n4p is not being used here, but I wonder if they could utilize it for a mid range refresh for manufacturing costs late next year.
Save a whole 50 bucks per card or something idk. 
Where I got 5nm number
https://en.wikichip.org/wiki/5_nm_lithography_process


for onlookers, dont forget, pi is 3, e is 3, sqrt of gravity is 3.  

[porina]

3 hours ago, ZetZet said:

And you use DLSS 3 which interpolates frames by actually delaying showing what frame comes next. Gamers love latency!

While we need better description my understanding is it wouldn't add latency. It is not interpolating, but extrapolating.

 

e.g. say we have traditionally rendered frames 1 and 2. DLSS 3 would take frames 1 and 2, and insert frame 2.5 before frame 3. It is not back-creating frame 1.5 which would necessitate delating showing of frame 2.

 

 

 

Also for @HenrySalayne and @starsmine saw following and wondered if it might be of interest to your discussion. 

Source:

 

 

[CTR640]

2 hours ago, starsmine said:

you have the identification labels right there.

Its 4080 12GB and 4080 16GB
You have your suffix right there that subdivides the 4080s. 

You forgot about the amount of cuda cores and other specs? Maybe the box and ID labels says the 4080 16GB and 4080 12GB but definitly not the differences between the cuda cores etc etc. Come on, give me a break. Stop licking Jensens arse...

 

	Nvidia GeForce RTX 4080 16GB	Nvidia GeForce RTX 4080 12GB
CUDA cores	9,728	7,680
Base clock	2,210MHz	2,310MHz
Maximum clock	2,510MHz	2,610MHz
Memory size	16GB GDDR6X	12GB GDDR6X
Memory bus	256-bit	192-bit
TDP	320 watts	285 watts

[StDragon]

4 hours ago, Arika S said:

that's more than i have ever spent on an entire PC

Nvidia is milking it, and they know it too. Q2 2023 should be where it gets real interesting. I expect some major price slashing by then.

[starsmine]

15 minutes ago, CTR640 said:

You forgot about the amount of cuda cores and other specs? Maybe the box and ID labels says the 4080 16GB and 4080 12GB but definitly not the differences between the cuda cores etc etc. Come on, give me a break. Stop licking Jensens arse...

 

	Nvidia GeForce RTX 4080 16GB	Nvidia GeForce RTX 4080 12GB
CUDA cores	9,728	7,680
Base clock	2,210MHz	2,310MHz
Maximum clock	2,510MHz	2,610MHz
Memory size	16GB GDDR6X	12GB GDDR6X
Memory bus	256-bit	192-bit
TDP	320 watts	285 watts

3080 12GB, and 3080
2060 and 2060 12GB and 2060 KO (which only EVGA gave it the KO name to distinguish it, otherwise we would have never known)
1060 3GB and 1060
1050 2GB and 1050 3GB
RX560 and RX560 <-- now that one is bad
RX 5700xt and  RX 5700 XT 50th Anniversary Edition
the four varients of GTX 460 that are all called 460 

23 minutes ago, porina said:

While we need better description my understanding is it wouldn't add latency. It is not interpolating, but extrapolating.

 

e.g. say we have traditionally rendered frames 1 and 2. DLSS 3 would take frames 1 and 2, and insert frame 2.5 before frame 3. It is not back-creating frame 1.5 which would necessitate delating showing of frame 2.

 

 

 

Also for @HenrySalayne and @starsmine saw following and wondered if it might be of interest to your discussion. 

Source:

 

 

Ah, thank you for the source there, so that density is them averaging all the logic, cache, and wires into one density number.

[Quackers101]

4 minutes ago, starsmine said:

3080 12GB, and 3080
2060 and 2060 12GB and 2060 KO (which only EVGA gave it the KO name to distinguish it, otherwise we would have never known)
1060 3GB and 1060
1050 2GB and 1050 3GB
RX560 and RX560 <-- now that one is bad
RX 5700xt and  RX 5700 XT 50th Anniversary Edition
the four varients of GTX 460 that are all called 460

a lot of that, which shouldn't happen and they know what they are doing. Very anti-consumer

[BiG StroOnZ]

Paul's Hardware Coverage

 

 

Linus Tech Tips Coverage

 

 

Hardware Unboxed Coverage

 

 

[Dracarris]

1 hour ago, starsmine said:

3080 12GB, and 3080

Why that one? they are apart from the VRAM capacity almost identical GPU-wise?

[Guest]

I'm concerned if DLSS3's optical flow stuff will make the video game equivalent of the soap opera effect, especially if the source frames are low enough.

 

But we'll see

[AnonymousGuy]

1 hour ago, Dracarris said:

Why that one? they are apart from the VRAM capacity almost identical GPU-wise?

No sireee.  The 3080 12GB is basically a 3080Ti (more vram, wider bus, more cuda).   This is a great table: https://en.wikipedia.org/wiki/GeForce_30_series#GeForce_30_(30xx)_series_for_desktops

[Dracarris]

37 minutes ago, AnonymousGuy said:

No sireee.  The 3080 12GB is basically a 3080Ti (more vram, wider bus, more cuda).   This is a great table: https://en.wikipedia.org/wiki/GeForce_30_series#GeForce_30_(30xx)_series_for_desktops

welp okay I think the difference in Cuda cores and SMs is rather small but the wider memory interface sure is a bigger difference.

[BiG StroOnZ]

NVIDIA detailed the design of its GeForce RTX 4090 Founders Edition Cooler, PCB and new Power Spike Management:

 

Quote

 

 

 

 

 

 

 

 

 

NVIDIA says it has made several changes to its design. The metal outer frame now comes with a more pronounced gunmetal tinge. The heatsink array underneath has been redesigned to improve airflow between the two fans.

 

The fan itself has been updated, too. NVIDIA says it tested as many as 50 new fan designs before choosing this one, which offers up to 20% higher airflow than the one you have in the RTX 3090 Ti. Besides the heatsink and fan, the company also redesigned the vapor-chamber plate that pulls heat from the GPU and surrounding memory chips; as well as the base plate drawing heat from the VRM components. The heatpipes follow an improved heat-distribution.

 

While the RTX 4090 operates at 450 W by default, the power delivery capability allows you to increase the power limit up to 600 W for overclocking. The card features a 23-phase VRM (20-phase GPU + 3-phase memory). NVIDIA claims that it has re-architected its VRM to offer superior transient voltage management. This is specifically to minimize the kind of spikes or excursions we've seen with previous-gen high-end graphics cards such as the RTX 3090 Ti and Radeon RX 6900 Series. These spikes often causes spontaneous shutdowns on some power supplies, even if they had a much higher wattage rating than required.

 

NVIDIA shows how current spikes get mitigated by their new VRM design, which uses a PID controller feedback loop. While the RTX 3090 bounces up and down, the RTX 4090 stays relatively stable, and just follows the general trend of the current loading pattern thanks to a 10x improvement in power management response time. It's also worth pointing out that the peak power current spike on the RTX 3090 is higher than the RTX 4090, even though the RTX 3090 is rated 100 W lower than the RTX 4090 (350 W vs 450 W).

 

https://www.techpowerup.com/299096/nvidia-details-geforce-rtx-4090-founders-edition-cooler-pcb-design-new-power-spike-management

[Alcarin]

The $900 4070 pretending to be a 4080 is especially ridiculous and hilarious, I actually laughed out loud at that. Oh and the "fancy" interpolation tech just screams extra latency to me. As per their Cyberpunk comparison slide, they make a big deal about that "58ms" latency. That's ridiculously high for that framerate, and would be very jarring to play. They would have to do some serious black magic rituals to get that sorted out.

 

Thoroughly unimpressed. When Nvidia slashes prices next year then I might look into possibly upgrading, but I'd rather just wait for the 50 series.