Discussion Ada/'Lovelace'? Next gen Nvidia gaming architecture speculation

[Karnak]

*3090Ti

But yeah, with the massive L2 increase, a +50% SM increase (going with a cut down, full AD102 which we know thanks to the NV leak has 144SM) and the rumored like >2.6GHz - with still most likely a lot more architectural improvements we don't know yet - only like +70-80% against the 3090Ti doesn't seem that much.

It's either more in games or the leaks were wrong. I'm still with the former for now.

 

[jpiniero]

I scored 14 977 in Time Spy Extreme

Intel Xeon W-3175X Processor, NVIDIA GeForce RTX 3090 Ti x 1, 49152 MB, 64-bit Windows 10}

www.3dmark.com

By comparison, the top score at 3dmark is 14600. That's a 3090 Ti overclocked to 2.8 Ghz using LN2.

 

[Karnak]

Ofc not a daily use case but still interesting if we're assuming the rumored clocks for Ada (which would be roughly similar to the LN2 3090Ti) are true. Then >19000 points for the 4090 are not that big of an improvement if you're going only by architectural changes.

Though if you're getting like 1GHz+ clocks without needing water or LN2 that's still part of the overall performance improvement.

Still interesting at least to me isolating clock speeds and looking at the architecture itself (though adding more CU/SM by itself is technically not an "architectural" thing).

 

[MrTeal]

If Kimi is right, where would that put the 4080 in TSE? The 4090 has 60% more SMs and 75% more memory bandwidth than a 4080 based on his previous leaks. TSE scales pretty linearly on Ampere with SM, so if all those numbers are right you might see the 4080 be pretty close to a 3090 Ti. That would be pretty terrible generational scaling as the 3080 was ~35% faster in TSE than a 2080 Ti and a 3070 was within a few percent of it.

 

[jpiniero]

https://twitter.com/x/status/1549051075243966464

Does say that 3+ Ghz is very doable. Maybe not at 128 SMs/450 W but doable.

Ofc not a daily use case but still interesting if we're assuming the rumored clocks for Ada (which would be roughly similar to the LN2 3090Ti) are true. Then >19000 points for the 4090 are not that big of an improvement if you're going only by architectural changes.

That would be 30% faster than that overclocked 3090 Ti with 50% more SMs.

I suppose it could be a power/clock locked ES. I am wondering about memory bandwidth too, wonder if they were expecting faster speeds to be available.

 

[biostud]

And now something showing +100% in control 4k vs 3090

Full NVIDIA "Ada" AD102 GPU reportedly twice as fast as RTX 3090 in game Control at 4K - VideoCardz.com

Alleged full AD102 GPU with 160+ FPS in Control at 4K Just a day after first performance leaks on GeForce RTX 4090, we now have rumors on the full AD102 GPU. The full AD102 GPU should not be confused with RTX 4090 non-Ti. The full GPU features as many as 18432 CUDA cores which is […]

videocardz.com

 

[jpiniero]

And now something showing +100% in control 4k vs 3090

Full NVIDIA "Ada" AD102 GPU reportedly twice as fast as RTX 3090 in game Control at 4K - VideoCardz.com

Alleged full AD102 GPU with 160+ FPS in Control at 4K Just a day after first performance leaks on GeForce RTX 4090, we now have rumors on the full AD102 GPU. The full AD102 GPU should not be confused with RTX 4090 non-Ti. The full GPU features as many as 18432 CUDA cores which is […]

videocardz.com

That's the full AD102 and also says "high power draw". So the implication is that it's 600 W... or maybe more?

 

[Saylick]

That's the full AD102 and also says "high power draw". So the implication is that it's 600 W... or maybe more?

Kopite7kimi says 800W? Not sure.

https://twitter.com/x/status/1549330721763823617

 

[Mopetar]

If people match that with a power hungry CPU, it's going to take one hell of a power supply to keep that system fed.

800W just seems outside of the realms of believability. Just think of all the connectors it would need and all of the VRMs and other components. That doesn't square with a "very small" PCB at all, even if we're speaking relative to other top end GPUs.

 

[KompuKare]

Nevermind the max power draw, what about the spikes?

One way to mitigate bursty loads might be to have a large PCB with plenty of smoothing capacitors to lessen the load on the PSU. Looks like the Nvidia reference design isn't going down that route.

Alder Lake is power hungry too and Raptor Lake looks worse. Even Zen4 looks like it will up is TDP.

So what happens if playing a game which is well-threaded but CPU bursty and GPU bursty, will a nominal 800W requirement actually require a 2000W power supply?

Are any shares in PSU manufacturers publicly traded?

 

[Saylick]

If people match that with a power hungry CPU, it's going to take one hell of a power supply to keep that system fed.

800W just seems outside of the realms of believability. Just think of all the connectors it would need and all of the VRMs and other components. That doesn't square with a "very small" PCB at all, even if we're speaking relative to other top end GPUs.

Yeah, I'd likely want a 2kW PSU just to be safe, now that I know transients can easily double the GPU power consumption. If you can afford the >$2000 full AD102, you can afford the 2000W PSU.

 

[jpiniero]

This was brought up before... 1500 is the practical max.

 

[Saylick]

This was brought up before... 1500 is the practical max.

PSU SLI here we come!

Buy a 1500W PSU just for the GPU, then another 500W PSU for the rest of the system.

 

[DeathReborn]

This was brought up before... 1500 is the practical max.

3500W is the practical max here with our juicy 240V 15A (most still wired for 230V 13A so 2900W).

 

[Tup3x]

300W is about the max I would consider acceptable for to end card. Even 250W card dumps too much heat for my taste.

 

[amenx]

Wonder if we see some WC models from AIBs.

 

[spaceman]

Let’s see
RTX 4050 =3060
RTX 4060 = 3070
We can fill in the blanks
It’s entirely predictable at this point might be a whopping 4% +and less power draw with some esoteric feature added. But it’s old

 

[spaceman]

But wait we’ll get the ti or some super variant to be up sold to when these models disappoint

 

[jpiniero]

Let’s see
RTX 4050 =3060
RTX 4060 = 3070
We can fill in the blanks
It’s entirely predictable at this point might be a whopping 4% +and less power draw with some esoteric feature added. But it’s old

I am not entirely convinced that you will see desktop parts for even AD106... and if there is, it's gonna be awhile. Plus both only have 25% more SMs compared to Ampere and are only 128 bit. I bet you will see the mobile versions launch at CES.

 

[xpea]

19k for Time spy extrm is Kimi sandbagging on cut down low frequency AD102. Full die (SM and all L2 cache activated) equivalent to AIB OC models will be above 25k at less than 600W (can't be more specific for obvious reasons)

 

[Leeea]

PSU SLI here we come!

Buy a 1500W PSU just for the GPU, then another 500W PSU for the rest of the system.

1500w is not the limit for the PSU.

1500w is the practical limit for your United States 120v 15 amp wall outlet!

It does not matter if you SLI your PSUs, you are just going to throw a breaker or blow a fuse.


You know that special circuit you installed to charge your car? Dual use, also is for your computer now .


I know I have linked this before, but it sure hits the nail on the head:

Spoiler: 4090 review video

That guy sure predicted the future!

 

[psolord]

^ I laugh in European Master Race fashion!

You peasant Americans, your American gpus run better in Europe!

har har har har

 

[Leeea]

^ I laugh in European Master Race fashion!

You peasant Americans, your American gpus run better in Europe!

har har har har

-eyes Psolord for a long moment-

Looks at EU electricity rates. Looks at my bill.

-cackles manically-

Good luck paying your electric bill after you sell your kidney to buy said GPU!

Spoiler: The US actually does get household 240v, and is safer to

US also gets 240v to our house. But where EU is a 2 pole 240v transformer, ours is a 3 pole 240v transformer. The center pole goes to ground, and either pole to ground then yields our standard 120v. Pole to Pole is the same 240v EU gets. The US system is safer.

We use the 240v for our stoves, air conditioners, dryers, etc. Each home usually gets at least 200 amp 240v service, although newer homes are starting to get 400 amp service.

Each home in the US gets its own transformer, which is very different from the EU. In the EU multiple homes will commonly share a secondary, here in the US that is rare.

To get 240v from our breaker box, each breaker alternates between the main transformer pole. So just tying two breakers together yields a 240v circuit. Sometimes we use a little plastic bridge for that.


This multi-grounded design offers many safety benefits for the US system. The ground acts as a physical barrier, and dangerous touch voltages are shorted to ground during faults.

If either main touches ground the breaker is thrown. The conduit and wiring is designed so the ground is exposed so any break touches ground first.

This also allows for gfci outlets that are far more sensitive, cheaper, and simpler to wire then their EU RCD equivalents. A 120v US GFCI will trigger at 5 milliamps, where a 240v EU RCD will trigger at 30 to 100 milliamps. In other words, you will get zapped between 12x-33x more with the EU system before it throws its breaker.

 

[psolord]

Well we have 40 million Poles in Europe and quite probably even more in countries other than Poland, so eat that! xD

 

[igor_kavinski]

The ARM acquisition could also have been Nvidia's ticket to low power ARM GPU cores, allowing them to ditch their power hungry cores and use multiple power sipping modified Mali cores developed exclusively for themselves. They could still do that through licensing but it wouldn't be the same as having the brilliant ARM engineers in-house to do their bidding.