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

[dr1337]

Lately for the past month or so infamous twitter leaker kopite7kimi has had a nice trickle of next generation nvidia leaks, haven't seen too many people talking about it.

https://twitter.com/kopite7kimi/status/1336900538185572357

https://twitter.com/kopite7kimi/status/1336900538185572357

https://twitter.com/i/web/status/1337739048270479361

https://twitter.com/i/web/status/1343462867811495937

And now today with further extrapolation by 3Dcenter.org and videocardz, the full AD102 die is pegged at 18,432 FP32 cores as a massive monolithic 5nm part. Though this news is really early considering Ampere launched only a few months ago, kopite has an excellent track record and also doesn't expect Lovelace for over another year anyways.

IMO this is starting to get interesting. Assuming the leaks are true it begs the question, is there something wrong with hopper? Or was hopper just too far out and Nvidia is refreshing the lineup with "ampere 2.0" in the mean time? (too much pressure from AMD?) And holy cow they're going to nearly double fp32 again?!?! Its hard not to be excited about such a monster graphics card 😆

What do you guys think? Is kopite off their rocker? Is nvidia really going to give us Big Ampere? I know its still really early for any legitimate speculation but its also so rare that we'd get such details so far out, though such a sudden shift in plans is also atypical from nvidia. Still, sounds very interesting, and its gonna be neat to see how this all pans out.

 

[guidryp]

Am I the only one that's a bit disappointed given the original (months ago) expectations? Everyone was talking about how this is basically a two node or 1.5 nodes jump because of going from "bad" Samsung 8nm to TSMC 5/4nm, and we're basically seeing 1.6X raster performance or 1.8x RT performance while using 2.7 times the amount of transistors, an insane cooler and 450 watts?

DLSS 3 is obviously a joke.

Node jumps aren't as impressive as they used to be. Dennard scaling has fallen by the wayside.

I don't put much stock in rumors, so don't get my expectations up.

 

[Mopetar]

Disappointing is a bit of a stretch. This card will let you run 4K with all the bells and whistles without needing to compromise. DLSS isn't even needed in many titles even when ray tracing is turned on in order to ensure minimum acceptable frame rates.

I'm more concerned with cards like the 4080 12 GB which won't look as impressive for their asking price.

 

[blckgrffn]

Am I the only one that's a bit disappointed given the original (months ago) expectations? Everyone was talking about how this is basically a two node or 1.5 nodes jump because of going from "bad" Samsung 8nm to TSMC 5/4nm, and we're basically seeing 1.6X raster performance or 1.8x RT performance while using 2.7 times the amount of transistors, an insane cooler and 450 watts?

DLSS 3 is obviously a joke.

Disappointed? IMO this is a full generational jump in performance that we could have had most of with Ampere if there hadn't been a process SNAFU. So we got both a refinement and a big silicon technology jump.

If anything, I think we should be prepared for underwhelming releases going forward (even a 5090). This same jump won't happen again for a while.

Which makes it even more clear to me - you are no fool if you buy this day one. It's a 1080ti type card that will likely be serviceable (laugh if you want) until 2030 if Nvidia chooses to support it that long, which I believe they will.

Or you could buy a middling card now, another one in three years, and another 3 years after that and finally, six years and three cards later you'll have something measurably faster. Congrats.

 

[Stuka87]

All this for $100 more than the original 3090, has buyers at this end of the market really well served.

You are ignoring the 3090Ti, which came out only 6 months ago, costs $2000, and is significantly slower than the 4090. People that bought these cards get nailed to the wall.

Not that I have any sympathy for somebody that spends 2K on a GPU, but it continues nVidia's "lets screw the customer to make a few more bucks" mindset. If the 3090Ti just replaced the 3090 at the same MSRP, it would not really be an issue. But nVidia priced it knowing that the 4090 was going to decimate it.

 

[linkgoron]

Disappointed? IMO this is a full generational jump in performance that we could have had most of with Ampere if there hadn't been a process SNAFU. So we got both a refinement and a big silicon technology jump.

So you're basically agreeing with me, if you're saying that Ampere would have been way closer with TSMC 7nm. For example, Navi 21 was twice as fast as Navi 10 at 4k with 2.7 times the amount of transistors, on the same node (according to TPU, 6900xt vs 5700xt).

 

[GodisanAtheist]

TPU has one of the most extensive game test suites and they show the 4090 capping out at about 45% faster than the 3090Ti @ 4K.

I suppose the increase in performance from review to review is going to depend much more heavily on the CPU in the test bench as well as the mix of older DX11 titles in the suite as well.

All told, 4090 looks like a hell of a card for the future, but sort of wasted on the games of today in anything shy of 4K+ resolution.

 

[n0x1ous]

4090 should only be tested with top of the line cpus like 5800x3d or raptor lake/zen 4. Its even bottlenecked at 4k by mortal cpus

 

[GodisanAtheist]

4090 should only be tested with top of the line cpus like 5800x3d or raptor lake/zen 4. Its even bottlenecked at 4k by mortal cpus

- The real news story is going to be the 4080's. I'm guessing the 4080 12GB, which is supposed to retail for an outrageous $800 will land somewhere in 3080 territory which is an absolute joke.

 

[TESKATLIPOKA]

RTX 4090 is a very good card. Maybe the performance increase is not as high as I originally expected, but It's still a lot better than Ampere, and they can release a full version with 20-25% more performance.
I love the much improved perf/W.
RTX4080 12GB looks pretty expensive for 3080-3080Ti level of performance.
AMD will have to try their best. Tomorrow we will know some new info about RDNA3 according to Kelper_L2.

 

[jpiniero]

Am I the only one that's a bit disappointed given the original (months ago) expectations? Everyone was talking about how this is basically a two node or 1.5 nodes jump because of going from "bad" Samsung 8nm to TSMC 5/4nm, and we're basically seeing 1.6X raster performance or 1.8x RT performance while using 2.7 times the amount of transistors, an insane cooler and 450 watts?

As I've been saying, I think it's a combination of memory bandwidth and maybe even CPU limitations at 4K.

Even then, from Compubase's review at 4K I counted 4 of the games were 80% faster than the highest Ampere card and Spider Man was 77%.

 

[lixlax]

Am I the only one that's a bit disappointed given the original (months ago) expectations? Everyone was talking about how this is basically a two node or 1.5 nodes jump because of going from "bad" Samsung 8nm to TSMC 5/4nm, and we're basically seeing 1.6X raster performance or 1.8x RT performance while using 2.7 times the amount of transistors, an insane cooler and 450 watts?

DLSS 3 is obviously a joke.

Thats exactly what I was thinking. The gains are a very decent generational leap, but if looking to the huge node jump, number of transistors, big jump in clockspeed, supposedly newer\better architecture it appears that it has some rather serious bottleneck somewhere. Perf per teraflop is much lower.

 

[Hans Gruber]

The 3090Ti has a core clock of 1900-2000mhz and the 4090 can push 3ghz. So what are we looking at here with the 4090? A huge node jump and huge GPU core clock increases. 65-70% better performance. Is the TSMC stuff 4nm? I know that Nvidia is one tick ahead of AMD RDNA3. I think the RDNA3 is on 5nm.

 

[blckgrffn]

So you're basically agreeing with me, if you're saying that Ampere would have been way closer with TSMC 7nm. For example, Navi 21 was twice as fast as Navi 10 at 4k with 2.7 times the amount of transistors, on the same node (according to TPU, 6900xt vs 5700xt).

But that didn't happen. So this is what we get. Had it been only 20-30% raster increase I think we should have freaked out. If they had dedicated themselves to raster performance, maybe we would have gotten a full 100% raster increase.

Obviously, they didn't.

This is what I thought it would be.... so not disappointed.

Thats exactly what I was thinking. The gains are a very decent generational leap, but if looking to the huge node jump, number of transistors, big jump in clockspeed, supposedly newer\better architecture it appears that it has some rather serious bottleneck somewhere. Perf per teraflop is much lower.

It looks like there is more headroom with even faster CPUs. We'll see.

 

[TESKATLIPOKA]

The 3090Ti has a core clock of 1900-2000mhz and the 4090 can push 3ghz. So what are we looking at here with the 4090? A huge node jump and huge GPU core clock increases. 65-70% better performance. Is the TSMC stuff 4nm? I know that Nvidia is one tick ahead of AMD RDNA3. I think the RDNA3 is on 5nm.

It's not 4nm, but custom 5nm from what I read on the internet.

 

[Yosar]

Thats exactly what I was thinking. The gains are a very decent generational leap, but if looking to the huge node jump, number of transistors, big jump in clockspeed, supposedly newer\better architecture it appears that it has some rather serious bottleneck somewhere. Perf per teraflop is much lower.

Yep, exactly. Disappointing. Comparing power efficiency against 3090 Ti is very wrong because 3090 Ti was already freak of the nature. Against 3090 we have 55-60% in 4K for 100 W (25%-30%) more power, 2.7 times more transistors, all those additional thousands of shaders, new cache, much better process and the result is monster card with only 55-60% better performance at best?
I fail to see almost any progress here.

 

[biostud]

Yep, exactly. Disappointing. Comparing power efficiency against 3090 Ti is very wrong because 3090 Ti was already freak of the nature. Against 3090 we have 55-60% in 4K for 100 W (25%-30%) more power, 2.7 times more transistors, all those additional thousands of shaders, new cache, much better process and the result is monster card with only 55-60% better performance at best?
I fail to see almost any progress here.

I think they are heavily invested in DLSS3 and raytracing, so developers and consumers can rejoice at the performance from a monopolistic ecosystem.

 

[igor_kavinski]

Maybe the 4090 is working in Ampere emulation mode?

 

[moonbogg]

- The real news story is going to be the 4080's. I'm guessing the 4080 12GB, which is supposed to retail for an outrageous $800 $900 ($1000+ AIB) will land somewhere in 3080 territory which is an absolute joke.

FTFY

 

[GodisanAtheist]

FTFY

Whups, I actually remembered in the wrong direction and the price still seemed bad.

 

[moonbogg]

I just want a 4060 that performs like a 3070 for the price of a 3080.

 

[aigomorla]

waiting on waterblocks for the 4090....

 

[TESKATLIPOKA]

RTX 4090TI info.
Twitter

Best dies are already saved for a 2.75Ghz boost (2.95 typ. gaming) 475W 18176 CUDA cores 96MB L2 cache 24Gb beast. Performance 10~20% above 4090

This looks very good. I don't want to know the price, but the card will be fantastic.

 

[xpea]

Yep, exactly. Disappointing. Comparing power efficiency against 3090 Ti is very wrong because 3090 Ti was already freak of the nature. Against 3090 we have 55-60% in 4K for 100 W (25%-30%) more power, 2.7 times more transistors, all those additional thousands of shaders, new cache, much better process and the result is monster card with only 55-60% better performance at best?
I fail to see almost any progress here.

Games don't use yet all features like SER, DMM and OMM. but still the performance is already excellent. Instead of looking at the half empty glass, maybe you should also look at the half full one...

RT perf (5 times faster than RDNA2)



And in rendering, it's another huge generational leap (6 times faster than RDNA2 is no slouch):

 

[exquisitechar]

Yep, exactly. Disappointing. Comparing power efficiency against 3090 Ti is very wrong because 3090 Ti was already freak of the nature. Against 3090 we have 55-60% in 4K for 100 W (25%-30%) more power, 2.7 times more transistors, all those additional thousands of shaders, new cache, much better process and the result is monster card with only 55-60% better performance at best?
I fail to see almost any progress here.

I don't know if we're just looking at different reviews, but in practice, it doesn't seem to be drawing 100 W more compared to the 3090.

 

[jpiniero]

Games don't use yet all features like SER, DMM and OMM. but still the performance is already excellent. Instead of looking at the half empty glass, maybe you should also look at the half full one...

RT perf (5 times faster than RDNA2)

With console gamers demanding 60 fps, it's safe to say that developers aren't going to go that far with RT. Mostly Raster with some Light RT is going to continue to be the norm.