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

[dr1337]

gaming flagship

Nvidia's core strategy right now is to reuse the gaming dies for workstation cards, or maybe its the other way around. Every 'pro' card they make except for the $15k HBM-clad A100 flagship shares a die with a gaming GPU. The 4090 has 24gb but right now you can buy the exact same chip with 48gb as the $5k A6000.

The 5090 will be capped at 32gb for sure. Gaming might not need the crazy bandwidth of 512bit GDDR7 but it makes the card sound extremely strong from a marketing perspective as being at ram parity with AMD right now is no good. 512 bit bus means leapfrogging them not only in capacity but raw bandwidth as well, its a no-lose situation. Then on the other side, moving up to 64gb seems mandatory for next gen AI cards and very conveniently fills the $10k gap in Nvidia's pro lineup, no extra dies or MCM packages required.

Though I do agree its likely going to be on some flavor of N3. Technically they have more room to squeeze out of N4 but I honestly don't expect 5000 series until 2H24. Surely by then the node will be in full production will all the kinks ironed out.

 

[Kepler_L2]

Very unlikely that a 5090 would use the full bus imo. Lineup is probably more like:

5090Ti/Titan GB102 - 512bit - 32GB
5090 GB102 - 448bit - 28GB
5080Ti GB103 - 384bit - 24GB
5080 GB103 - 320bit - 20GB

 

[jpiniero]

Micron's roadmap does say 3 GB GDDR7 chips is possible.

https://images.anandtech.com/doci/18981/HBM3%20Gen2%20Press%20Deck_7_25_2023_10.png

 

[CakeMonster]

Very unlikely that a 5090 would use the full bus imo. Lineup is probably more like:

5090Ti/Titan GB102 - 512bit - 32GB
5090 GB102 - 448bit - 28GB
5080Ti GB103 - 384bit - 24GB
5080 GB103 - 320bit - 20GB

Judging by this cycle, I'd be very surprised if they offered a 5080Ti unless it was priced ridiculously to make the 5090 look more appealing.

 

[Aapje]

Judging by this cycle, I'd be very surprised if they offered a 5080Ti unless it was priced ridiculously to make the 5090 look more appealing.

It all depends on the professional sales. If there is enough demand to sell all full dies as AI cards and the like, then gamers will only ever get the partially disabled dies.

But if the AI market collapses or a competitor suddenly outcompetes Nvidia, we'll probably see a 5090 Ti to sell those top tier dies.

 

[igor_kavinski]

Geforce RTX 5000 series: Every third frame is a REAL frame.

nGreedia loyalists: Thank you for this amazing tech and sky high FPS, Lord Jensen!

Jensen: You ain't seen nothing yet! Stay tuned for even more fake frames in the future. Our tech will be so advanced that you won't be able to tell the difference between fake frames and real frames. Even the raytracing will be done using a best guess AI algorithm. The future is FAKE and AWESOME!!!!

Rabid gamers: (orgasming in unison) OHHHH AHHHH OOOHHHHH!!!!!!

 

[Ranulf]

Geforce RTX 5000 series: Every third frame is a REAL frame.

It's 33.33% pure frame goodness! For only a 33.33% price increase! The more you buy, the more you save.

 

[tajoh111]

I could see 512bit being true.

The very highest end market is the one thriving the most at the moment and Nvidia generally wants the crown at all costs.

In addition, Nvidia has the professional market to branch off into which makes the investment generally worth it. They can attack the high end enthusiast market and the professional market simultaneously. Also unlike AMD, the strong presence in both segments means pretty much guaranteed sales and thus lower risk.

AMD on the hand doesn't have this guarantee so they generally hedge their bets by not going as extreme. They admit that they do not want to go after the extreme market like the RTX 4090(maybe this is a lie to cover up the lack of competitiveness of the 7900 XTX).

I also think Nvidia actually needs the bandwidth because if they don't want to dedicate anymore die space to huge cache stores, as a result, they will need more actual bandwidth. More so than AMD since AMD can use stacking manufacturing on their MCD to ensure adding cache isn't expensive.

Nvidia I am guessing will go 512bit bus but similar cache while AMD sticks with 384bit bus and more cache.

One more thing I suspect next generation is AMD will not have a clock advantage over Nvidia next generation much like this generation. Generally in the past aside from when Nvidia manufactured on SAMSUNG, LN2 clocks generally give a preview of next gen clocks. This has been consistent with Nvidia for the last 10 years aside from the Samsung years. With Ada hitting higher clocks than Navi 3 under LN2, unless Nvidia changes things with Blackwell which they might, I see the same thing happening.

 

[Ajay]

Well, if the next gen is still going to be stuck on GDDR6x/w, then wider buses are the only way to improve memory bandwidth significantly. I don't know about the other memory companies, but Samsung has mentioned that they need to get power (via lower voltage) down for client products (consumer - less of a problem for HPC/AI goliaths). If that tracks, GDDR6 may still be sticking with us for another generation, in some variation or another.

 

[SmokSmog]

Speculation

BW102
N4 800mm2
192SM
16*3GB 32Gb/s GDDR7 chips
48GB (96GB clamshell mode for Quadro)
2TB/s ( 2x of the 4090)

Maybe they would remove some L2 with that huge bandwidth.
BW102 with 2TB/s raw bandwidth and 64MB L2 ???

 

[Mopetar]

Even with a bigger bus, they'll probably still keep the larger caches. They can make a big chip and cut it a lot of ways. At some point we likely will see a Titan (or whatever they call it) with a 512-but bus, but that may only be towards the end of of the lifecycle.

Full dies are going to datacenter customers or professional users who'll pay several times over even the inflated prices that gamers pay. Even the cut down version of the die that NVidia sells out of the gate is going to beat AMD who does not seem to have much interest is matching NVidia on die size.

 

[SteinFG]

Speculation

BW102
N4 800mm2
192SM
16*3GB 32Gb/s GDDR7 chips
48GB (96GB clamshell mode for Quadro)
2TB/s ( 2x of the 4090)

Maybe they would remove some L2 with that huge bandwidth.
BW102 with 2TB/s raw bandwidth and 64MB L2 ???

Don't think 3GB GDDR is coming this soon. I mean, we just got to 2GB in last gen, even then a lot of cards (like 3060Ti and up, or A750/770 8GB) still used 1GB. NVidia probably knows the same, and increasing bus width to make 5090 more attractive.

Samsung is saying that they just made 2GB GDDR7 chips, and micron only mentions that they have plans of making 2-3GB GDDR7, without any timeframe for 3GB chips. I think we'll see a repeat of GDDR6: first introduce faster memory with same capacity (RX 5000, RTX 2000), then increase it (RX6000, RTX 3000)

 

[jpiniero]

Don't think 3GB GDDR is coming this soon. I mean, we just got to 2GB in last gen, even then a lot of cards (like 3060Ti and up, or A750/770 8GB) still used 1GB. NVidia probably knows the same, and increasing bus width to make 5090 more attractive.

Micron's roadmap mentions 3 GB chips.

Speculation

BW102
N4 800mm2
192SM
16*3GB 32Gb/s GDDR7 chips
48GB (96GB clamshell mode for Quadro)
2TB/s ( 2x of the 4090)

I thought about it... and it makes sense for this to be a *101 chip. There's been rumors that Blackwell Server is going to be chiplets... and knowing nVidia you figure it's going to be stupidly expensive and power hungry. An inbetween product might be needed, and 2 TB/sec bandwidth is nothing to scoff at.

 

[Aapje]

I thought about it... and it makes sense for this to be a *101 chip. There's been rumors that Blackwell Server is going to be chiplets... and knowing nVidia you figure it's going to be stupidly expensive and power hungry. An inbetween product might be needed, and 2 TB/sec bandwidth is nothing to scoff at.

It makes a lot more sense than having both AD106 and AD107 chips, especially since AD106 is only in the 4060 Ti on desktop. Perhaps it's worth it due to the mobile 4070 sales, but it does mostly seem like a waste of design effort.

A 512 bit 101 chip can at least be sold for mucho $$$ for AI.

 

[jpiniero]

NVIDIA introduces RTX 5000 with 32GB memory, RTX 4500 24GB and RTX 4000 20GB ADA workstation series - VideoCardz.com

NVIDIA RTX 5000, RTX 4500, RTX 4000 get ADA update Three new workstation GPUs from NVIDIA. NVIDIA RTX 4500 ADA NVIDIA’s highly anticipated update to its workstation series has finally arrived, introducing three new models to the lineup. Leading the pack is the RTX 5000, which boasts the ADA...

videocardz.com

New Quadros announced. Weirdly no AD103 model, and the cut AD102 model is cut very far (100 SMs and 256-bit memory)

 

[Saylick]

Nvidia announced DLSS3.5, which includes some optimizations for Ray Tracing in the form of Ray Reconstruction. Basically using a neural network trained to preserve high frequency content in RT effects better since the denoising step traditionally washes out that higher frequency content. Looks like an improvement in both the RT effects and higher fps.

NVIDIA announces DLSS 3.5 with Ray Reconstruction, launches this fall - VideoCardz.com

NVIDIA DLSS 3.5 announced A new upscaling technology has emerged, which is essentially an enhancement to NVIDIA’s existing DLSS3. NVIDIA DLSS 3.5, Source: NVIDIA While there haven’t been any updates regarding AMD FSR3, and XeSS going open-source, NVIDIA is determined to make headlines with the...

videocardz.com

 

[Heartbreaker]

Looks like an improvement in both the RT effects and higher fps.

Really just a image quality improvement. Sometimes delivers small performance improvement, if you replace several denoise steps with one AI step.

 

[igor_kavinski]

Wow. Will work across ALL RTX GPUs. Looks like Jensen wants to score some points with his customer base.

 

[Panino Manino]

Wow. Will work across ALL RTX GPUs. Looks like Jensen wants to score some points with his customer base.

Makes sense as it replaces their older denoise.
Which remembers that famous phrase... "the competition can't compete".

 

[HurleyBird]

This looks game changing, way more significant than framegen. Sure, the examples are probably cherry picked a bit, but what is shown is like comparing lowest settings to highest settings in terms of IQ. Hopefully AMD can come up with a comparable solution for FSR.

 

[igor_kavinski]

Hopefully AMD can come up with a comparable solution for FSR.

Don't hold your breath

 

[Heartbreaker]

This looks game changing, way more significant than framegen. Sure, the examples are probably cherry picked a bit, but what is shown is like comparing lowest settings to highest settings in terms of IQ. Hopefully AMD can come up with a comparable solution for FSR.

Unfortunately it looks like a lot of work for each game, so it will likely be a limited number of games.

 

[MoogleW]

Unfortunately it looks like a lot of work for each game, so it will likely be a limited number of games.

What makes you say that? It should be implemented similarly by developers as every DLSS tech branded before it like super resolution and frame gen(eventually with simple plugins). None of which have issues being implemented or adapted in games.

 

[Heartbreaker]

What makes you say that?

The lack of games listed with it planned.

 

[Tup3x]

The lack of games listed with it planned.

Only games that actually heavily use RT would benefit (also I don't except Remedy to update Control at this point for example). I'd assume that implementation would be rather straight forward - just replace the denoiser(s) with this one. It likely isn't just that simple but it necessarily isn't much more complicated.