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

[GodisanAtheist]

My new GPU upgrade requirements are gonna turn into Double Performance for $300 @ 300W...

 

[SmokSmog]

Pessimistic:
RTX 4090 AD102 140SM 24/48GB?
RTX 4080Ti AD102 128SM 384Bit 24GB
RTX 4080 AD103 84SM 256Bit 16GB
RTX 4070 AD104 60SM 192Bit 12GB - faster or as fast as RTX 3090
RTX 4060Ti AD104 52SM 192Bit 12GB
RTX 4060 AD106 36SM 128Bit 8GB
RTX 4050Ti AD107 24SM 128Bit 8GB



Optimistic:
RTX 4090 AD102 140SM 384Bit 24/48GB?
RTX 4080Ti AD102 128SM 384Bit 24GB
RTX 4080 AD102 114SM 320Bit 20GB
RTX 4070Ti AD103 84SM 256Bit 16GB
RTX 4070 AD103 80SM 256Bit 16GB
RTX 4060Ti AD104 60SM 192Bit 12GB faster or as fast as RTX 3090
RTX 4060 AD104 52SM 192Bit 12GB
RTX 4050Ti AD106 36SM 128Bit 8GB
RTX 4050 AD107 24SM 128Bit 4/8GB

It all depends on the RDNA3 performance, if it's slow then pessimistic variant, if fast then optimistic variant.

 

[Timmah!]

If Lovelace has 2x full FP32 units like Hopper seems to have, do you reckon the INT32 units will still have the partial FP32 capability like on Ampere? Basically having potentially 3xFP at times (compared to pre Ampere architectures with their 1:1 FP32 to integer ratio?)

 

[gdansk]

I wonder if it is a blend of the two. Nvidia may make a gap to drive people upward from the xx70. A 4070 with only 60 SM 12GB and a 4080 with 102 or 114 SM and 20GB. If the 4080 is $1000 and the 4070 $600 it would push people toward the much higher ASP 4080. Nvidia can always fill the gap later with a Ti version if the competition aims there. And maybe I'm too optimistic on pricing lol

 

[igor_kavinski]

Nvidia got screwed when it decided to go with Samsung. It's almost a pattern with Intel/Nvidia. Everything seems to be going great for them and then a misstep dethrones them. AMD is the benefactor in both cases. It's like the Universe is keeping AMD alive.

 

[CakeMonster]

It all depends on the RDNA3 performance, if it's slow then pessimistic variant, if fast then optimistic variant.

Isn't timetable a bit too close that they would be able to revamp the models and bins based on AMD's lineup?

 

[CakeMonster]

Nvidia got screwed when it decided to go with Samsung. It's almost a pattern with Intel/Nvidia. Everything seems to be going great for them and then a misstep dethrones them. AMD is the benefactor in both cases. It's like the Universe is keeping AMD alive.

I'm sure they would have loved to have more volume but looking as sales figures NV didn't do too bad in '20 and '21 as far as I can tell. And I'm not sure what alternative foundry would have given their more volume over competitors, TSMC 7nm would just have made everyone miserable.

 

[maddie]

Nvidia got screwed when it decided to go with Samsung. It's almost a pattern with Intel/Nvidia. Everything seems to be going great for them and then a misstep dethrones them. AMD is the benefactor in both cases. It's like the Universe is keeping AMD alive.

Dethrone? Do I need a revised dictionary?

 

[igor_kavinski]

Dethrone? Do I need a revised dictionary?

Performance / Watt:



In my dictionary, that's a dethrone. I know which one I would buy if I had cash to burn. More VRAM and lower power consumption? Yes, please. I would gladly pay $200 more for the 6900 XT. Don't want to pay top dollar for only 10GB RAM that will hurt later if I want to try some heavy duty texture mods.

 

[Mopetar]

Nvidia got screwed when it decided to go with Samsung. It's almost a pattern with Intel/Nvidia. Everything seems to be going great for them and then a misstep dethrones them. AMD is the benefactor in both cases. It's like the Universe is keeping AMD alive.

Nvidia really only has themselves to blame. They didn't want to pay what TSMC was asking and opted to go with Samsung instead. Frankly I don't think Nvidia stumbled so much as AMD got their act together and put together a really compelling product. Both Zen (particularly Zen 2 onwards) and RDNA 2 are just solid designs.

Hopefully the renewed competition has lit a fire under team green because prior the the pandemic and the mining boom, a 3080 (which traditionally would be more like a Ti card since it was released on the 102 die as opposed to the 104 die like the xx80 cards have been for several generations) for $700 was some of the best value Nvidia has offered in quite a while.

The 4000 series of cards may wind up being an even better value, particularly if the 4060 Ti winds up offering a similar value as the 3060 Ti did. If we get something that can offer 3080 Ti performance in 1440p at $400 that's going to be a sweet spot for many gamers putting together a new build.

 

[igor_kavinski]

If we get something that can offer 3080 Ti performance in 1440p at $400 that's going to be a sweet spot for many gamers putting together a new build.

I hope the thermals/power consumption are not similar to 3080 Ti.

 

[jpiniero]

Going with Samsung was genius IMO.

The 4000 series of cards may wind up being an even better value, particularly if the 4060 Ti winds up offering a similar value as the 3060 Ti did. If we get something that can offer 3080 Ti performance in 1440p at $400 that's going to be a sweet spot for many gamers putting together a new build.

Full AD106 extreme best case is maybe close to 3080 and $500 min.

 

[Glo.]

Full AD106 extreme best case is maybe close to 3080 and $500 min.

2304 ALUs/128 bit bus, as fast as 4352 ALUs/320 bit bus.

Perf/ALU would have to basically double, with memory bandwidth efficiency would also have to double, at least.

Im not saying its impossible.

But historically this has NEVER happened.

Tame your expectations, people. Especially when 144 SM AD102 is supposed to bring 80% performance increase over 84 SM GA102.

Nvidia REALLY has this irrational perception...

P.S. 2304 ALUs/128 bit bus with 2.3 GHz clock speeds - we are looking at 3060 Ti-3070 performance, if 80% for AD102 is correct number, with 80% more SMs over GA102.

 

[jpiniero]

2304 ALUs/128 bit bus, as fast as 4352 ALUs/320 bit bus.

Perf/ALU would have to basically double, with memory bandwidth efficiency would also have to double, at least.

Performance gap between the 38 SM 3060 Ti and the 3080 is about 45%, from Compubase's review of the 3090 Ti. Obviously the effective clock would need to be a lot higher than 2.3 and the L2 increase would have to support the memory bandwidth, along with using faster GDDR.

 

[Glo.]

Performance gap between the 38 SM 3060 Ti and the 3080 is about 45%, from Compubase's review of the 3090 Ti. Obviously the effective clock would need to be a lot higher than 2.3 and the L2 increase would have to support the memory bandwidth, along with using faster GDDR.

Ahem.

3080 has 68 SMs. So its 80% more SMs for 45% more performance. Memory bus width is 256 vs 320 bit bus. Here is the culprit.

2304 ALUs/128 bit bus. You get lower memory bandwidth, but higher L2 cache capacity.

There is literally zero chance that AD106 will achieve 3080 performance. Zero. Specs gap is too wide, unless there is something that we don't know, like: Nvidia doubled the ALU count per SM.

This is the only possibility that 36SM/128 bit bus, 32 MB L2 cache GPU will even touch 3080 performance.

However, again. AD102 is supposed to be 80% faster than 3090, while having 80% more SMs. So pretty much we can exclude this possibility.

 

[Mopetar]

It's really hard to say without actually nailing down exactly what resolution we're considering and other speculations such as memory speed. If the clock speeds on both core and memory are high enough it could get close to a 3080 in 1080p. The 3060 Ti was GA-104 so really we should be talking about the dies because Nvidia can shuffle that around. If the 4060 Ti (or whatever they call the cut down AD-104 die) still comes in at that same price and part of the product stack, I wouldn't be surprised if they have a ~$400 card that beats the 3080 at lower resolutions.

 

[SmokSmog]

Imho RTX 4070 and RTX 4060Ti should get at least AD104
AD103 is the new GA104.

Look at the SM ratio between GA102 and GA104 then AD102 and AD103

Ada Lovelace
AD102 =144SM 100%
AD103=84SM 58,3%
AD104=60SM 41,66%
Now Ampere
GA102=84SM 100%
GA103= 60SM 71,42%
GA104=48SM 57,14%<------ RTX 3070Ti
GA106= 30SM 35,71%<------RTX 3060

AD104 looks like it could be in RTX 4060.

RTX 4070Ti AD103 84SM 256Bit 16GB
RTX 4070 AD103 80SM 256Bit 16GB
RTX 4060Ti AD104 60SM 192Bit 12GB faster or as fast as RTX 3090
RTX 4060 AD104 52SM 192Bit 12GB
RTX 4050Ti AD106 36SM 128Bit 8GB
RTX 4050 AD107 24SM 128Bit 4/8GB

RTX 4060 getting AD106 with 128Bit bus and 8GB vram would have been a downgrade from 12GB RTX 3060
AD106 and 107 look like really low,low end chips for RTX 4050Ti/non ti.

Remember, Nvidia is preparing 450-600W AD102 GPU, that mean they must be worried about RDNA3 performance.

 

[jpiniero]

Allegedly AD103 is intended to be mobile only.

 

[GodisanAtheist]

Allegedly AD103 is intended to be mobile only.

-That's a MASSIVE drop from the AD102 to the AD104. NV could fit a solid 2 chips between those two parts.

Is NV planning to just harvest everything inbetween?

 

[jpiniero]

-That's a MASSIVE drop from the AD102 to the AD104. NV could fit a solid 2 chips between those two parts.

I imagine they would want to get every last chip possible on AD102. But yields are said to be so good on N5 that they might not have to cut that far.

You're right it does cause a big gap. But that might be needed to help justify the prices, if we are talking $1500 for the lowest AD102.

 

[GodisanAtheist]

I imagine they would want to get every last chip possible on AD102. But yields are said to be so good on N5 that they might not have to cut that far.

You're right it does cause a big gap. But that might be needed to help justify the prices, if we are talking $1500 for the lowest AD102.

- Yeah but $1500 for the bottom bin AD102 (assuming bottom bin 102 performs ~3090 + 30%) to say $800 on the top bin AD104 (assuming AD104 performs ~3080 territory) is still a helluva price gap.

 

[Aapje]

Allegedly AD103 is intended to be mobile only.

That makes no sense. Why wouldn't they make cards with it?

 

[jpiniero]

That makes no sense. Why wouldn't they make cards with it?

Don't need to?

- Yeah but $1500 for the bottom bin AD102 (assuming bottom bin 102 performs ~3090 + 30%) to say $800 on the top bin AD104 (assuming AD104 performs ~3080 territory) is still a helluva price gap.

Full AD104 should def be faster than 3080 - 3080 is 68 SMs so only 13% more.

 

[Glo.]

Nvidia Ada Lovelace Leaked Specifications, Die Sizes, Architecture, Cost, And Performance Analysis

In depth breakdown of Nvidia's Lovelace including architecture, die size, cost, and performance.

semianalysis.substack.com

Great analysis of AD Lovelace specs by Dylan and Locuza.

Edit:

Also video on the topic.

 

[jpiniero]

Nvidia Ada Lovelace Leaked Specifications, Die Sizes, Architecture, Cost, And Performance Analysis

In depth breakdown of Nvidia's Lovelace including architecture, die size, cost, and performance.

semianalysis.substack.com

Great analysis of AD Lovelace specs by Dylan and Locuza.

I believe it does support GDDR7 too. Just won't be available in time for launch. The mobile parts may end up using it depending on their timing. Also I have a feeling each die is bigger than they are forecasting. That's just a hunch though.

As for the clock speeds, I feel like 2.2-2.3 is way too low. nVidia is getting 1.9-2 on the crappy SS8. I feel like fill rate is a bit of a bottleneck, and while nVidia could up the ROPs/TMUs per SM, I somehow doubt that will happen. Hell even Intel is getting more than that.