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

[igor_kavinski]

also I don't except Remedy to update Control at this point for example

They might if they are bored and have nothing better to do. Not a bad strategy to want your game back in the news.

 

[Heartbreaker]

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.

But then you need two different denoise paths, one for RTX cards and one for everyone else, and that's a problem if as they indicate they apply in different places, as would be the case if they are replacing multiple denoisers, with one for RTX, but everyone else is still using the multiple steps. This is VERY messy.

I think this will be limited to partner games where they get heavy funding/help from NVidia.

 

[Dribble]

But then you need two different denoise paths, one for RTX cards and one for everyone else, and that's a problem if as they indicate they apply in different places, as would be the case if they are replacing multiple denoisers, with one for RTX, but everyone else is still using the multiple steps. This is VERY messy.

I think this will be limited to partner games where they get heavy funding/help from NVidia.

Nvidia is probably 90% of the RT market and this works with all Nvidia RTX cards not just 4 series so everyone else isn't a lot. They can manage to support DLSS and FSR, this will be similar I suspect - all games supporting it other than AMD exclusives.

 

[Heartbreaker]

Nvidia is probably 90% of the RT market and this works with all Nvidia RTX cards not just 4 series so everyone else isn't a lot. They can manage to support DLSS and FSR, this will be similar I suspect - all games supporting it other than AMD exclusives.

No one is going to release a game that doesn't work with RT on AMD/Intel cards, so they have to do two different paths.

Also this isn't really like scaling, which uses all the same inputs at the same location in the code, so those are trivial to support. This happens at different points in the code. It's messy.

 

[jpiniero]

NVIDIA reportedly working on GeForce RTX 4080 Ti with AD102 GPU for early 2024 release - VideoCardz.com

NVIDIA RTX 4080 Ti incoming Leaker claims NVIDIA is working on a faster RTX 4080. MEGAsizeGPU, a highly reliable hardware leaker, has hinted at the possibility of NVIDIA introducing a fresh addition to its graphics card lineup, named either the RTX 4080 Ti or RTX 4080 SUPER. According to the...

videocardz.com

Sketch Rumor that nVidia is working on releasing a deeper cut AD102 part for release in Early 2024. Says price would be in the 4080's MSRP range. I dunno... the 4090 is deeply cut as it is so I'm not sure it's necessary. Presumably they would also cut the 4080's MSRP to $999 at the same time.

 

[Tigerick]

NVIDIA reportedly working on GeForce RTX 4080 Ti with AD102 GPU for early 2024 release - VideoCardz.com

NVIDIA RTX 4080 Ti incoming Leaker claims NVIDIA is working on a faster RTX 4080. MEGAsizeGPU, a highly reliable hardware leaker, has hinted at the possibility of NVIDIA introducing a fresh addition to its graphics card lineup, named either the RTX 4080 Ti or RTX 4080 SUPER. According to the...

videocardz.com

Sketch Rumor that nVidia is working on releasing a deeper cut AD102 part for release in Early 2024. Says price would be in the 4080's MSRP range. I dunno... the 4090 is deeply cut as it is so I'm not sure it's necessary. Presumably they would also cut the 4080's MSRP to $999 at the same time.

It is actually making sense based on future GPU lineups. I am maintaining table of Blackwell GPU series in the frontpage; it is nearly complete, have a look to see the reasons behind. Even with the releasing of 4080Ti with 20GB 320-bit support, there is still headroom for upcoming 5080Ti cause of 24GB 384-bit memory bus support.

OTOH, there is no headroom for 4070 series even though it is being outperformed by RX7800XT. NV won't response cause if they come out 4070S with higher CUDA core counts, the performance would be too near to 4070Ti. They have to deal with it, they won't cut price officially cause NV wants to maintain the price line until 5070 comes out. AMD finally win over NV with N32 series...

 

[Heartbreaker]

OTOH, there is no headroom for 4070 series even though it is being outperformed by RX7800XT. NV won't response cause if they come out 4070S with higher CUDA core counts, the performance would be too near to 4070Ti. They have to deal with it, they won't cut price officially cause NV wants to maintain the price line until 5070 comes out. AMD finally win over NV with N32 series...

2060 Super nearly matched 2070 performance, so an update being "too near" another card is not something they are worried about.

That being said, the rumor is probably a fake. Wait for some leak from kopite7kimi.

 

[Tigerick]

2060 Super nearly matched 2070 performance, so an update being "too near" another card is not something they are worried about.

That being said, the rumor is probably a fake. Wait for some leak from kopite7kimi.

When I said no headroom, I also meant headroom for future RTX5070. If my speculation is correct, RTX5070 will maintain same memory bus with 12GB GDDR6X support. Thus, there won't be much improvement in rasterization, RT & tensor performance are different story but that one have to wait...

That's mean die size of RTX5070 would be smaller than current AD104, which explains product positioning...

I would love to be proven wrong but if NV is going to announce RTX4080Ti without RTX4070S, then you probably will see replacement, RTX5070 in 2025...

As for amount of L2 cache, I will explain in Blackwell thread later...

 

[Heartbreaker]

When I said no headroom, I also meant headroom for future RTX5070. If my speculation is correct, RTX5070 will maintain same memory bus with 12GB GDDR6X support. Thus, there won't be much improvement in rasterization, RT & tensor performance are different story but that one have to wait...

Or GDDR7 which has 50%+ more bandwidth, which is plenty of headroom.

 

[Tigerick]

Or GDDR7 which has 50%+ more bandwidth, which is plenty of headroom.

Nah, I still think NV will be using GDDR6X as standard memory choice. That explains the large amount of L2 cache used to offset bandwidth deficiency against GDDR7. I know it sounds unreality especially RDNA5 will most likely be using GDDR7, but with the rumor of 512-bit memory bus and my table, that convince me that NV will milk GDDR6X for one more generation...

 

[Heartbreaker]

Nah, I still think NV will be using GDDR6X as standard memory choice. That explains the large amount of L2 cache used to offset bandwidth deficiency against GDDR7. I know it sounds unreality especially RDNA5 will most likely be using GDDR7, but with the rumor of 512-bit memory bus and my table, that convince me that NV will milk GDDR6X for one more generation...

Rumors say 5000 series is only in 2025. By then GDDR7 will be ubiquitous.

 

[Tigerick]

Rumors say 5000 series is only in 2025. By then GDDR7 will be ubiquitous.

Not exactly in the beginning of year 2025; NV will most likely launch flagship and mobile GPU at the early of 2025. And remember NV has control over 80% of GPU market share (over 95% in mobile GPU), that means NV needs millions of memory chips to support the launch. That's why the safe bet is actually support GDDR6 and GDDR6X. AMD OTOH....

Come to think of NV's current product lineup, I think NV is planning Blackwell series with GDDR6X at the beginning stage, that explain weird lineup in this generation. All RTX4000 models except 4090 are only showing little improvement over Ampere lineup. With Blackwell, NV could finally show the full potential of generation leaps...we shall see

 

[Heartbreaker]

Not exactly in the beginning of year 2025; NV will most likely launch flagship and mobile GPU at the early of 2025. And remember NV has control over 80% of GPU market share (over 95% in mobile GPU), that means NV needs millions of memory chips to support the launch. That's why the safe bet is actually support GDDR6 and GDDR6X. AMD OTOH....

Come to think of NV's current product lineup, I think NV is planning Blackwell series with GDDR6X at the beginning stage, that explain weird lineup in this generation. All RTX4000 models except 4090 are only showing little improvement over Ampere lineup. With Blackwell, NV could finally show the full potential of generation leaps...we shall see

NVidia typically always pushes bleeding edge on available memory.

 

[Tigerick]

NVidia typically always pushes bleeding edge on available memory.

Yeah, they also care about profit margin, with higher bus of GDDR6X & large amount of L2 cache, NV is well positioned against RDNA5 with chiplet solutions. Remember, NV has to pay more for monolithic design with N3E process, thus usage of GDDR6X somehow remedy the profit margin which NV cares the most....

 

[jpiniero]

2060 Super nearly matched 2070 performance, so an update being "too near" another card is not something they are worried about.

That being said, the rumor is probably a fake. Wait for some leak from kopite7kimi.

The only reason I won't say it's totally fake is because nVidia clearly reserved the 4080 Ti name for such a product. It'd probally be 10-15% faster than the 4080.

 

[Ajay]

Yeah, they also care about profit margin, with higher bus of GDDR6X & large amount of L2 cache, NV is well positioned against RDNA5 with chiplet solutions. Remember, NV has to pay more for monolithic design with N3E process, thus usage of GDDR6X somehow remedy the profit margin which NV cares the most....

True, but how much are the extra packaging costs for AMD - also, more complex packaging leads to a higher number of rejected GPUs (on top of defective dice). Not sure how that all squares up. That said chiplet/tiles are the only way to move forward once TSMC moves to High NA EUV. Maybe NV is waiting till they have no choice.

 

[Tigerick]

True, but how much are the extra packaging costs for AMD - also, more complex packaging leads to a higher number of rejected GPUs (on top of defective dice). Not sure how that all squares up. That said chiplet/tiles are the only way to move forward once TSMC moves to High NA EUV. Maybe NV is waiting till they have no choice.

Yeah, NV and AMD are both preparing new generation of GPU lineup with different solutions:-

NV continuous using monolithic design (which costs a bomb with five different dies) with usage of GDDR6X try to offset the BOM. With single die, NV have an advantage on power and bandwidth, thus explains usage of high power GDDR6X.

AMD try to use chiplet design (savings of different die design) with GDDR7. With chiplet design, there are some trade-off like power and bandwidth deficiency; that explains usage of GDDR7.

If you ask me, who will win? I would say if both generations are hitting their full potential, then they may be tie to each other. But with NV's years of experience of monilithics design and GDDR6X support, I am more confident with NV's design. AMD has screwed up the design of N31 once and delay RDNA4 into RDNA5 with chiplet design plus new memory technology, man I am not sure AMD will pull through....plus release timings ????

 

[Aapje]

My assumption is that Blackwell was pushed to 2025 to be able to use GDDR7, at least for the high end. They will probably use N3E and then probably another N3 process for Blackwell+, like N3P or N3X. So they can they make monoliths for another two generations.

Of course, the current rumors are that they will make a separate design for AI and such, so we may see both a monolith for gaming and a chiplet design for compute.

@Ajay

Chiplet-designs should produce fewer low quality GPUs as they can test and bin the chips before combining them. So they can they combine chiplets that are binned the same. And they can toss only the truly broken chiplets, instead of having to scrap an entire monolothic chip.

 

[jpiniero]

I assume Blackwell Server will be MCM.

 

[Tigerick]

After we are done with first rumor, RGT has posted video with more SKU. I just posted the specs here for discussions:-

• There are 3 sets of SKUs with same memory sizes. Does it make sense?
• The specs of RTX4080Ti is different than first rumor just posted, of course 20GB version is making more sense cause NV will most likely drop the price of current RTX4080 to $999
• Then what is the point of RTX4070Ti with 16GB??? And RTX4060Ti with 12GB???
• RTX4070S is the topic I questioned: if NV does release Super version, then there are going to have 3 SKUs with AD104 with difference only in 46, 52 and 60 SM...
• Well, laptop wise I am not so sure cause I know AMD going to release N43 with 192-bit memory and 7900M with 256-bit bus. Maybe NV will response???? Brainstorming times??

 

[jpiniero]

RGT is dumb.

The 100 SM / 256-bit configuration is at least plausible for the 4080 Ti since that's what the RTX 5000 Ada is.

 

[Ajay]

Chiplet-designs should produce fewer low quality GPUs as they can test and bin the chips before combining them. So they can they combine chiplets that are binned the same. And they can toss only the truly broken chiplets, instead of having to scrap an entire monolothic chip.

Well, that's the plan. It has held up well with the simpler packaging being used now. Fast forward a bit into the future - and we will see. These sort of complexities are fine for something like the AMD MI300. The difficulty is getting it done on high volume GPUs. The big RDNA4 GPUs were to have more complex packaging. Now, the problem seems that it was just more time consuming to design to hit the required market window. But I am curious what the yield would be after packaging. Guess we will see when RDNA5 comes out (hopefully, as AMD knew they were running behind long before we did, RDNA5 was able to be pushed up a bit - we shall see).

 

[Ajay]

WRT Nvidia's large monolithic chip design - there is an option, once we get to high NA EUV, of using stitching to make dice with the same reticle limit that we have today. One half of each die will be exposed with one mask, and the other half, across all dice, will be exposed with a second mask. Make packings simpler and get rid of some of the problems with chiplets having to due with the structure of the DX 11/12 APIs. It's an open question what the yields would be for this sort of process. It also increases the wafer completion time to close to double that of the normal process. Anywho - lots going on over the next 5 years.

 

[Tigerick]

I am trying to compile upcoming mobile GPU from both nVidia and AMD based on leaks, well it does make sense than desktop GPU. Basically, nVidia is going to release Ti models to counter AMD's offerings either through better specs or lower price tier. The specs are not final, so pending modifications:-

AMD	Die	CU	Memory	TGP		nVidia	Die	SM	Memory	TGP
?	N43	36	10GB 160-bit GDDR6	?		RTX4070	AD106	36	8GB 128-bit GDDR6	115W
						RTX4070Ti	AD104	46	12GB 192-bit GDDR6	150W
?	N43	40	12GB 192-bit GDDR6	?		RTX4080	AD104	58	12GB 192-bit GDDR6	150W
7900M	N31	72	16GB 256-bit GDDR6	180W		RTX4080Ti	AD103	68	12GB 192-bit GDDR6	150W
						RTX4090	AD103	76	16GB 256-bit GDDR6	150W
						RTX4090Ti	AD102	100	16GB 256-bit GDDR6	200W

 

[Kepler_L2]

I am trying to compile upcoming mobile GPU from both nVidia and AMD based on leaks, well it does make sense than desktop GPU. Basically, nVidia is going to release Ti models to counter AMD's offerings either through better specs or lower price tier. The specs are not final, so pending modifications:-

AMD	Die	CU	Memory	TGP		nVidia	Die	SM	Memory	TGP
7800M	N43	36	10GB 160-bit GDDR6	?		RTX4070	AD106	36	8GB 128-bit GDDR6	115W
						RTX4070Ti	AD104	46	12GB 192-bit GDDR6	150W
7800M XT	N43	40	12GB 192-bit GDDR6	?		RTX4080	AD104	58	12GB 192-bit GDDR6	150W
7900M ?	N31	?	?	175W		RTX4080Ti	AD103	68	12GB 192-bit GDDR6	150W
7900M XT	N31	72	16GB 256-bit GDDR6	175W		RTX4090	AD103	76	16GB 256-bit GDDR6	150W
						RTX4090Ti	AD102	100	16GB 256-bit GDDR6	200W

????? Navi4 is not 7000 series.