NVidia Ada Speculation, Rumours and Discussion

Oh no not in client it is not.

Nah.

Basically everything.
Xtor scaling is rapidly dying in just about every way possible.

Proprietary cc links uhuhuhuhuhuhuhh.

 

You won't get these in mass market PC platforms any time soon.
The best we can hope for here is some type of a MCM "APU" with HBM "cache" for graphics and 4/8-channel connection to DDR5 memory.
But I have doubts about even such overengineered config being able to beat the usual CPU+dGPU combo, especially in costs.

 

Tbh they're trivial to implement.
Current xGMI runs on pretty pedestrian PCBs and all.

Why even HBM, just pile on some SRAM tiles.
Think more AMD c'mon; they make the best stuff.

 

Will likely be too expensive and an overkill in terms of bandwidth. We're not not talking about cache but more about a ~8GBs of RAM in an MCM.

 

Nope.
Those tiles are cheap.

Oh hell no non-unified memory is a big no-no.

 

Nvidia plans are 2022 and 2024. But all throughout history they had releases to respond to competition. There are people who expect nvidia to launch new cards in 2023 due to shortages, good current sales, etc. I think its the opposite. AMD is somewhat a competitor after 7 or 8 years, this will accelerate both their programs, its not gonna translate to the longest gpu cycle ever, next to Turing, from 2020 to 2023.

 

2022 is for GA100 successor, 2024 is for a successor to that successor.
The roadmap with "Nexts" they've shown isn't for gaming markets, it's for server space.
Gaming lineups are likely linked to these but the gap between launches there can be up to a year depending on market situation and the amount of difference between the lineups.

 

Technical threads are NOT the place to discuss market share. Remain on topic.

 

Oh come on, you don't have hindsights for >2025..
Regardless, I was just thinking out loud on possibilities for faster PCIe bandwidth (or some possible successor) and 3D memory stacking.

I remember when AMD bought ATi the topic of socketable GPUs also came up. Maybe in the AMD Fusion days that was somewhat of a plan, but DDRx DIMMs never really kept up with GPU demands..

 

We have OAM now!
Try not to crack bare dies installing the things tho.

Which is why SRAM piles woohoo.

 

It would be nice, if we could go back on topic, not shred another thread with the ever same market share, miners, GPU shortage discussion.

 

This thread is about speculation of the Nvidia RTX 4000 series. The market breakdown, pricing conditions, shortages, and miners impact belongs to their own threads elsewhere.

 

I don't know if this is the right place to ask, but do we know if all Ampere blocks (rt, alu, tensors...) are running at the same speed ?

If so, is "double pumped" some units possible, like in the g80 days ? I guess in some situations it would be more performant than double the units. I don't think they need a lot more alu raw power for rtx 4xxx, but they need another big RT jump imo, even if they're already ahead.

 

I think the video codec blocks have traditionally run at different clock domains, to keep the chip from consuming large amounts of power just to decode video. Of course the memory PHYs need to synchronize with the external VRAM chips, the display output blocks need to run at that standard's clock range, same with PCIe, etc.

But AFAIK the blocks you mentioned like SMs, caches, ROPs, etc. are all running at the same clocks.

 

Seems like the RT cores are rarely the bottleneck in actual games. We need more alu power, just needs to be consistently usable and not bottlenecked by whatever is happening with Ampere.

 

Regfile ain't made of rubber and gum.

 

The twitterati are hinting that Nvidia will add another layer to the compute hierarchy for its DC parts - the CPC. I can only guess that this is referring to a "cache processing cluster". Presumably this is another cache level shared by multiple TPCs similar to RDNA's L1.

 

If I were to take a guess from what is leaked so far I'd say that we won't get 40 series till 2023.
2022 - GH100 (Hopper) and 30 Super
2023 - Lovelace
So presumably Nv will repeat the cadence of Volta->Turing with HPC->gaming parts.

 

So what is 30 Super going to be? Refresh at 6 nm?

 

Same 8N process, addition of GA103 GPU (~3070Ti?), more VRAM, lower prices.