Speculation and Rumors: Nvidia Blackwell ...

  • Thread starter Deleted member 2197
  • Start date
DegustatoR said:
This looks like a rather meaningless marketing placeholders tbh. All videocards are capable of "neural rendering" and of course there will be "advanced DLSS" and "enhanced RT" on future products.
Click to expand...
No no, something is coming ... NVIDIA just released a new SDK for "In-Game Inference".

Integrated with Graphics Pipelines. Offers native integration into game pipelines and simultaneous CUDA and graphics execution with low latency
Click to expand...

supports any cloud API endpoint including NVIDIA NIM, as well as local execution on PCs. For local execution, developers can utilize either an in-process execution method to integrate directly with latency-sensitive applications, or an out-of-process execution method that integrates as a service within an application
Click to expand...

developer.nvidia.com

NVIDIA In-Game Inferencing SDK

Integrate AI models into apps to manage deployment across devices.
developer.nvidia.com developer.nvidia.com
 
DavidGraham said:
RTX 5000 will come with a new "neural rendering capabilities" feature, according to marketing points leaked by INNO3D. Also "advanced' DLSS to offer more image quality and faster fps.

Click to expand...

It’s the software side that’s most interesting to me. I’ll also be curious how many of the software improvements are backwards compatible vs just for 5000 series.

Of course a new gen of RT cores is great to see.
 
RobertR1 said:
It’s the software side that’s most interesting to me. I’ll also be curious how many of the software improvements are backwards compatible vs just for 5000 series.

Of course a new gen of RT cores is great to see.
Click to expand...
I read somewhere that NVIDIA acutally has more software engineers than hardware enigineers, so would not surpise me one bit if they try and leverage that fact.
 
RobertR1 said:
It’s the software side that’s most interesting to me. I’ll also be curious how many of the software improvements are backwards compatible vs just for 5000 series.

Of course a new gen of RT cores is great to see.
Click to expand...
I mean Blackwell is intriguing because the new rumoured architectural overhaull which I wonder how well it will be put to use with RT & PT improvements (though Rubin will architecturally just be Blackwell on TSMC N3P for Consumer Desktop, maybe even Chiplets?). Though I suspect GDDR7 will do a some of the heavy lifiting in ray & path-tracing uplift.
 
I hope for a significant architecture overhaul, not a Ada+. Given how they don't have a node jump this year they need it more than ever.
 
Dampf said:
I hope for a significant architecture overhaul, not a Ada+. Given how they don't have a node jump this year they need it more than ever.
Click to expand...
It seems to be a thing Nvidia does in that everytime the node stagnates is when they do a big architectural leap. Like Maxwell, Turing and now Blackwell.

My prediction is RTX 60/Rubin & RTX 70/Rubin successor will iterate on Blackwell (but probably go chiplet based on RTX 60 or at the latest 70) and then RTX 80 stays on the same node as 70 but does a new big architectural jump and GDDR8.
 
DegustatoR said:
I do wonder when we might see x1.5 capacity chips though and if that can bring some changes to the Tis and/or a possible Super refresh in 2026.
Click to expand...

My understanding is that these are expected to be retail ready around Q3 2025, definitely in time for a refresh. NVIDIA could be keeping the memory speed lower on the 5080 to keep costs down and give them segmentation room for a 24 GB 5080 Super (not that it will matter much since the chip is already fully enabled).
 
Broopster said:
My understanding is that these are expected to be retail ready around Q3 2025, definitely in time for a refresh. NVIDIA could be keeping the memory speed lower on the 5080 to keep costs down and give them segmentation room for a 24 GB 5080 Super (not that it will matter much since the chip is already fully enabled).
Click to expand...
I mean the "400W" rumoured TBP is probably room to upclock for a 5080 Super IMV. Along with a 5090 being "600W" where a 48GB Super refresh with more SMs enabled.
 
Dangerman said:
I mean the "400W" rumoured TBP is probably room to upclock for a 5080 Super IMV. Along with a 5090 being "600W" where a 48GB Super refresh with more SMs enabled.
Click to expand...
I’m sure it’ll be clocked a little higher but I doubt it’ll make a big difference (look at the 4080 Super). The big draw will be the extra VRAM.

I doubt they will release a refresh for the 5090, if they do it would probably just have more SM and faster memory. I do think they’ll release the pro models with GB202 with 3GB memory modules though.
 
Last edited:
trinibwoy said:
Yeah I meant lots of alternative options at a price if your heart is set on more than 8GB. The vast majority of games and gamers will be fine with 8GB entry level cards.
Click to expand...
That would be fine if it was being sold as a 5050/5050Ti for less than $240, as it should be. Because yes, this is indeed an entry level GPU.

But being sold as a 5060 and for $300+ is firmly into lower midrange territory and in the year 2025, over four years after the new generation consoles have come out, should have the expectation of getting somebody a decent experience in the latest games without having to play with lower-than-console visuals.
 
Let's hope Intel continues to provide us with an inexpensive >8GB VRAM alternatives. I'd also hope AMD sees this new "memory war" as an easy opportunity to pry a few NV fans away...
 
Albuquerque said:
Let's hope Intel continues to provide us with an inexpensive >8GB VRAM alternatives. I'd also hope AMD sees this new "memory war" as an easy opportunity to pry a few NV fans away...
Click to expand...
It should be an easy win for AMD. A 12GB card for $300 is well within the realm of possibility. Though they may have designed their stuff (memory bus) with the assumption that higher capacity memory modules would be available. That would leave them in an awkward spot same as NVIDIA.
 
Albuquerque said:
Let's hope Intel continues to provide us with an inexpensive >8GB VRAM alternatives. I'd also hope AMD sees this new "memory war" as an easy opportunity to pry a few NV fans away...
Click to expand...
Intel had to inflate their BMG-G21 die to ~AD104 level to fit the 192 bit bus. Other options of providing more VRAM are limited to using G7's new chip densities - and it seem to be Nvidia exclusive for now.
N44 is 128 bit so as 7600 either 8 or 16GBs, the latter is likely an overkill for such performance.
N48 is 256 bit and this posit an interesting question on how AMD will position 9070 and 9070XT - the latter should be at 7900GRE level which also has 16GB but the former will likely be close to 7800XT while having 192 bit bus and thus - could result in a downgrade from 16 to 12GBs.
 
@DegustatoR reminds us of a relevant physical point: the actual die has to be big enough to fit all the necessary memory pinouts for whatever memory we are hoping for. As lithography processes permit smaller and smaller dies, we absolutely run into a problem with there simply not being enough literal room on the substrate to connect all the memory pins we need to get the capacity we desire. We've run into this problem before, and it's an interesting problem to be sure.
 
This, and also the fact that PHYs don't scale that well since about 7nm which means that the more of these you put into a chip - the more expensive transistors you're spending on something which hasn't been an improvement for years now.

It's two main reasons why the last couple of generations now have opted for narrower buses and bigger caches. The inability to put more VRAM on that is a side effect of chip production economics. Hopefully everyone adopting G7 in a couple of years will take care of that.
 
Is there not room around something like AD106 for 6 32bit interfaces? I can't find a die shot of AD106 and the shots I found of AD102 say they're not to scale.
 
You must log in or register to reply here.
Back
Top