Nvidia Blackwell Architecture Speculation

  • Thread starter Deleted member 2197
  • Start date
AMD has me worried we won’t be getting any actual Blackwell info.

There are already several Chinese stores showing boxes of the product, the only thing missing is the official launch.

GgTSow7XUAAFNyQ
GgnGLhybQAABkR6
GgnQLXcWsAAMrSw
 
Did NVIDIA just open by referring to tokens helping robots to imitate their “masters?” Cancelled! 😝. I would also like to know how much Jensen paid Michael Jackson’s estate for that jacket.

Hopefully we can get to some actual hardware soon.
 
That guy that made the pricing video is an idiot. Click-baiting like crazy.

We'll see what happens once these instantly sell out and scalpers get their hands on them.
 
That guy that made the pricing video is an idiot. Click-baiting like crazy.

We'll see what happens once these instantly sell out and scalpers get their hands on them.

Figure around $3K for the next year to actually get ahold of a 5090 Card?
 
We'll have to see if adoption of Neural texture compression takes off like DLSS eventually did. I wonder how much effort it will require to implement?

Yah. I'd say I'm a little skeptical, because what's the fallback for your game if the gpu doesn't support it? You ship the full textures alongside it? I'm sure they'll throw money at some companies to use it and we'll at least get to see what it's like.
 
Yah. I'd say I'm a little skeptical, because what's the fallback for your game if the gpu doesn't support it? You ship the full textures alongside it? I'm sure they'll throw money at some companies to use it and we'll at least get to see what it's like.
With AMD on AI train too now chances of these ML things being standardized and used by everyone are improving I'd say.
 
Our new frame generation AI model is 40% faster, uses 30% less VRAM, and only needs to run once per rendered frame to generate multiple frames. For example, in Warhammer 40,000: Darktide, this model provided a 10% faster frame rate, while using 400MB less memory at 4K, max settings, using DLSS Frame Generation.

1736220027581.png
 
The offical links are starting.




Seems like Nvidia is switching from CNNs to ViTs for some of this stuff.
 
Interesting little tidbits from the main announcement link.

With up to 92 billion transistors, Blackwell is the most powerful consumer GPU ever created. The Blackwell streaming multiprocessor (SM) has been updated with more processing throughput, and a tighter integration with the Tensor Cores in order to optimize the performance of neural shaders. Blackwell is enhanced by several hardware and software innovations to improve Shader Execution Reordering. The reorder logic is twice as efficient, increasing the speed and precision of reordering which accelerates the performance of neural shaders.

New Blackwell Tensor Cores have been built with a massive amount of AI processing horsepower, and support accelerated processing of FP4 precision models. With FP4, Blackwell Tensor Cores can not only process models faster, but can do so while using less graphics memory.

To help support the frame pacing requirements of next generation DLSS Multi Frame Generation, the Blackwell architecture was built with enhanced hardware flip metering capabilities to provide the speed and accuracy required for a smooth, high-quality experience.

Blackwell also features brand new RT Cores designed to ray trace massive amounts of detailed geometry. The RT Cores have 2X the ray triangle intersection rate of the previous generation, and enhanced compression designed to reduce memory footprint. This allows Blackwell GPUs to ray trace levels of geometry that were never before possible.

Blackwell has also been enhanced with PCIe Gen5 and DisplayPort 2.1b UHBR20, driving displays up to 8K 165Hz.

For GeForce RTX 50 Series laptops, new Max-Q technologies such as Advanced Power Gating, Low Latency Sleep, and Accelerated Frequency Switching increases battery life by up to 40%, compared to the previous generation.

And in order to feed all this processing power, Blackwell is equipped with the world's fastest memory - GDDR7 with speeds up to 30Gbps. With G7 memory, Blackwell GPUs can deliver up to 1.8TB/s of memory bandwidth.
 
Back
Top