Discussion Nvidia Blackwell in Q1-2025

[adroc_thurston]

The MS one is probally gonna be over a grand.

Why the hell would it be?

 

[jpiniero]

Why the hell would it be?

N3 is expensive.

Plus it's widely rumored that MS will let you use it as a PC and install Steam, etc.

 

[adroc_thurston]

N3 is expensive.

It's a tiny tiled part.

Plus it's widely rumored that MS will let you use it as a PC and install Steam, etc.

Nope. It's an xbox.

 

[jpiniero]

https://videocardz.com/newz/nvidia-rubin-cpx-render-hints-at-possible-link-to-future-geforce-rtx-6090-design

This guy seems to think it's 16 GPCs x 6 TPCs. Which would keep it at 192 SMs. GB202 is 12x8.

 

[basix]

Well, 16 GPCs are visible on the diagram. It does not mean, that the chip will really look like this.

But I think 16 GPCs is reasonable. I can think of a few reasons for that:

• Better scaling in general, because 1.33x GPCs (ROPs anyone? )
• Power of 2 scaling for ML/AI Workloads: A100, H100 and B200 all have 8x GPCs. Rubin HPC might have 8 or 16 and Rubin CPX would have 16 GPC as well. ML/AI workloads like power of 2 divisions (MI350X did go back to 256 CU because of that, MI300X featured 320 CU)
• GauRast: Gaussian Splatting Acceleration (Neural Rendering) within the rasterizer. 1.33x GPCs will bring a boost there (see the GauRast paper https://arxiv.org/html/2503.16681v1)
• Transformer Attention Acceleration (see GB300 or Rubin CPX) does benefit from exponential functions. On GB300 Nvidia says they pimped the SFU within the SM for that (or emulate SFU EX2 functions). GauRast from the previous point will introduce additional EXP-Units within the rasterizer (GPC frontend). This might create some synergies between Neural Rendering and general ML/AI.

 

[fastandfurious6]

xbone really doesnt make any sense anymore... just get a pc lol

 

[Win2012R2]

xbone really doesnt make any sense anymore... just get a PS5 lol

FTFY

 

[poke01]

https://videocardz.com/newz/seasonic-lists-geforce-rtx-5070-ti-super-and-rtx-5070-super-in-its-psu-calculator

The only blackwell cards worth buying are coming soonish

 

[jpiniero]

https://videocardz.com/newz/seasonic-lists-geforce-rtx-5070-ti-super-and-rtx-5070-super-in-its-psu-calculator

The only blackwell cards worth buying are coming soonish

I do agree that nVidia delayed it. Maybe CES.

 

[Win2012R2]

It's still a bad arch even with more memory.

 

[jpiniero]

It's still a bad arch even with more memory.

Tough to get that much better performance while using the same node and a mandate to keep transistor budgets flat if not lower.

It definitely would have looked better if they had used N4P instead of N5.

 

[Win2012R2]

Tough to get that much better performance while using the same node and a mandate to keep transistor budgets flat if not lower.

Dropping 32bit PhysX was Borderlands 2 criminal...

But seriously the whole package clearly does not work as they intended, maybe it's -10% clocks, but just feels something gone wrong there, with all that "AI" money Nvidia should be able to afford making a proper gaming card.

It definitely would have looked better if they had used N4P instead of N5.

Ada was designed using special TSMC 4N, no chance they used N5 for Blackwell.

 

[jpiniero]

Ada was designed using special TSMC 4N, no chance they used N5 for Blackwell.

4N is an nVidia N5 special.

Rubin, btw, looks like what I thought Blackwell is going to be. It looks like they will double down (if not more) on tensor cores, leaving little room for raster or RT improvements beyond higher clocks. And that's assuming they don't cut core counts at the lower end to keep costs in line with N3's way higher prices.

 

[Win2012R2]

leaving little room for raster or RT improvements beyond higher clocks

Great for RDNA5 - hopefully AMD won't blow their true real chance to take the lead.