• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Blackwell gpus

He's a really bad presenter tonight isn't he, he joked about 'this is what hapens when you don't rehearse' and it's really showing.
A tonne of forced pauses for undeserved applause, nothing of much substance (that I'm looking for anyway) yet.

"New GPU go vroom, line on chart go up"
 
Last edited:
Yup.

"The first model we delivered to OpenAI I could carry, now the new one is bigger and I can't carry it"

Errr...good for you?
 
He's a really bad presenter tonight isn't he, he joked about 'this is what hapens when you don't rehearse' and it's really showing.
A tonne of forced pauses for undeserved applause, nothing of much substance (that I'm looking for anyway) yet.

"New GPU go vroom, line on chart go up"

He tries to be like Jobs in the way he presents, but he's literally a poundland grade version.
 
Blackwell B100/B200 specs, it's not much but all we know for now
  • Custom TSMC 4NP Process Node
  • Multi-Chip-Package GPU with 2 GPU Die
  • 208 Billion Transistors (108b per die)
  • 160 SMs (20,480 Cores)
  • 192 GB HBM3e Memory with 8 TB/s Memory Bandwidth over a 8192-bit Memory Bus Interface
  • PCIe 6.0 Support
  • 700W peak TDP


Claimed performance:
  • 20 PFLOPS FP8 (2.5x Hopper)
  • 20 PFLOPS FP6 (2.5x Hopper)
  • 40 PFLOPS FP4 (5.0x Hopper)
 
Last edited:
Blackwell B100/B200 specs, it's not much but all we know for now
  • Custom TSMC 4NP Process Node
  • Multi-Chip-Package GPU with 2 GPU Die
  • 208 Billion Transistors (108b per die)
  • 160 SMs (20,480 Cores) per die
  • 192 GB HBM3e Memory with 8 TB/s Memory Bandwidth over a 8192-bit Memory Bus Interface
  • PCIe 6.0 Support
  • 700W peak TDP


Claimed performance:
  • 20 PFLOPS FP8 (2.5x Hopper)
  • 20 PFLOPS FP6 (2.5x Hopper)
  • 40 PFLOPS FP4 (5.0x Hopper)


Here is the thing though: I suspect these performance numbers are from comparing the old GPU which had one die to the new GPU with two dies and the new die has 25% more cores than the old dies. So when you attribute how much of the performance gain is from the the extra cores, it's almost all of it - Blackwell doesn't seem like it has big IPC or clock improvements, the extra performance is from the extra die and cores, and the gaming GPUs won't get the extra die and cores I'm not sure about because B100/B200 is right at the reticule limit, I.e it's expensive as hell to build

Based on this, I suspect most of Blackwell gaming GPU performance gains would come from AI stuff like new software or improvements in existing AI functions, raster performance improvements may be minimal guys
 
Last edited:
Based on this, I suspect most of Blackwell gaming GPU performance gains would come from AI stuff like new software or improvements in existing AI functions, raster performance improvements may be minimal guys

So like 40 series blackwell may be another hype for AI and we'll all pay heavily for it, best get the KY jelly ready and a re-morgage organised, if you have spare organs i guess thats a path?. or as jolly Jensen would say "the more you buy the more you save"
 
Last edited:
Here is the thing though: I suspect these performance numbers are from comparing the old GPU which had one die to the new GPU with two dies and the new die has 25% more cores than the old dies. So when you attribute how much of the performance gain is from the the extra cores, it's almost all of it - Blackwell doesn't seem like it has big IPC or clock improvements, the extra performance is from the extra die and cores, and the gaming GPUs won't get the extra die and cores I'm not sure about because B100/B200 is right at the reticule limit, I.e it's expensive as hell to build

Based on this, I suspect most of Blackwell gaming GPU performance gains would come from AI stuff like new software or improvements in existing AI functions, raster performance improvements may be minimal guys
Interesting thought, I'd not been able to pay attention enough to realise this.
 
It will just be a better cooler, some faster clocks with GDDR7 memory increasing the memory bandwidth and different VRAM configuration with the 5080 now coming in with 18/21GB and the 90 class with more?

And obviously with that comes more heat from the higher TDP.
 
Back
Top Bottom