Speculation and Rumors: Nvidia Blackwell ...

Kaotik said:
www.tomshardware.com

Nvidia addresses significant Blackwell yield issues, production ramps in Q4

Hopper set to remain Nvidia's datacenter working horse for this year.
www.tomshardware.com www.tomshardware.com

NVIDIA has confirmed the delay, they've had to respin both B100 & B200 to improve yields. Shipping starts Q4 fiscal year apparently, which means it might be 2025 calendar already.
Click to expand...

There may be a delay but Nvidia sure didn’t confirm it. All they said is that there was a respin and that mass production starts Q4. The rest is creative journalism unless there was a prior announcement of mass production starting earlier than Q4.
 
From the wording the respin seem to be aimed at increasing yields which means that they can produce and ship chips without it (and they do actually) but are making a respin for higher volume. Don't know if that even counts as "a delay".
 
improving yields could be an euphemism. Could very well be that it would damage the margins too much to ship any significant volumes in current state
 
The phrasing is just atrocious but seems more clickbait than anything else. Either ways I'd be more interested in the chip density than power consumption.
 
I'll stick with "AI datacenter bin is around 600w, max consumer bin is more like 500w".

600w is a lot to put through a consumer power cable and ask from a consumer cooling setup. Making a chip like that for existing datacenter on the other hand makes a lot of sense. Then you just toss the worse candidates in a "5090/ti" bin and sell it to consumers.
 
Last edited:
It's rumored to be a 4nm process (5nm class) from TSMC but I don't know if that specifically means it's the exact same process. TSMC itself had several iterations since Ada launched such as N4P (maybe what 4NP is dervied from), N4X (stated for volume production in 2024) and N4C (stated for 2025, cost reduced iteration). Nvidia says Blackwell (B100) is on TSMC 4NP instead of 4N for what it's worth.

I don't think Nvidia and/or TSMC have ever been specific in terms of details on what Ada (and Hopper, and Blackwell) actually use beyond the term 4N (and 4NP for Blackwell) and "custom" thrown around. So I don't know how much going to specifics matters here as I don't think there is anything substantial in terms of actual details and public terms could be as much marketing/obfuscation, not just technical meaning.
 
Last edited:
arandomguy said:
It's rumored to be a 4nm process (5nm class) from TSMC but I don't know if that specifically means it's the exact same process. TSMC itself had several iterations since Ada launched such as N4P (maybe what 4NP is dervied from), N4X (stated for volume production in 2024) and N4C (stated for 2025, cost reduced iteration). Nvidia says Blackwell (B100) is on TSMC 4NP instead of 4N for what it's worth.

I don't think Nvidia and/or TSMC have ever been specific in terms of details on what Ada (and Hopper, and Blackwell) actually use beyond the term 4N (and 4NP for Blackwell) and "custom" thrown around. So I don't know how much going to specifics matters here as I don't think there is anything substantial in terms of actual details and public terms could be as much marketing/obfuscation, not just technical meaning.
Click to expand...
No significant advancement from the manufacturing process though I take it. So it's either a power increase and/or architecture improvement to increase perf.
 
techuse said:
No significant advancement from the manufacturing process though I take it. So it's either a power increase and/or architecture improvement to increase perf.
Click to expand...

Assuming similar characteristics here (eg. 4N = N4, 4NP = N4P) depending on the figures used it could be 6-10% performance from process.

GDDR7 should be a factor here as well in terms of possible gains.

Blackwell is rumoured to different significantly architecturely along the lines of Maxwell or Turing.
 
With AMD being MIA in top tier the only reason why Nvidia would increase power limit on 5090 is if it wouldn't be much of an upgrade over 4090 at the same power. The same rumor however says that 5080 at 400W should be 10% faster than 4090 (a 450W SKU) suggesting that the architectural changes and the process improvements should be enough to put 5090 considerably ahead of 4090 on the same power. So why would Nvidia increase the power limit on 5090 in comparison to 4090? It's possible of course that they want to go all out with the SKU from the start instead of adding some 5090Ti later when RDNA5 will hit the market. But this seems like a weird thing to do prior to knowing anything about the top RDNA5 SKU.
 

DegustatoR said:
With AMD being MIA in top tier the only reason why Nvidia would increase power limit on 5090 is if it wouldn't be much of an upgrade over 4090 at the same power. The same rumor however says that 5080 at 400W should be 10% faster than 4090 (a 450W SKU) suggesting that the architectural changes and the process improvements should be enough to put 5090 considerably ahead of 4090 on the same power. So why would Nvidia increase the power limit on 5090 in comparison to 4090? It's possible of course that they want to go all out with the SKU from the start instead of adding some 5090Ti later when RDNA5 will hit the market. But this seems like a weird thing to do prior to knowing anything about the top RDNA5 SKU.
Click to expand...

How much performance does a 4090 lose if capped to 400W though? Given diminishing returns at higher wattage a 10% advantage at 50W less may not be anything to write home about.
 
trinibwoy said:
How much performance does a 4090 lose if capped to 400W though? Given diminishing returns at higher wattage a 10% advantage at 50W less may not be anything to write home about.
Click to expand...
That's a different question and we have no way of answering that without knowing how much performance would 5080 and 5090 lose in the same way.
 
DegustatoR said:
That's a different question and we have no way of answering that without knowing how much performance would 5080 and 5090 lose in the same way.
Click to expand...

We can try to extrapolate architectural efficiency by looking at 400W performance though. If the 5080 really does need 400W to beat the 4090 by 10% that doesn’t seem very impressive. At 400W the 4090 barely loses any performance.
 
You must log in or register to reply here.
Back
Top