Speculation and Rumors: Nvidia Blackwell ...

  • Thread starter Deleted member 2197
  • Start date
I think the 5090 will be a chiplet/MCM. With the rumor that the 5080 has a 256-bit bus, the 5090 will adjust be 2 of these glued together with the new chip interconnect use in B100. This will be Nvidias version of the Apple Ultra chips.

1 tapeout and they can create a multitude of products across the gaming lines and AI products.
 
You all forget one important thing. When Blackwell was designed, the top SKU was supposed to fight a monster RDN4 MCM and Nvidia didn't want to loose the halo battle. That's why the gaming Blackwell MCM was designed.
Now that AMD gave up on this segment for this generation, the dynamic of the market is totally different. In other words, Nvidia has many possibilities but knowing Jensen a bit, they will just follow their initial plan with some adjustments on SKUs/prices/timeframe...
 
The 5080 being 1199 at least for a 16GB card with a die that's smaller than AD103 supposedly is depressing. If MLID (I know, I know) saying he reckons they're probably doing it for Chinese market for those people to snap it up. Like if GB203 is smaller than AD103 with a process that's dropped in price quite a bit in 24/25 they could do $799 with a full die, two slot cooler @ 250W with much more better Perf Per Watt (and less overbuilt board).

I mean I am planning to wait late next year when RDNA 5 launches and then Grab a 5090 Ti/Super/5090 Ti Super whichever uses a full(ish) GB202 die but still annoyed how Nvidia could just craft a really good lineup but just gets greedy for a quarter or two until Market correction. Wouldn't be suprised if this comes the norm that Nvidia released an overpriced-ish launch lineup and a year later there abouts when AMD does RDNA 6 etc then the market corrects.
 
DegustatoR said:
Somehow I really doubt that GB203 will be smaller than AD103 on the same process.
Click to expand...
Well @xpea has implied it will be since they've been able to increase transistor density with AI based innovations. I mean 30% with B100 so I guess GB203 is probably less but still substantial (maybe 10-20%?), So could be 350mm2 or there abouts? Basically smaller than AD103 (376mm2).
 
Dangerman said:
Well @xpea has implied it will be since they've been able to increase transistor density with AI based innovations. I mean 30% with B100 so I guess GB203 is probably less but still substantial (maybe 10-20%?), So could be 350mm2 or there abouts? Basically smaller than AD103 (376mm2).
Click to expand...
Higher density on the same process could be just a result of using different process libraries and not necessarily related to any innovations. Blackwell is a new architecture so whatever they will save this way will likely be eaten up by additional features. This transition will probably be similar to Pascal->Turing, maybe not as extreme but comparable. Thus chips being smaller is something I find unlikely.
 
Dangerman said:
The 5080 being 1199 at least for a 16GB card with a die that's smaller than AD103 supposedly is depressing. If MLID (I know, I know) saying he reckons they're probably doing it for Chinese market for those people to snap it up. Like if GB203 is smaller than AD103 with a process that's dropped in price quite a bit in 24/25 they could do $799 with a full die, two slot cooler @ 250W with much more better Perf Per Watt (and less overbuilt board).

I mean I am planning to wait late next year when RDNA 5 launches and then Grab a 5090 Ti/Super/5090 Ti Super whichever uses a full(ish) GB202 die but still annoyed how Nvidia could just craft a really good lineup but just gets greedy for a quarter or two until Market correction. Wouldn't be suprised if this comes the norm that Nvidia released an overpriced-ish launch lineup and a year later there abouts when AMD does RDNA 6 etc then the market corrects.
Click to expand...
Nvidia isn't going to offer better value if consumers like yourself are just going to give in anyways. Raising prices by a foot, then conceding an inch or so is not at all a 'correction', either. It's just giving us crumbs as they gorge themselves at the banquet!

DegustatoR said:
Somehow I really doubt that GB203 will be smaller than AD103 on the same process.
Click to expand...
Proportionally, GB203 is supposed to be smaller compared to the top die than with Lovelace, but we dont know top die size. Either way, if it's really just 96SM's, that's not much more than AD103. I reckon it'll be pretty similar in size to AD103. Smaller isn't impossible.
 
wccftech.com

NVIDIA Shares Blackwell GPU Compute Stats: 30% More FP64 Than Hopper, 30x Faster In Simulation & Science, 18X Faster Than CPUs

NVIDIA has shared more performance statistics of its next-gen Blackwell GPU architecture which has taken the industry by storm.
wccftech.com wccftech.com
In a new blog post, NVIDIA has shared how Blackwell GPUs are going to add more performance to the research segment which includes Quantum Computing, Drug Discovery, Fusion Energy, Physics-based simulations, scientific computing, & more. When the architecture was originally announced at GTC 2024, the company showcased some big numbers but we have yet to get a proper look at the architecture itself. While we wait for that, the company has more figures for us to consume.
...
NVIDIA also sheds light on the double-precision of FP64 (Floating Point) capabilities of its Blackwell GPUs which are rated at 30% more TFLOPs than Hopper. A single Hopper H100 GPU offers around 34 TFLOPs of FP64 compute and a single Blackwell B100 GPU offers around 45 TFLOPs of compute performance. Blackwell mostly comes in the GB200 Superchip which includes two GPUs along with the Grace CPU so that's around 90 TFLOPs of FP64 compute capabilities. A single chip is behind the AMD MI300X and MI300A Instinct accelerators which offer 81.7 & 61.3 TFLOPs of FP64 capabilities on a single chip.
...
NVIDIA quickly shifts gears and brings us AI performance once again where its Blackwell GB200 GPU platform once again reigns supreme with a 30x gain over H100 in GPT (1.8 Trillion Parameter). The GB200 NVL72 platform enables up to 30x higher throughput while achieving 25x higher energy efficiency and 25x lower TCO (Total Cost of Operation). Even putting the GB200 NVL72 system against 72 x86 CPUs yields an 18x gain for the Blackwell system and a 3.27x gain over the GH200 NVL72 system in Database Join Query.
NVIDIA-Blackwell-GPU-Specs.png

But NVIDIA isn't stopping any time soon as the company is anticipated to start production of its next-gen Rubin R100 GPUs by as early as late 2025 and the initial details sound insane.
Click to expand...
 
Look at how impressive our numbers are if we multiply them juuuust right, ignore that you have to re-write your entire codebase to use our new formats that might be entirely unusable for your models!

Jim Keller said this chip cost 10x too much to develop, which considering a lot of the actual performance gains comes from taping two giant chips together, is probably generous. I expected Nvidia to slow down as it got too big and too rich to focus well, but this is a bit on the surprisingly fast end of the scenarios I had in mind.

Maybe consumer will be more impressive.
 
Frenetic Pony said:
Jim Keller said this chip cost 10x too much to develop, which considering a lot of the actual performance gains comes from taping two giant chips together, is probably generous. I expected Nvidia to slow down as it got too big and too rich to focus well, but this is a bit on the surprisingly fast end of the scenarios I had in mind.

Maybe consumer will be more impressive.
Click to expand...


The big cherry picked numbers are obviously dubious. Nvidia should be ashamed to use such trashcan level marketing tactics...

But Mr Keller is also disconnected from the real world. The R&D cost should not be seen in a vacuum but always put in relation to the ROI. And Blackwell will deliver so much revenue that its development cost is a drop in the 100+ billion dollar ocean of expected generated profit...
 
Frenetic Pony said:
Look at how impressive our numbers are if we multiply them juuuust right, ignore that you have to re-write your entire codebase to use our new formats that might be entirely unusable for your models!

Jim Keller said this chip cost 10x too much to develop, which considering a lot of the actual performance gains comes from taping two giant chips together, is probably generous. I expected Nvidia to slow down as it got too big and too rich to focus well, but this is a bit on the surprisingly fast end of the scenarios I had in mind.

Maybe consumer will be more impressive.
Click to expand...
Jim said this ironically. I doubt it that CowoS L cost as much as 10 billions.
 
xpea said:
The big cherry picked numbers are obviously dubious. Nvidia should be ashamed to use such trashcan level marketing tactics...

But Mr Keller is also disconnected from the real world. The R&D cost should not be seen in a vacuum but always put in relation to the ROI. And Blackwell will deliver so much revenue that its development cost is a drop in the 100+ billion dollar ocean of expected generated profit...
Click to expand...

ROI is missing the point, what matters is what it could be done for and what it was done for. Spending 10x as much as needed is an "opportunity cost", they let 9 billion slip for no reason, and this sort of thing companies on the downslope do. The prime example would be Apple, it's spent the past decade not caring that it's near, or at, the biggest company around. Tim kept costs down, continued to wring every penny out of consumers he could, and ended up being rewarded for it by fighting for the biggest publicly listed company around.

Nvidia paying out million dollar+ salaries to employees is great for the employees. But by accounts of people working with them it's made some of those same managers/high level employees tend towards slacking off and not caring. They've got their house paid off and enough in stock options to retire today, what do they care if the company succeeds a bit more or a bit less? They don't anymore, but these are the very people Nvidia would be relying on the keep growing.

Which is why I'm going to guess Blackwell consumer isn't going to be super impressive either. Enough to fight AMD's hyper truncated RDNA4, and Intel is still struggling with software and perception with Battlemage. Not like they're going to collapse quite yet (let's see what RDNA5 is). But as a 2 year update probably not the most impressive thing around.
 
You must log in or register to reply here.
Back
Top