I'm not super knowledgeable on networking stuff but standards-wise we went from 1gbps in 1998 to 10gbps 2002. So 4years for a 10x increase.
How long it took for it to actually materialize and be beneficial... I have no idea.
How long it took for it to actually materialize and be beneficial... I have no idea.
10GBps in 2002 was over fiber. 10GBase-T was released in 2006, 4 years later. It seems that ~10X over 8 years is the norm over different technologies.
It sounded to me (based on Jen Hsun’s insistence they will push everything including process node to the limit) that Blackwell Ultra might actually be a N3E or N3P shrink (maybe with more SMs but maybe not?) rather than just Blackwell with faster memory.
That would make a lot more sense than the previous rumours implying they’d have a new generation every year; more of a tick-tock with different teams working on them, which is also what NVIDIA did very early on (see: TNT->TNT2, GF1->GF2, GF3->GF4, etc… none of them had any big changes besides the process, except maybe the GF2 going from 4x1 to 4x2 pipelines, but there were rumours at the time GF1 was supposed to be 4x2 but bugged).
That would make a lot more sense than the previous rumours implying they’d have a new generation every year; more of a tick-tock with different teams working on them, which is also what NVIDIA did very early on (see: TNT->TNT2, GF1->GF2, GF3->GF4, etc… none of them had any big changes besides the process, except maybe the GF2 going from 4x1 to 4x2 pipelines, but there were rumours at the time GF1 was supposed to be 4x2 but bugged).
Rubin is similar still different than Blackwell in order to offer substantial performance improvement.
Similar:
2 dies linked by a mega bandwidth Silicon bridge
8 stacks of HBM
Different:
I/O have now their own dies (*2) so the main 800mm2 dies on N3P are compute only for maximum density
HBM is now version "4" but it's custom for NV to reduce power and accommodate their aggressive schedule
Similar:
2 dies linked by a mega bandwidth Silicon bridge
8 stacks of HBM
Different:
I/O have now their own dies (*2) so the main 800mm2 dies on N3P are compute only for maximum density
HBM is now version "4" but it's custom for NV to reduce power and accommodate their aggressive schedule
Last edited:
VR200 is the successor of GB200
VR for Vera (CPU) Rubin (GPU)
I got the same feeling from the presentation. Would also fit with Rubin N3P and Rubin Ultra N2.
I always figured this was gonna be the case. They weren't gonna have the ability to do completely new architectures every year, nor would that even make a lot of sense. They only ever said they wanted to have a new 'generation' every year or whatever. But generation doesn't have to mean some entirely new architecture or anything. Intel 14th generation CPU's, for instance....
So yea, a proper new architectural leap, and then a more optimized/refined/spec-improved/process-leap version of it in between.
3 months later:
forum.beyond3d.com
forum.beyond3d.com
N100 is not Blackwell successor. It's Rubin
Rubin is not a 4x830mm2 die monstrosity (it never existed)
Just to say. Believe who you want...
Speculation and Rumors: Nvidia Blackwell ...
Considering there's apparently a 30% transistory density increase I do wonder; where does that put the 96SM rumoured GB203? Because if that is technically smaller than AD103 like maybe 350mm2 I do wonder if that means a 700mm2 GB202 which on N4P/X sounds way more economical than a 600mm2 N3E...
forum.beyond3d.com
Speculation and Rumors: Nvidia Blackwell ...
Considering there's apparently a 30% transistory density increase I do wonder; where does that put the 96SM rumoured GB203? Because if that is technically smaller than AD103 like maybe 350mm2 I do wonder if that means a 700mm2 GB202 which on N4P/X sounds way more economical than a 600mm2 N3E...
forum.beyond3d.com
Rubin is not a 4x830mm2 die monstrosity (it never existed)
Just to say. Believe who you want...
DavidGraham
Veteran
Jensen is claiming that the days of 1 million GPUs in a single data center are coming. And Elon Musk is claiming that his cluster of 100K H100s is going to be online in a few months, after that he is planning for a 300K B200s cluster in 2025.
DegustatoR
Legend
Isn't it just B200?
B200 Ultra is same B200 silicon but paired with 12-Hi stack HBM to reach 288GB RAM (vs 192GB for B200)
B200 refers to B for Blackwell generation GPU and 200 for 2* 100 GPU dies. No relationship with X86
GB200 is G for Grace CPU, B for Blackwell GPU and 200 for 2* 100 GPU dies
VR200 is V for Vera CPU, R for Rubin GPU and 2* 100 GPU dies
Not for nVidia. B100 (700W) and B200 (1000W) is Blackwell combined with x86. GB200 is the Grace-Blackwell Superchip: https://resources.nvidia.com/en-us-blackwell-architecture
well their marketing is not consistent. But internally it's like I said
I don’t think it’s possible to tell from that keynote whether Blackwell-Ultra is “just another H200” (which was also more bandwidth, and possibly a respin, not just more capacity) or if it is indeed a process shrink or more. Besides the insistence on pushing the limits of everything, the only other hint is H100 and H200 have the same “mini die shot” on the slide while Blackwell/Blackwell-Ultra are different, which is hardly conclusive.
However if Blackwell-Ultra is a H200-level change or less, then let me say bluntly that the “yearly cadence” would be complete BS. No one inside a GPU company would consider that a yearly cadence, especially when there have been companies with true yearly cadences in the mobile market with architectural changes nearly every year.
Honestly NVIDIA has too much money *not* to be making a new flagship chip every year. They need to increase their R&D as much as possible as quickly as possible to maintain their competitive advantage, but it’s hard to greatly increase team sizes in a way that doesn’t result in diminishing returns and other inefficiencies (never mind just replacing the people leaving because they made more than enough money to retire on their stock grants!)
Creating a process shrink of the flagship on a tick-tock cadence is the probably the easiest way to spend more money by creating a new independent team without affecting the rest of the roadmap. That doesn’t mean they are actually doing it, but I could not reasonably defend their failure to do so, it would highlight a catastrophic failure of execution to still only have one new flagship every two years while soon making nearly $100B revenue a year from that single chip…
Is there no updated CPU with Blackwell? Grace Next was initially scheduled for 2025 IIRC. Vera will be a 2026 part.
