NVIDIA discussion [2025]

[DavidGraham]

How is this any different to the GB200 NVL4?

The NVL4 is a special variant of GB200, the base model of GB200 has only 2 B200 GPUs. Now the base model of GB300 has 4 B300 GPUs.

Unwrapping the NVIDIA B200 and GB200 AI GPU Announcements

NVIDIA on Monday, at the 2024 GTC conference, unveiled the "Blackwell" B200 and GB200 AI GPUs. These are designed to offer an incredible 5X the AI inferencing performance gain over the current-gen "Hopper" H100, and come with four times the on-package memory. The B200 "Blackwell" is the largest...

www.techpowerup.com

On other news, Mediatek will have access to NVIDIA IP (NVLink and routers/modems) for it's ASIC business.

https://twitter.com/x/status/1902532405455720628

 

[vola]

one Rubin chip seems to have 112 SMs. blackwell was 80 SMs per chip so going from N4 to N3 gives a 40% increase in SM count at the same die size.

 

[Kaotik]

Der8auer commenting (bashing really) on NVIDIA representatives answers in recent interview

 

[Potato Head]

Der8auer commenting (bashing really) on NVIDIA representatives answers in recent interview

Why should I care?

 

[trinibwoy]

Der8auer commenting (bashing really) on NVIDIA representatives answers in recent interview

Is it true that there are no examples of the Nvidia cable melting? According to Der8auer the Nvidia adapter isn’t built any better than other cables so theoretically it should melt too.

The Nvidia guy has been in marketing for 20 years and still completely fumbled those questions about supply and availability. He should’ve just said that they’re working hard to increase supply instead of claiming lots of cards are being shipped and it’s retailers fault scalping is happening. That was tone deaf. Scalping doesn’t happen if they flood the zone with cards.

Why should I care?

5090 owners should probably care if it’s true that Nvidia cheaped out on the GPU connector. The missing shunt sounds like such an amateur mistake that it’s really mind boggling that Nvidia and all their manufacturing partners decided to repeat the same mistake after all the 4090 drama. Haven’t seen a good explanation for why they would do that. Either Der8auer is wrong or Nvidia is incredibly malicious and the AIBs are complicit. Why would you not fix a very small thing that would avoid major PR issues and RMA headaches and possibly lawsuits if someone’s house burns down? It makes no sense.

 

[Kaotik]

Is it true that there are no examples of the Nvidia cable melting? According to Der8auer the Nvidia adapter isn’t built any better than other cables so theoretically it should melt too.

No, NVIDIA cables can melt just like any other, happened already with RTX 40 series too

5090 owners should probably care if it’s true that Nvidia cheaped out on the GPU connector. The missing shunt sounds like such an amateur mistake that it’s really mind boggling that Nvidia and all their manufacturing partners decided to repeat the same mistake after all the 4090 drama. Haven’t seen a good explanation for why they would do that. Either Der8auer is wrong or Nvidia is incredibly malicious and the AIBs are complicit. Why would you not fix a very small thing that would avoid major PR issues and RMA headaches and possibly lawsuits if someone’s house burns down? It makes no sense.

They actually made it even weaker than RTX 40 series in the 5090 FE with just 1 shunt vs 2 (though it makes little difference since the 2 in RTX 40 weren't actually separate. RTX 30 was the last where it had some sense in it with 3 shunts, 2 lines per shunt separated all the way.

(from

)

 

[troyan]

They didnt cheap out. It is just a connector. Its like blaming a hair dryer company for not safe guarding when someone is throwing it into a full bathtube...

The problem is that this connector needs more space and bending can result in loosening the connection:

If I bend the RTX 40 Series adapter will it cause issues with the connection between my PSU and GPU?
The adapter has been proven to work in a wide variety of conditions. Please follow our compatibility diagram: plan a minimum of 1.4” or 36mm clearance above the top of the graphics card for cable bend and airflow clearance.

 

[Kaotik]

They didnt cheap out. It is just a connector. Its like blaming a hair dryer company for not safe guarding when someone is throwing it into a full bathtube...

The problem is that this connector needs more space and bending can result in loosening the connection:

They did cheap out on the PCB side pushing everything through 1 shunt resístor.
What you're describing as the problem is indeed a problem but not the only problem. The whole 12VHPWR / 12V-2x6 is a problem in itself with ridiculously low safety tolerances and indeed capability of pushing even all the damn 600 watts through one cable without card or PSU having any idea it's happening (which obviously doesn't fit into tolerances either and causes melting).

 

[homerdog]

They didnt cheap out. It is just a connector. Its like blaming a hair dryer company for not safe guarding when someone is throwing it into a full bathtube...

The problem is that this connector needs more space and bending can result in loosening the connection:

The issue on NVIDIA and partners' side is that the cards have no way to detect when something has gone wrong. "Gone wrong" meaning the load is not evenly distributed across the wires, for any reason.

 

[troyan]

It is not their job. It should be in the specification and mandatory.

Reminds me of OLEDs which can burn in. Nothing will ever pretend it. So in the end it is the end user who has to be careful.

 

[digitalwanderer]

They didnt cheap out. It is just a connector. Its like blaming a hair dryer company for not safe guarding when someone is throwing it into a full bathtube...

The problem is that this connector needs more space and bending can result in loosening the connection:

If they're just connecting all the power wires to the same bus when they connect to the card it's not exactly the cable's fault anymore. The 40 and 50 series have a design flaw in how they handle the incoming power lines.

 

[Albuquerque]

So, who in this thread has access to the PCI-SIG specification document for the 12VHPWR cable? Because we can solve this "they're doing it wrong" vs "no they're not" argument really fast if anyone can provide the actual reference implementation documents.

By the way, they're here: 12VHPWR Sideband Allocation and Requirements but only accessible if you're a paying member. A cursory glance has yielded no results in trying to find a "free" version of that document.

 

[Arun]

Yep, Jensen said they're sticking with the old naming for Blackwell and changing it for Rubin. Something about aligning the naming with NVLink topology.

It’s extremely confusing, there’s also the fact B300 NVL16 is really 16 *single-die* Blackwell Ultras (with 128GiB or 144GiB of HBM3e each - not 100% sure, probably the latter) while GB300 uses dual-die packages like B200/GB200.

I expect we’ll always need to carefully look at memory capacity/teraflops/etc to sanity check what a SKU actually is going forward… Server/HPC GPUs are rapidly catching up with notebook GPUs in terms of branding insanity.

 

[Granath]

It’s extremely confusing, there’s also the fact B300 NVL16 is really 16 *single-die* Blackwell Ultras (with 128GiB or 144GiB of HBM3e each - not 100% sure, probably the latter) while GB300 uses dual-die packages like B200/GB200.

I expect we’ll always need to carefully look at memory capacity/teraflops/etc to sanity check what a SKU actually is going forward… Server/HPC GPUs are rapidly catching up with notebook GPUs in terms of branding insanity.

Isn't it 4 GPU with 4 dies each?

 

[Kaotik]

It’s extremely confusing, there’s also the fact B300 NVL16 is really 16 *single-die* Blackwell Ultras (with 128GiB or 144GiB of HBM3e each - not 100% sure, probably the latter) while GB300 uses dual-die packages like B200/GB200.

I expect we’ll always need to carefully look at memory capacity/teraflops/etc to sanity check what a SKU actually is going forward… Server/HPC GPUs are rapidly catching up with notebook GPUs in terms of branding insanity.

Isn't it 4 GPU with 4 dies each?

First we need to agree on what's a GPU, even NVIDIA has several meanings for it. If you're referring to single packaging, 4 dies comes in Rubin Ultra.

But regardless, B300 NVL16 at least should have 32 GPU dies, since it's mentioned to use "Blackwell Ultra" which is 2 GPU dies in single package. Jensen mentioned something about not changing the naming midgen and how NVL72 should have already been 144 too, since each "GPU" has 2 dies.

 

[trinibwoy]

First we need to agree on what's a GPU, even NVIDIA has several meanings for it. If you're referring to single packaging, 4 dies comes in Rubin Ultra.

I think it should be software defined since physical packaging and interconnects will continue to evolve.

How many compute devices does a 4-die Rubin expose to CUDA? If it’s just one then that’s one GPU.

 

[DavidGraham]

After many setbacks and delays to it's AI initiatives, Apple is reportedly ordered 1 billion dollars worth of GB300 GPUs from NVIDIA.

https://twitter.com/x/status/1904460171180658949

 

[trinibwoy]

Does this mark a shift in strategy for Apple? What’s Siri running on now?

 

[Kaotik]

Does this mark a shift in strategy for Apple? What’s Siri running on now?

Training was done on Google chips, couldn't find solid info on if they're running it on them too or not.

 

[Albuquerque]

Yeah there's a number of online AIs out there which run on Google's Tensor chips. As a simple example, https://pi.ai is one such model which all inference nodes are powered by Tensor. I think the actual training is still done on NV hardware tho...