News Spitballing Nvidia's GB202 GPU die manufacturing costs — die could cost as little as $290 to make

[nogaard777]

I would assume all the imperfect dies are being sold in China as 5090Ds so they can still charge full price for them. Not to mention there's been pictures popping up of entire PALLETS of 5090Ds coming out of China.

Whether all the dies are going there or they're simply focusing on AI enterprise GPU I'm glad I'm not looking to buy one because this was a BS paper launch.

 

[MobileJAD]

All that r and d plus they threw everything at it inculding the kitchen sink to barely get 30% atypical delta gains means the margins are atypically even higher. The already recuperated the r and d cost 1000 fold over!
How do you guarantee the public pay for such atypical margin ask? You create fud around scarcity and throttle all stock to maximize the maximum profits per silicon sold. Nvidia tactics 101!

Does Nvidia even have to do anything special for RTX 5090 scarcity to happen? I would have thought that Nvidia would have put all of its effort into supplying AI datacenters before worrying about gamer class video cards.
I have been on Nvidia most of my gaming life, but now AMD is looking like the better deal.

 

[aberkae]

Does Nvidia even have to do anything special for RTX 5090 scarcity to happen? I would have thought that Nvidia would have put all of its effort into supplying AI datacenters before worrying about gamer class video cards.
I have been on Nvidia most of my gaming life, but now AMD is looking like the better deal.

True but why not be open about it instead of having getting everyone hyped for a vapor launch. This is all for the stock to get pumped.
Once the ai datacenters get saturated with supply and demand starts to plautau guess who they come back begging for their business? Fingers crossed their strategy doesn't backfire.
One thing we learned this week is that their stock is very vulnerable and volatile.

 

[bit_user]

True but why not be open about it instead of having getting everyone hyped for a vapor launch.

Yeah, I think they do owe the public an explanation for why the launch inventories were so low and what they're doing about it.

Worst case: it could be due to the PCIe 5.0 problems. Maybe they're having to do extra validation and divert some units which don't pass. In that case, maybe they don't want to explain it for fear of looking incompetent.

If you haven't been following the ramp of the datacenter Blackwell chips, they've been encountering one problem after another.

https://www.tomshardware.com/tech-i...e-now-fixed-blackwell-chips-are-in-production

https://www.tomshardware.com/tech-i...lume-ramp-up-of-blackwell-machines-trendforce
​

If they have too many more problems with new product introductions, I think investors might start to get worried.

 

[Gaidax]

I would assume all the imperfect dies are being sold in China as 5090Ds so they can still charge full price for them. Not to mention there's been pictures popping up of entire PALLETS of 5090Ds coming out of China.

Whether all the dies are going there or they're simply focusing on AI enterprise GPU I'm glad I'm not looking to buy one because this was a BS paper launch.

No, all the imperfect dies are sold to us.

Keep in mind that what we get in 5090 is not a full B202. They will probably save full B202 for professional cards.

 

[Gaidax]

Yeah, I think they do owe the public an explanation for why the launch inventories were so low and what they're doing about it.

I mean, does the explanation matter? You will get a bunch of corpo PR speak, explanation is the least anyone needs here.

It's really simple, they need to fix the whatever mess there is that caused this and start pumping out them GPUs. And I bet they would if they could too, so something obviously went wrong there. Fix that and can keep explanations to themselves.

 

[bit_user]

No, all the imperfect dies are sold to us.

Keep in mind that what we get in 5090 is not a full B202. They will probably save full B202 for professional cards.

Have they released specs on the 5090D? If it's restricted to PCIe 4.0, then maybe that's why so many are being diverted. Or, if it has 384-bit memory interface, then any problem in the memory controller would mean that GB202 can't go into workstation or normal RTX 5090 cards. It used to be that Nvidia disabled a 32-bit memory channel on the top-end mainstream card - maybe with good reason?

 

[bit_user]

I mean, does the explanation matter?

Yes, if it included some information about when the situation would be remedied and how much additional product should be moving into the channel. That would be a best case, not necessarily what I expect. Still, it would be nice to give their customers some idea of when the stock situation should improve.

 

[Gaidax]

Have they released specs on the 5090D? If it's restricted to PCIe 4.0, then maybe that's why so many are being diverted. Or, if it has 384-bit memory interface, then any problem in the memory controller would mean that GB202 can't go into workstation or consumer cards. It used to be that Nvidia disabled a 32-bit memory channel on the top-end mainstream card - maybe with good reason?

From what I understand, memory and CUDA cores are the same, only AI processing is hard limited in some way.

 

[Gaidax]

Yes, if it included some information about when the situation would be remedied and how much additional product should be moving into the channel. That would be a best case, not necessarily what I expect. Still, it would be nice to give their customers some idea of when the stock situation should improve.

Yes, they are already rushing in to give you all the details on the guts of their logistics and operation.

As I said, even if they will give you an explanation, it will be nothing more than a generic PR speak about difficulties in the supply chain/logistics and such. Will it help? Not really.

 

[bit_user]

From what I understand, memory and CUDA cores are the same, only AI processing is hard limited in some way.

Edit: sorry, I totally missed your point. I'm leaving my original reply, in case someone finds it informative.
---
The memory controller is hard-wired logic.

As for the rest, here are the fixed-function units, to the best of my knowledge (I might be leaving some out):

• TMUs - texture fetch, decompression, and interpolation.
• ROPs - polygon rasterization and blending.
• RT - bounding volume hierarchy (BVH) and polygon ray intersection tests.
• Tessellation - subdivision of parametric patches into polygons.

I didn't list Tensor cores, because those are sort of like just another arithmetic pipeline integrated into the Streaming Multiprocessors (SMs).

You can see the memory controllers labelled in this annotated GB202 die shot:

Source:

https://twitter.com/x/status/1883153126011892140

View: https://x.com/Kurnalsalts/status/1883153126011892140/photo/1

​

You can see the GDDR7 controllers along the periphery of the die. I think the Fixed Function units I listed above should be in the blocks labelled "Raster Engine / 3D FF".

Note that Toms also covered this story, but didn't include a very high-res image or much other info. So, I just linked to the source.

Also, props to Videocardz, who covered this same topic with a bunch of additional context:

https://videocardz.com/newz/nvidia-...die-shot-revealed-24756-cores-and-512-bit-bus
​

 

[aberkae]

Derbauer's take on it.

View: https://youtu.be/3zdBHgquvDg?si=8bY49j265gDtvrhx

Pretty much matches exactly what I said.

 

[Gaidax]

Well, I mean it does not take a genius to figure out there was some sort of supply chain failure and when there is a huge demand and barely any supply prices skyrocket. Nobody cares about MSRP, that's just a guideline not reality.

I bet it will take a year before 5090s will even start reasonably moving towards MSRP, if at all. 5080 will go there faster, as it's somewhat of a disappointment overall and there will be much more of them anyway.

 

[bit_user]

Well, I mean it does not take a genius to figure out there was some sort of supply chain failure and when there is a huge demand and barely any supply prices skyrocket. Nobody cares about MSRP, that's just a guideline not reality.

MSRP tells us something about the prices where Nvidia feels its AIB partners should be profitable. It sets a sort of price floor.