• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Nvidia delays Blackwell GPUs

erek

erek

8=D
2FA
Joined
Dec 19, 2005
Messages
17,892
“Product pushed back at least three months due to design problems”

“Google, Meta, and Microsoft are among those betting billions on Nvidia's GPUs amid an AI arms race. Google has ordered more than 400,000 GB200 chips, The Information reports, in a deal valued well north of $10 billion.

Meta also has a $10bn order, while Microsoft was expecting to have 55-65,000 GB200 GPUs ready for OpenAI by the first quarter. That now seems unlikely.

The production issue was discovered by manufacturer TSMC, and involves the processor die that connects two Blackwell GPUs on a GB200.

Nvidia is now adjusting the design and will need to run a new production test with TSMC before it can mass produce more.

It is also considering producing a version of the chip that only contains one Blackwell GPU to speed up delivery.

Due to the last-minute delay, TSMC will also have to leave production lines idle until the issue is fixed.”

1722696796137.png

Source: https://www.datacenterdynamics.com/...pacting-hyperscaler-data-center-plans-report/
 
erek said:
“Product pushed back at least three months due to design problems”

“Google, Meta, and Microsoft are among those betting billions on Nvidia's GPUs amid an AI arms race. Google has ordered more than 400,000 GB200 chips, The Information reports, in a deal valued well north of $10 billion.

Meta also has a $10bn order, while Microsoft was expecting to have 55-65,000 GB200 GPUs ready for OpenAI by the first quarter. That now seems unlikely.

The production issue was discovered by manufacturer TSMC, and involves the processor die that connects two Blackwell GPUs on a GB200.

Nvidia is now adjusting the design and will need to run a new production test with TSMC before it can mass produce more.

It is also considering producing a version of the chip that only contains one Blackwell GPU to speed up delivery.

Due to the last-minute delay, TSMC will also have to leave production lines idle until the issue is fixed.”

View attachment 670214
Source: https://www.datacenterdynamics.com/...pacting-hyperscaler-data-center-plans-report/
Click to expand...

Design Issues May Postpone Launch of NVIDIA's Advanced Blackwell AI Chips​

UPDATED by Nomad76 Today, 03:53 Updated: Today, 07:15 Discuss (35 Comments)
NVIDIA may face delays in releasing its newest artificial intelligence chips due to design issues, according to anonymous sources involved in chip and server hardware production cited by The Information. The delay could extend to three months or more, potentially affecting major customers such as Meta, Google, and Microsoft. An unnamed Microsoft employee and another source claim that NVIDIA has already informed Microsoft about delays affecting the most advanced models in the Blackwell AI chip series. As a result, significant shipments are not expected until the first quarter of 2025.
 
Wow this is pretty big news tbh. If AMD was smart (Which is hard for them sometimes) they would be shipping those MI300 cards as fast as they can.

This also means the 5000 series will now be pushed back even more....Plus it looks like the AI bubble is going bye bye.

Not good news at all.
 
Brackle said:
Wow this is pretty big news tbh. If AMD was smart (Which is hard for them sometimes) they would be shipping those MI300 cards as fast as they can.

This also means the 5000 series will now be pushed back even more....Plus it looks like the AI bubble is going bye bye.

Not good news at all.
Click to expand...
Bloomberg

1722697689068.png
 
Brackle said:
This also means the 5000 series will now be pushed back even more....Plus it looks like the AI bubble is going bye bye.
Click to expand...
I wonder if this might actually accelerate the release of the 5000 series. The cards running a single die wouldn't be impacted by this, so shipping those would provide some revenue while the multi-die versions get sorted out.
 
“Update 2:
SemiAnalysis's Dylan Patel reports in a message on Twitter (now known as X) that Blackwell supply will be considerably lower than anticipated in Q4 2024 and H1 2025. This shortage stems from TSMC's transition from CoWoS-S to CoWoS-L technology, required for NVIDIA's advanced Blackwell chips. Currently, TSMC's AP3 packaging facility is dedicated to CoWoS-S production, while initial CoWoS-L capacity is being installed in the AP6 facility.
OPBAK9K8BzMPgFhE_thm.jpg
Additionally, NVIDIA appears to be prioritizing production of GB200 NVL72 units over NVL36 units. The GB200 NVL36 configuration features 36 GPUs in a single rack with 18 individual GB200 compute nodes. In contrast, the NVL72 design incorporates 72 GPUs, either in a single rack with 18 double GB200 compute nodes or spread across two racks, each containing 18 single nodes.”

From TPU source linked above
 
The article says "design flaw" but that says TSMC ramp issues with their new advanced packaging facility. It was reported months back that TSMC was encountering unexpected setbacks and issues with their new CoWoS-L packaging process, and that is the process that makes the Nvidia interlink interposer that combines the two chips.
The CoWoS-L process allows TSMC to build interposers that are larger than the previous interposer reticle limit while packaging more communication lanes. Nvidia is 100% reliant on TSMC for these parts and is likely helping them save face, the design "flaw" is likely the new -L process has a slightly different practical limit than their design software indicated so they have to revise the design with the new specs the process is running at.
 
“As we've stated before, Hopper demand is very strong, broad Blackwell sampling has started, and production is on track to ramp in 2H. Beyond that, we don't comment on rumors”

NVIDIA Spokesperson
 
Does this just involve the data center high density AI chips, while consumer grade ones are unaffected or...? The summary kind of read like that.
 
Brackle said:
Wow this is pretty big news tbh. If AMD was smart (Which is hard for them sometimes) they would be shipping those MI300 cards as fast as they can.

This also means the 5000 series will now be pushed back even more....Plus it looks like the AI bubble is going bye bye.

Not good news at all.
Click to expand...
Well not good news for Nvidia anyway.
 
StoleMyOwnCar said:
Does this just involve the data center high density AI chips, while consumer grade ones are unaffected or...? The summary kind of read like that.
Click to expand...
Just the datacenter, none of the consumer parts are MCM so they don’t need the communications components that are causing the problem.
 
ChadD said:
Well not good news for Nvidia anyway.
Click to expand...
Not good for AI, the NVidia cards are still the go to product for mining, energy, physics research and simulation, CUDA does a lot AI is just the current hot topic.
 
Lakados said:
The article says "design flaw" but that says TSMC ramp issues with their new advanced packaging facility. It was reported months back that TSMC was encountering unexpected setbacks and issues with their new CoWoS-L packaging process, and that is the process that makes the Nvidia interlink interposer that combines the two chips.
The CoWoS-L process allows TSMC to build interposers that are larger than the previous interposer reticle limit while packaging more communication lanes. Nvidia is 100% reliant on TSMC for these parts and is likely helping them save face, the design "flaw" is likely the new -L process has a slightly different practical limit than their design software indicated so they have to revise the design with the new specs the process is running at.
Click to expand...
In the industry is there insurance for situations like this?

How do they effectively mitigate such a profound impact & loss?

I know there’s rumors and whispers of other segments in the energy industry that will charge tax payers for the companies loss of profits for other reasons (them saving money)

https://lailluminator.com/2024/07/2...rced-to-pay-utility-company-for-lost-profits/

https://www.businessreport.com/arti...-want-customers-to-pay-for-lost-profits?amp=1

Just curious because this would seem to be a big deal and potentially billions of losses for nvidia?

What’s the reality
 
erek said:
In the industry is there insurance for situations like this?

How do they effectively mitigate such a profound impact & loss?

I know there’s rumors and whispers of other segments in the energy industry that will charge tax payers for the companies loss of profits for other reasons (them saving money)

https://lailluminator.com/2024/07/2...rced-to-pay-utility-company-for-lost-profits/

https://www.businessreport.com/arti...-want-customers-to-pay-for-lost-profits?amp=1

Just curious because this would seem to be a big deal and potentially billions of losses for nvidia?

What’s the reality
Click to expand...
That Louisiana thing I thought was a joke, like an Onion article spread out of control….
Normally energy suppliers just raise rates to cover the gap. People use less so they produce less so demand remains the same but the supply is lower therefore price goes up and margins are maintained.

But Nvidia might take a hit on their margin but the orders aren’t going anywhere. An extra 3 months on a project that will take multiple months just to configure and get online isn’t overly significant in the grand scheme of things. Companies buying these don’t have an abundance of options anyways, it’s not like anybody else builds CUDA compatible hardware.
 
erek said:
“Update 2:
SemiAnalysis's Dylan Patel reports in a message on Twitter (now known as X) that Blackwell supply will be considerably lower than anticipated in Q4 2024 and H1 2025. This shortage stems from TSMC's transition from CoWoS-S to CoWoS-L technology, required for NVIDIA's advanced Blackwell chips. Currently, TSMC's AP3 packaging facility is dedicated to CoWoS-S production, while initial CoWoS-L capacity is being installed in the AP6 facility.
View attachment 670267
Additionally, NVIDIA appears to be prioritizing production of GB200 NVL72 units over NVL36 units. The GB200 NVL36 configuration features 36 GPUs in a single rack with 18 individual GB200 compute nodes. In contrast, the NVL72 design incorporates 72 GPUs, either in a single rack with 18 double GB200 compute nodes or spread across two racks, each containing 18 single nodes.”

From TPU source linked above
Click to expand...
That might be the important quote right there: "This is partially offset by [redacted] which will have a decent revenue contribution and [redacted] is driving the increase in 3Q revenue"

I'm not holding my breath, but I am hoping that [redacted] is the gaming product. Of course, in order to have any meaningful impact on 3Q financials, that product would have to launch, like, today.
 
Video cards will be using Single die chips for the GPU's, so video card releases I doubt are affected.

The bloomberg post even says "AI chip is delayed".
 
GoodBoy said:
Video cards will be using Single die chips for the GPU's, so video card releases I doubt are affected.

The bloomberg post even says "AI chip is delayed".
Click to expand...
Well AI is the new buzz word. It makes for a much better and impactful headline than “gaming chip delayed”
 
(unlike the B100-B200) Nvidia did not even hint a possible release date or took order for the L40 replacement and the gaming card, making it hard to talk about being delayed for a trade headline. Maybe it will (maybe the issue is larger than what they say how just fab timing will push everything back or something and we will never know, if they launch in say March 2025 were they delayed by this or not ? Who would know)
 
Last edited:
Jensen needs his Ego taken down a notch. I'm okay with this.
 
OFaceSIG said:
Jensen needs his Ego taken down a notch. I'm okay with this.
Click to expand...
seems GPUs now can also get CTE (akin or analogous to Chronic traumatic encephalopathy) problems (i jest) , pretty interesting

"The bridge die placement requires very high levels of accuracy, especially when it comes to the bridges between the two main compute dies as these are critical for supporting the 10 TB/s chip-to-chip interconnect. A major design issue rumored is related to the bridge dies. These bridges need to be redesigned. Also rumored is a redesign of the top few global routing metal layers and bump out of the Blackwell die. This is a primary cause of the multi-month delay.

There has also been the issue of TSMC not having enough CoWoS-L capacity in aggregate. TSMC built up a lot of CoWoS-S capacity over the last couple years with Nvidia taking the lion’s share. Now with Nvidia quickly moving their demand to CoWoS-L, TSMC is both building a new fab, AP6, for CoWoS-L and converting existing CoWoS-S capacity at AP3. TSMC needs to convert the old CoWoS-S capacity as otherwise it would be underutilized and the ramp of CoWoS-L would be even slower. This conversion process makes the ramp very lumpy in nature.

Combine these two issues and it’s clear that TSMC will not be able to supply enough Blackwell chips as Nvidia would like. Consequently, Nvidia is focusing what capacity they have almost entirely on GB200 NVL 36x2 and NVL72 rack scale systems. HGX form-factors with the B100 and B200 are effectively now being cancelled outside of some initial lower volumes."
 

“NVIDIA's New B200A Targets OEM Customers; High-End GPU Shipments Expected to Grow 55% in 2025​

PRESS RELEASE by TheLostSwede Today, 06:57 Discuss (0 Comments)
Despite recent rumors speculating on NVIDIA's supposed cancellation of the B100 in favor of the B200A, TrendForce reports that NVIDIA is still on track to launch both the B100 and B200 in the 2H24 as it aims to target CSP customers. Additionally, a scaled-down B200A is planned for other enterprise clients, focusing on edge AI applications.“
 
Armenius said:
It's a TSMC issue, not a NVIDIA issue.
Click to expand...
With that being said, and TSMC a third party vendor for Nvidia

You’d wonder if they assessed liability insurance with tsmc or risk accepted the gap?

Would be interesting to understand their third party engagement processes

—-
 
Lakados said:
The article says "design flaw" but that says TSMC ramp issues with their new advanced packaging facility. It was reported months back that TSMC was encountering unexpected setbacks and issues with their new CoWoS-L packaging process, and that is the process that makes the Nvidia interlink interposer that combines the two chips.
The CoWoS-L process allows TSMC to build interposers that are larger than the previous interposer reticle limit while packaging more communication lanes. Nvidia is 100% reliant on TSMC for these parts and is likely helping them save face, the design "flaw" is likely the new -L process has a slightly different practical limit than their design software indicated so they have to revise the design with the new specs the process is running at.
Click to expand...

Nah, no company is taking a fall for the other. If they say it's a issue with Nvidia design than it most likely is if Nvidia does not deny it. Neither one wants bad press.
 
Gideon said:
Nah, no company is taking a fall for the other. If they say it's an issue with Nvidia design than it most likely is if Nvidia does not deny it. Neither one wants bad press.
Click to expand...
Well it turns out the “design issue” is the TSMC CoWoS-L interposer has a small organic component, and it heats and cools at a different rate than the rest of the interposer. It causes it to warp over time. So they need to redesign their circuitry so they can minimize the warping. So TSMC issue…
 
GoodBoy said:
Video cards will be using Single die chips for the GPU's, so video card releases I doubt are affected.

The bloomberg post even says "AI chip is delayed".
Click to expand...

Considering how many financial media journalists still call the company “Nuh-vidia”, I wouldn’t rely on that as an accurate guideline regarding which specific products are impacted by the delay. Bloomberg calling a GPU an “AI chip” is basically like your mom calling every console you ever owned a “Nintendo” regardless of the brand.
 
Armenius said:
It's a TSMC issue, not a NVIDIA issue.
Click to expand...
Not all problems are production problems. From what I've read production has been halted. However production can be halted because of a design flaw just as quickly, as a production flaw.
 
Lakados said:
Well it turns out the “design issue” is the TSMC CoWoS-L interposer has a small organic component, and it heats and cools at a different rate than the rest of the interposer. It causes it to warp over time. So they need to redesign their circuitry so they can minimize the warping. So TSMC issue…
Click to expand...

Unless Nvidia specified that component. But generally, if another company is causing a delay in product they will call them out, to show investors why there is a delay.
 
Gideon said:
Unless Nvidia specified that component. But generally, if another company is causing a delay in product they will call them out, to show investors why there is a delay.
Click to expand...
It's the thing that makes the -L process different from the regular one.

CoWoS-L: This uses local silicon interconnect (LSI) along with RDL interposer together forming a reconstituted interposer (RI). In addition to the RDL interposer, it also the preserves the attractive feature of CoWoS-S in the form of through silicon vias (TSVs). This also mitigates the yield issues arising due to the use of large silicon interposer in CoWoS-S. In some implementations, it may also use through insulator vias (TIVs) instead of TSVs to minimize the insertion loss.

cowos-L-package.png
 
GoldenTiger said:
They do? I've never felt that way as a customer, and they must not be too bad to their partners with the scale they work with them on.
Click to expand...
Talk to EVGA about how Nvidia treats partners.

Let's not bring up how they treated customers regarding the 970......Yeah I know we both got a check from Nvidia on that one.

But, what I will say I do like Nvidia products and I always pick vendor based on their warranty.
 
GoldenTiger said:
They do? I've never felt that way as a customer, and they must not be too bad to their partners with the scale they work with them on.
Click to expand...
Companies will do business with the winner. They want to make money. I'm not saying they don't make money, don't make excellent products. Engineering, etc, Nvidia is top notch. But their business practices are garbage, have been garbage. Unless you have scale you don't get to do business with Nvidia anymore. Look at all the vendors that were either pushed out of marketing Nvidia products or chose not to because of their predatory practices.
 
  • Like
Reactions: kac77
like this
Brackle said:
Talk to EVGA about how Nvidia treats partners.

Let's not bring up how they treated customers regarding the 970......Yeah I know we both got a check from Nvidia on that one.

But, what I will say I do like Nvidia products and I always pick vendor based on their warranty.
Click to expand...
EVGA went out of the videocard business due to their own issues and financials.

I got 20% back and $30 on each of my GTX 970 purchases. I was ultimately satisfied, but disappointed they lied about the specs upfront.

I too like Nvidia products, and have always picked vendor based on warranty. Evga had a great one that was ultimately too costly alongside their step up program for them. :)
 
erek said:
With that being said, and TSMC a third party vendor for Nvidia

You’d wonder if they assessed liability insurance with tsmc or risk accepted the gap?

Would be interesting to understand their third party engagement processes

—-
Click to expand...
It’s a managed risk sort of thing, but there’s so much money and time
Involved that nobody wants to piss anybody off so TSMC is generally good for trying to make things right.
 
You must log in or register to reply here.
Back
Top