[GOSSIP] Nvidia rushed TSMC through the Blackwell production process, causing failure during some common data center usage

Marees

Supreme [H]ardness
Joined
Sep 28, 2018
Messages
4,945

Nvidia and TSMC’s Lucrative AI Alliance Shows Signs of Stress​


Nvidia’s Blackwell chip troubles, revealed in August, hurt relations with manufacturing partner TSMC, leading Nvidia to consider having Samsung make its latest gaming GPUs (non-AI related chips), The Information reports. Nvidia engineers testing initial Blackwell chips found they failed during some common data center usage, possibly from a design flaw.

But one early media report blamed TSMC’s advanced packaging process. Unnamed TSMC engineers said Nvidia rushed TSMC through the Blackwell production process, causing strain, as there was less time to work kinks out.

Note: Nvidia has used Samsung as a manufacturing partner for gaming GPUs before, so it's not that strange to turn to them now.

https://www.theinformation.com/articles/nvidia-and-tsmcs-lucrative-ai-alliance-shows-signs-of-stress

https://x.com/dnystedt/status/1847063385332240540
 
Related threads. Hopefully they are based on confirmed news

"It appears like 2024 will go down as the second consecutive year without any new GPU generation launch from either NVIDIA or AMD. Kopite7kimi, a reliable source with NVIDIA leaks, says that the GeForce RTX 50-series "Blackwell" generation won't see a debut before the 2025 International CES (January 2025). It was earlier expected that the company would launch at least its top two SKUs—the RTX 5090 and RTX 5080—toward the end of 2024, and ramp the series up from 2025. There is no explanation behind this "delay." Like everyone else, NVIDIA could be rationing its foundry allocation of the 3 nm wafers from TSMC for its high-margin "Blackwell" AI GPUs. The company now makes over five times the revenue from selling AI GPUs than it does from gaming GPUs, so this development should come as little surprise."

Source: https://www.techpowerup.com/324785/nvidia-geforce-blackwell-wont-arrive-before-january-2025

“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/
 
You just put sand in a hydraulic press and make chips right? Samsung is Korean I rather trust their chips.
 
If they heavily test and spend billion-month fixing it, that not really taking a shortcut, maybe it is Nvidia and TSMC trying to take one at best but not pulling it of, it is not like the alchemist magic they are trying to make is not more and more complicated, Blackwell, 4np, etc... was not made faster than previous gen.
 
Last edited:
My question for the unnamed TSMC engineer?
Was it Nvidia pushing things forward?
Or
Was it some TSMC manager above them that had already messed up with the LSI warping problem and they didn’t want to take another L so they were pushing their team harder in an attempt to look better and they were telling the people below them the pressure was coming from Nvidia, because the previous fuckup had already pushed them back 3 months.
 
If they heavily test and spend billion-month fixing it, that not really taking a shortcut.
Right the early reports were saying the new Interposer type TSMC designed heated and cooled differently which caused warping which would break solder connections and cause the unit to fail.

That oversight required Nvidia to redesign their layout to better spread out hotspots and move the LSI component of the Interposer to match and no longer warp.

Sure was it “rushed” because TSMC wasn’t aware their brand new packaging process had a quirk they missed? maybe.
But that’s what the testing is for, crap comes up in testing all the time, this is just a very problematic issue to encounter that late into the game.

Encountering something like that after rollout would be devastating, encountering it during load testing is borderline normal.

The bigger question for TSMC would be how much testing had they done on their new Interposers before offering it as a solution to Nvidia to integrate into their new chips? Because this sounds like a materials problem they should have caught early.
 
My experience with large corporations: never time to do it right, but always time and money to do it over after meeting the previous deadline with less than optimal product.
Mostly saw this in aerospace industry.

Kid
 
OH my gosh wait so this means that EVERY SINGLE CHIP produced is flawed?
And the entire 50-series is actually fatally flawed?

I'm sure they'll downplay it ... But I bet this is much much worse than they say heheh.
OH yes, this might be a really bad thing for Nvidia for years to come.

Hopefully AMD takes the time to do things right.
 
OH my gosh wait so this means that EVERY SINGLE CHIP produced is flawed?
And the entire 50-series is actually fatally flawed?

I'm sure they'll downplay it ... But I bet this is much much worse than they say heheh.
OH yes, this might be a really bad thing for Nvidia for years to come.

Hopefully AMD takes the time to do things right.
No.
 
OH my gosh wait so this means that EVERY SINGLE CHIP produced is flawed?
And the entire 50-series is actually fatally flawed?

I'm sure they'll downplay it ... But I bet this is much much worse than they say heheh.
OH yes, this might be a really bad thing for Nvidia for years to come.

Hopefully AMD takes the time to do things right.
1729367005529.png
 
OH my gosh wait so this means that EVERY SINGLE CHIP produced is flawed?
And the entire 50-series is actually fatally flawed?

I'm sure they'll downplay it ... But I bet this is much much worse than they say heheh.
OH yes, this might be a really bad thing for Nvidia for years to come.

Hopefully AMD takes the time to do things right.
This is only data centre (packaging defect)

Doesn't affect gaming chips

NvidIa I believe didn't release these chips but delayed the delivery
 
More Coverage:

Nvidia might be in the blame game with TSMC over Blackwell chip failures, but that doesn't mean it will go back to Samsung​


Tech site The Information (via Business Korea) reports that "during testing of the Blackwell chips in the weeks after Huang's announcement, Nvidia engineers discovered that the chips failed in the types of high-voltage environments that are common in data centers". This is supposedly according to "two people with direct knowledge of the problem".

The two sides are apparently blaming each other for this, with TSMC saying Nvidia rushed production and Nvidia saying the defects are due to TSMC's packaging tech. Whoever's to blame, the result seems to be tension in a three decade-long relationship between the two companies.

NVIDIA and TSMC have maintained a partnership for nearly 30 years, a relationship underpinned by a unique network within the Greater China region. NVIDIA's CEO Huang is a Taiwanese-American, and TSMC is a symbol of Taiwan. It is a well-known anecdote that Huang once remarked, "Without TSMC, there would be no NVIDIA today."

https://www.pcgamer.com/hardware/gr...-that-doesnt-mean-it-will-go-back-to-samsung/
 
100% my fault — Jensen, confesses.

“"We had a design flaw in Blackwell," Huang said. "It was functional, but the design flaw caused the yield to be low. It was 100% Nvidia's fault."
According to media reports, opens new tab, the delay in production had caused tensions between Nvidia and TSMC but Huang dismissed that as "fake news".
"In order to make a Blackwell computer work, seven different types of chips were designed from scratch and had to be ramped into production at the same time," he said.”

Source: https://www.reuters.com/technology/...kwell-ai-chips-now-fixed-ceo-says-2024-10-23/
 
Any impact on NVidia's stock price? On Jensen's personal worth?

Any speculation as to what the AIB vendors will do besides wait?

Any impact on NVidia's sales to corporates for AI work?
With hardforum’s SEO it’s certainly plausible that your very sentiment could have a small role in the influence of said stock, Hmm

1729713381587.png
 
With hardforum’s SEO it’s certainly plausible that your very sentiment could have a small role in the influence of said stock, Hmm
Maybe, maybe not. But I don't own any NVidia stock, because I don't own any stocks. :rolleyes:
 
Back
Top