Microsoft to Unveil Custom AI Chips to Fight NVIDIA's Monopoly

[AleksandarK]

According to sources close to The Information, Microsoft is supposed to unveil details about its upcoming custom silicon design for accelerating AI workloads. Allegedly, the incoming chip announcement is scheduled for November during Microsoft's annual Ignite conference. Held in Seattle from November 14 to 17, the conference is supposed to show all of the work that the company has been doing in the field of AI. The alleged launch of an AI chip will undoubtedly take center stage in the announcement, as the demand for AI accelerators has been so great that companies can't get their hands on GPUs. The sector is mainly dominated by NVIDIA, with its H100 and A100 GPUs powering most of the AI infrastructure worldwide.

With the launch of a custom AI chip codenamed Athena, Microsoft hopes to match or beat the performance of NVIDIA's offerings and reduce the cost of AI infrastructure. As the price of H100 GPU can get up to 30,000 US Dollars, building a data center filled with H100s can cost hundreds of millions. The cost could be winded down using homemade chips, and Microsoft could be less dependent on NVIDIA to provide the backbone of AI servers needed in the coming years. Nevertheless, we are excited to see what the company has prepared, and we will report on the Microsoft Ignite announcement in November.



View at TechPowerUp Main Site | Source

 

[Dimitriman]

Oh how I would love to see Jensen's bubble burst.

 

[Bwaze]

Oh how I would love to see Jensen's bubble burst.

Oh how I'd love to see AI bubble burst! As long as there will be such hype about this pattern recognition "artificial inteligence" I don't think we'll see any progress on anything else, right now everything and everybody is focusing on it, how they can cash in from this miracle tool that will make human creativity obsolete...

 

[Dimitriman]

Oh how I'd love to see AI bubble burst! As long as there will be such hype about this pattern recognition "artificial inteligence" I don't think we'll see any progress on anything else, right now everything and everybody is focusing on it, how they can cash in from this miracle tool that will make human creativity obsolete...

Ai is unfortunately not going away, because it is the main engine powering ads and ad delivery mediums (social media, streaming, VR, etc). This is the core reason Ai exists, any other justification for Ai existence is so irrelevant that it is borderline misinformation.
One can only hope that GPU's become obsolete for for Ai much like what happened to them when Bitcoin ASICS came to market and GPU Bitcoin farming died.

 

[Deleted member 185088]

I feel big companies will push for custom silicon.

Oh how I'd love to see AI bubble burst! As long as there will be such hype about this pattern recognition "artificial inteligence" I don't think we'll see any progress on anything else, right now everything and everybody is focusing on it, how they can cash in from this miracle tool that will make human creativity obsolete...

It is sad really, we have better technologies yet we still waste more than 40h a week working instead of living. AI is another missed opportunity, it can really speed up a lot of things, e.g. easy part of code could be written by AI and developers work less.

 

[Space Lynx]

who is going to make these chips? cause I am pretty sure Nvidia can afford to just buy out the entire node now like Apple does for TSMC production lines. so that leaves Intel... cause globalfoundry isn't advanced enough?

 

[DeathtoGnomes]

The custom chip market wont stop here, there will be more players entering the hot tub eventually.

 

[mechtech]

"As the price of H100 GPU can get up to 30,000 US Dollars"

wow

See this happening everywhere for everything since and post covid. The place I work for has general contractors doing work for us charging out their labourers for $200/hr, and I would bet their labours don't even see 20% of that. That pricing is starting a change for a lot of companies to start hiring and doing things internally.

If prices are exuberantly high it's only a matter of time before companies and people start to do more on their own.

 

[MrDweezil]

And they're going to compete with Nvidia with Microsoft's extensive experience in chip design? Makes sense for competitors to enter this market, but Microsoft hardly seems equipped to do it.

 

[ymdhis]

This is bitcoin all over again. First it was cpu only, then gpus, then custom asics.

 

[AnotherReader]

who is going to make these chips? cause I am pretty sure Nvidia can afford to just buy out the entire node now like Apple does for TSMC production lines. so that leaves Intel... cause globalfoundry isn't advanced enough?

Why would Nvidia buy out the entire node? They will only buy enough for their needs. Besides, Nvidia is limited by COWOS, also known as interposers. TSMC was expected to scale N5 up to 150,000 wafers per month in late 2022. The process used for Nvidia is a variant of N5 so its capacity is presumably included in that. With H100's die size of 814 mm^2, and a defect rate of 0.09 per cm^2 for N5 (probably lower by now), the yield is 50% for perfect dies. Since Nvidia doesn't use perfect dies for the H100, the yield is even higher, but let's assume that they only get 29 H100 per N4 wafer. With about $10 billion of data center revenue per quarter, only 2.55% of TSMC's wafer output would yield enough H100 GPUs for that $10 billion.

Besides, Microsoft is much richer than Nvidia and could afford to buy it outright if they wanted to even though that is rather unlikely. Microsoft's net income is about 4 times higher than Nvidia's and revenue is also about 4 times greater. Microsoft is no AMD; it's vastly richer than any of the hardware companies bar Apple.

 

[TumbleGeorge]

cause globalfoundry isn't advanced enough?

Why do you think that? Because they not producing bleeding edge of fake number of nanometers?

 

[AsRock]

Oh how I would love to see Jensen's bubble burst.

As muchh as i would like that, look who's saying it. Microsoft the master of monopoly's and deception.

 

[Bjprice]

Microsoft:

there’s money to be made <insert tech space>

Builds least capable version possible and it fails. Iterate minimally. Keep costs down.

Lobby government claiming the monopolistic habits of competitors causes failure.

Literally use your monopoly to force people to use your stuff.

Microsoft is much closer to hedge fund than tech company. They do the tech stuff because they have to. They’d much, much rather just acquire and exploit than innovate.

 

[thesmokingman]

I've been saying this for a while now. First it was Tesla, now MS, next stop Google. You can book it.

 

[Space Lynx]

Why would Nvidia buy out the entire node? They will only buy enough for their needs.

their needs have no limits, everyone and their mom is lining around the block for their AI cards and getting tired of waiting.

 

[geniekid]

their needs have no limits, everyone and their mom is lining around the block for their AI cards and getting tired of waiting.

Seriously. My mom won't shut up about it.

 

[Space Lynx]

Seriously. My mom won't shut up about it.

you would be surprised how her small cooking business will take off with AI bro

 

[ARF]

Why would Nvidia buy out the entire node? They will only buy enough for their needs. Besides, Nvidia is limited by COWOS, also known as interposers. TSMC was expected to scale N5 up to 150,000 wafers per month in late 2022. The process used for Nvidia is a variant of N5 so its capacity is presumably included in that. With H100's die size of 814 mm^2, and a defect rate of 0.09 per cm^2 for N5 (probably lower by now), the yield is 50% for perfect dies. Since Nvidia doesn't use perfect dies for the H100, the yield is even higher, but let's assume that they only get 29 H100 per N4 wafer. With about $10 billion of data center revenue per quarter, only 2.55% of TSMC's wafer output would yield enough H100 GPUs for that $10 billion.

Besides, Microsoft is much richer than Nvidia and could afford to buy it outright if they wanted to even though that is rather unlikely. Microsoft's net income is about 4 times higher than Nvidia's and revenue is also about 4 times greater. Microsoft is no AMD; it's vastly richer than any of the hardware companies bar Apple.

And yet, AMD can't afford to make the Navi 33 on anything better than N6 that can't even show its existence in the overall Steam hardware survey. What a joke !

Seriously though, where or what for do all these up to 150,000 wafers monthly go?

 

[AnotherReader]

And yet, AMD can't afford to make the Navi 33 on anything better than N6 that can't even show its existence in the overall Steam hardware survey. What a joke !

Seriously though, where or what for do all these up to 150,000 wafers monthly go?

There's a big computing world outside Nvidia and AMD. Apple, for one, until recently, was using N5 exclusively. They sold 225.3 million iPhones in 2022. Assuming that all of them used the A15, with a die size of 107.68 mm^2 and assuming a defect rate of 0.09 cm^2, this yields 489 perfect dies per wafer. Of course, this doesn't account for any redundancy which no one outside Apple would know. In any case, this means that Apple's iPhones alone consumed about four months of N5 capacity in 2022. Remember that TSMC was at 120k N5 wafers per month for much of 2022. I haven't included Mac and iPads in those numbers. Including those would probably make it closer to 5.5 months of TSMC's output.

Microsoft:

there’s money to be made <insert tech space>

Builds least capable version possible and it fails. Iterate minimally. Keep costs down.

Lobby government claiming the monopolistic habits of competitors causes failure.

Literally use your monopoly to force people to use your stuff.

Microsoft is much closer to hedge fund than tech company. They do the tech stuff because they have to. They’d much, much rather just acquire and exploit than innovate.

The only thing missing from your post is M$. Microsoft had $56 billion of revenue in its last quarter, and that money didn't come from investments in other companies. Please read the linked article before replying.

 

[MxPhenom 216]

Time to buy Microsoft stock

 

[HisDivineOrder]

Custom AI chips are the future. General purpose AI chips like Jensen's peddling will go away and I don't see him making the changes that deny him the option to use them as GPU's in case of emergency. He doesn't want to commit to just AI like he should if he truly believed it was the promised land.

In the end, his wariness will lead companies away from his overpriced general process units and force him back to gaming yet again. It's happened before and it'll happen again. He's always on the cusp of greatness and then has to "just" settle for a few billion by the end.

I wish he'd set his sights on AI, spin off the gaming GPU business, and leave it to its fate. Then the AI bubble could burst and he could watch the gaming industry continue to accelerate away from him, his greed having finally beaten him into submission after his long hard fought battle against reality.

 

[thesmokingman]

Custom AI chips are the future. General purpose AI chips like Jensen's peddling will go away and I don't see him making the changes that deny him the option to use them as GPU's in case of emergency. He doesn't want to commit to just AI like he should if he truly believed it was the promised land.

In the end, his wariness will lead companies away from his overpriced general process units and force him back to gaming yet again. It's happened before and it'll happen again. He's always on the cusp of greatness and then has to "just" settle for a few billion by the end.

I wish he'd set his sights on AI, spin off the gaming GPU business, and leave it to its fate. Then the AI bubble could burst and he could watch the gaming industry continue to accelerate away from him, his greed having finally beaten him into submission after his long hard fought battle against reality.

Concur. The NV marketing machine has given casuals the impression that NV's gpus are somehow gifted at AI which is far from the truth. It's just the best of the crap that is out there at an insane mark up, so insane in fact that trillion dollar companies are bringing it in house.

In regards to spinning off, I think it's too late now. Compute consumers have been gouged for years leading up to this. H100 is 30% improved over A100 yet the price went from $10K to $40K, smh. The market will not forget this.

 

[ARF]

There's a big computing world outside Nvidia and AMD. Apple, for one, until recently, was using N5 exclusively. They sold 225.3 million iPhones in 2022. Assuming that all of them used the A15, with a die size of 107.68 mm^2 and assuming a defect rate of 0.09 cm^2, this yields 489 perfect dies per wafer. Of course, this doesn't account for any redundancy which no one outside Apple would know. In any case, this means that Apple's iPhones alone consumed about four months of N5 capacity in 2022. Remember that TSMC was at 120k N5 wafers per month for much of 2022. I haven't included Mac and iPads in those numbers. Including those would probably make it closer to 5.5 months of TSMC's output.

What a bizarre world! Computing and phones in one paragraph. What a waste of resources!
Apple must be banned from using the state-of-the-art process nodes. 28nm and 22nm would be ideal for the iphone users...

 

[MxPhenom 216]

Microsoft:

there’s money to be made <insert tech space>

Builds least capable version possible and it fails. Iterate minimally. Keep costs down.

Lobby government claiming the monopolistic habits of competitors causes failure.

Literally use your monopoly to force people to use your stuff.

Microsoft is much closer to hedge fund than tech company. They do the tech stuff because they have to. They’d much, much rather just acquire and exploit than innovate.

This has to be the most uninformed post I have ever seen in my life.