Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/microsofts-new-internally-designed-ai-and-cloud-server-cpu-chips.19146/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Microsoft's new internally designed AI and Cloud Server CPU chips

blueone

Well-known member
Tucked away on Microsoft’s Redmond campus is a lab full of machines probing the basic building block of the digital age: Silicon. This multi-step process meticulously tests the silicon, in a method Microsoft engineers have been refining in secret for years.

Today at Microsoft Ignite the company unveiled two custom-designed chips and integrated systems that resulted from that journey: the Microsoft Azure Maia AI Accelerator, optimized for artificial intelligence (AI) tasks and generative AI, and the Microsoft Azure Cobalt CPU, an Arm-based processor tailored to run general purpose compute workloads on the Microsoft Cloud.

 
Maia is manufactured by TSMC N5. I'm not sure about the Cobalt 100 but it's probably also using TSMC N5.

TSMC N4. I had thought Microsoft was a strong candidate for the IFS whale customer for 18A but from I have heard they are still with TSMC. Microsoft and Intel have a long standing relationship and with Pat's charm it should be a lock, right?
 
TSMC N4. I had thought Microsoft was a strong candidate for the IFS whale customer for 18A but from I have heard they are still with TSMC. Microsoft and Intel have a long standing relationship and with Pat's charm it should be a lock, right?
I am kind of 50 50 on if the domain specific foundry customers would want to use intel. On the one hand intel is promoting this system foundry approach and is willing to licence/co-develop IP and for their customers to use their software stack (which I guess means oneAPI but I don't know). You would think for firms that don't have a core competency like the cloud guys that this would be a god sent (and google is already doing that for their phones with Samsung). On the other hand their design inexperience would make me think they would want to stay on the TSMC ecosystem so they can leverage the cornucopia of off the shelf IP that is already there from EDA vendors (rather than whatever intel or the big EDA guys have already ported).

I suppose the solution to this problem would be to skip 18A and consider intel's next "full node". Let the IFS ecosystem mature, intel build their foundry fundamentals, and have the "whales" pave the way.
 
I am kind of 50 50 on if the domain specific foundry customers would want to use intel. On the one hand intel is promoting this system foundry approach and is willing to licence/co-develop IP and for their customers to use their software stack (which I guess means oneAPI but I don't know). You would think for firms that don't have a core competency like the cloud guys that this would be a god sent (and google is already doing that for their phones with Samsung). On the other hand their design inexperience would make me think they would want to stay on the TSMC ecosystem so they can leverage the cornucopia of off the shelf IP that is already there from EDA vendors (rather than whatever intel or the big EDA guys have already ported).

I suppose the solution to this problem would be to skip 18A and consider intel's next "full node". Let the IFS ecosystem mature, intel build their foundry fundamentals, and have the "whales" pave the way.
OneAPI is open source.

I suspect IFS would pay for the port of any IP a big customer would want to Intel process, if that was a key factor in closing a deal.
 
Cloud is at such a critical point why risk IFS? It’s just not prudent to go full hog with your first chip there I reckon.
Dip your toes in slowly and try it out. Makes sense to go with an established foundry and ecosystem first
 
Last edited:
Intel will be the back up plan and will fund the development work as backup so it is minimal cost to MS. If the backup is needed/chosen it will be a great investment by Intel
 
TSMC N4. I had thought Microsoft was a strong candidate for the IFS whale customer for 18A but from I have heard they are still with TSMC. Microsoft and Intel have a long standing relationship and with Pat's charm it should be a lock, right?
Apparently Charm is not as important as capability and experience. Who knew?

Reminder: TSMC has shipped 10s (100s?) of millions of chips to end customer products in use today on N5/4. Intel plans to start shipping Intel4 next month in limited quantity. Fun Fact: Now through Q2 2024, Intel will ship twice as much N4/5 silicon from TSMC as they do Intel4 Silicon. Intel 18A ships to end customers in products in mid 2025 if all goes to plan
 
I suspect IFS would pay for the port of any IP a big customer would want to Intel process, if that was a key factor in closing a deal.

This is how Samsung foundry works. Unfortunately, it takes time to port IP and silicon validate it and you don't have the experience of multiple customers maturing the IP. TSMC has it easy, all IP originates at TSMC and with the 99% market share of N3 it will be hard for Intel and Samsung to compete with IP offerings. Actually it is impossible.
 
This is how Samsung foundry works. Unfortunately, it takes time to port IP and silicon validate it and you don't have the experience of multiple customers maturing the IP. TSMC has it easy, all IP originates at TSMC and with the 99% market share of N3 it will be hard for Intel and Samsung to compete with IP offerings. Actually it is impossible.
Until IFS gets up to speed, the primary reasons to choose IFS over TSMC will be companies forced to reduce geopolitical risk or something national security related. Or IFS might make them an offer they can't refuse to prove themselves a worthy supplier. After Xi's quip in SF about China and US being partners or adversaries, with no middle ground option, I suspect IFS's opportunity space just got wider. Probably so did Samsung's. Personally, I think that's silly, but in the business world optics are more important than reality. But you know that.
 
Until IFS gets up to speed, the primary reasons to choose IFS over TSMC will be companies forced to reduce geopolitical risk or something national security related. Or IFS might make them an offer they can't refuse to prove themselves a worthy supplier. After Xi's quip in SF about China and US being partners or adversaries, with no middle ground option, I suspect IFS's opportunity space just got wider. Probably so did Samsung's. Personally, I think that's silly, but in the business world optics are more important than reality. But you know that.

IFS has the NOT TSMC market to bootstrap business. The rest will have to be earned the good old fashioned way, superior PDKs.

You should have seen the security for Xi's visit, even on the water. Lots of restrictions of where we could sail. They even closed down lanes on the Bay Bridge so it was work from home week. The word on the street (Silicon Valley) is that Xi's visit is going well. Lot's of billionaires here (not me) and where there is money there is political clout. Personally I think there will be less semiconductor geopolitical risk once the 2024 elections are over (Taiwan and US). And if you think there will be a new POTUS there are closed lanes on the Bay Bridge that I can sell you. :ROFLMAO:
 
IFS has the NOT TSMC market to bootstrap business. The rest will have to be earned the good old fashioned way, superior PDKs.
Agreed. A better way of stating what I did.
You should have seen the security for Xi's visit, even on the water. Lots of restrictions of where we could sail. They even closed down lanes on the Bay Bridge so it was work from home week. The word on the street (Silicon Valley) is that Xi's visit is going well. Lot's of billionaires here (not me) and where there is money there is political clout. Personally I think there will be less semiconductor geopolitical risk once the 2024 elections are over (Taiwan and US). And if you think there will be a new POTUS there are closed lanes on the Bay Bridge that I can sell you. :ROFLMAO:
Ugh. Betting on a presidential election? No thanks. You can keep your bridge lanes in inventory. Who knows what the electorate will do choosing between :poop: and 🤮.
 
Agreed. A better way of stating what I did.
Ugh. Betting on a presidential election? No thanks. You can keep your bridge lanes in inventory. Who knows what the electorate will do choosing between :poop: and 🤮.

Agreed, a better way of stating how I feel: Who knows what the electorate will do choosing between :poop: and 🤮. I just wanted to work the bridge lane thing into a joke.
 
IFS has the NOT TSMC market to bootstrap business. The rest will have to be earned the good old fashioned way, superior PDKs.

You should have seen the security for Xi's visit, even on the water. Lots of restrictions of where we could sail. They even closed down lanes on the Bay Bridge so it was work from home week. The word on the street (Silicon Valley) is that Xi's visit is going well. Lot's of billionaires here (not me) and where there is money there is political clout. Personally I think there will be less semiconductor geopolitical risk once the 2024 elections are over (Taiwan and US). And if you think there will be a new POTUS there are closed lanes on the Bay Bridge that I can sell you. :ROFLMAO:
Polling heavily indicates that there will be a new POTUS.
 
TSMC N4. I had thought Microsoft was a strong candidate for the IFS whale customer for 18A but from I have heard they are still with TSMC. Microsoft and Intel have a long standing relationship and with Pat's charm it should be a lock, right?
I've seen a Microsoft slide that says TSMC 5nm.
 
Unfortunately, it takes time to port IP and silicon validate it and you don't have the experience of multiple customers maturing the IP.
Simple answer... make shuttles, especially those of us who can make PDKs (models, stdcells, pcells), layout automation, DRC and LVS checkers, etc.

What IP takes a long time to port over?
 
I would hope Americans arent so stupid as to vote for that orange cretin again. Mindblowing how so many can support such a vile man
It's voter turnout that decides elections and Trump's supporters turn out very strongly compared to Democratic voters who in many cases are not fully behind Joe Biden except as an alternative to Trump. Even when Biden is ahead in polls you have to assume that Trump will outperform by a few percentage points because of turnout, however right now Biden is behind in polls and it is not looking very good, and will probably get worse since Trump is likely to get a primary boost after he secures the Republican nomination (almost a sure thing).
 
Back
Top