Nvidia CEO: Intel Test Chip Results for Next Gen Process Look Good

[Daniel Nenni]

Intel IP is not necessarily a good fit for fabless companies. I remember the big hoohaaa about all of the Intel IP IFS can offer to customer. As it turns out there is not much IP that would work for the big fabless companies. Established fabless companies have a tried and true IP library that they will not give up. And if they did chose Intel IP they would be stuck at IFS which is not a good position to be in. Intel can, however, port customer IP over easy enough so IP is not a roadblock but the Intel IP catalog is not a big part of the IFS value proposition.

 

[delong.height]

1. A dedicated foundry, like TSMC, builds fab capacity based on customers' commitment and prepayments. On the other hand Intel chose to make IFS to serve both internal and external customers. IMO, it's a tough situation for Intel to deal with when time is good or bad. Will IFS sue internal Intel brothers and sisters for not fulfilling their order commitments?

2. Probably many people here do not agree with me on this. IFS is forced to operate under an awkward situation. There are many conflict of interests and contradicting goals they need to pursue. What you have pointed out is just one of the difficulties Intel is going to face. To me the best solution is to spin off IFS to set them free. It's good for both Intel product divisions and IFS.

But Pat Gelsinger thinks Intel must keep both product divisions and IFS together in order to survive. I think it's a high risk and unnecessary gamble.

When people buy products from AMD, Apple, Qualcomm, Nvidia, and Mediatek, they don't really care who actually manufacturered those products. They buy a particular product is because the superior price, performance, quality, features, and availability. None of the above mentioned companies operates a fab and all of them are doing just fine.

Intel is a product company, not a service company, because majority of Intel revenue are generated from selling semiconductor products. Intel biggest problem is that it doesn't have a killer product. To spend billions to create a new service business (IFS) is placing precious money, time, and management attention on wrong priority. IFS won't be able to bring revenue to Intel that is large enough and quick enough to offset Intel product division's shortfall.

the fact is even if Intel did spin up its IFS unit, it's no match when compared to TSMC given its scale, ecosystem, and customers. The best and deep pocket customers are already partnering with TSMC. Spinning up IFS is basically making Intel much weaker and cut off its strongest arm. How can the R&D ever be shared? What could happen to Core product line, and Xeon product line when the only competitiveness is their ability to their economies of scale and vertical integration, and the ability to flood the market with their in-house manaufactured chips. AMD, Apple, Qualcomm, Nvidia, and MediaTek, and others only care about if their products can be produced on time, with a healthy yield. They don't give a shit about IFS. Whether it lives or die, it's not their business. And how can they hand off massive orders to IFS when IFS no longer have this strong backing from Intel? when they have no money to invest in next generation of EUV, process technologies, packaging technologies, retaining talent, building more shells. Your advice will have the same result as AMD spinning off Global Foundries, making the latter no longer can afford go into more advanced processes. And Intel will forever doomed

 

[blueone]

Intel IP is not necessarily a good fit for fabless companies. I remember the big hoohaaa about all of the Intel IP IFS can offer to customer. As it turns out there is not much IP that would work for the big fabless companies. Established fabless companies have a tried and true IP library that they will not give up. And if they did chose Intel IP they would be stuck at IFS which is not a good position to be in. Intel can, however, port customer IP over easy enough so IP is not a roadblock but the Intel IP catalog is not a big part of the IFS value proposition.

Completely agree. Intel's IP they are willing to license is probably not interesting, but if a big IFS customer is involved I suspect Intel will be willing to fund the port. When I say "willing to license", I suspect Intel will not be willing to license proprietary IP that is fundamental to the Intel product groups' competitiveness. For example, Intel's excellent state machine-based compression and encryption blocks.

 

[blueone]

the fact is even if Intel did spin up its IFS unit, it's no match when compared to TSMC given its scale, ecosystem, and customers. The best and deep pocket customers are already partnering with TSMC. Spinning up IFS is basically making Intel much weaker and cut off its strongest arm. How can the R&D ever be shared? What could happen to Core product line, and Xeon product line when the only competitiveness is their ability to their economies of scale and vertical integration, and the ability to flood the market with their in-house manaufactured chips. AMD, Apple, Qualcomm, Nvidia, and MediaTek, and others only care about if their products can be produced on time, with a healthy yield. They don't give a shit about IFS. Whether it lives or die, it's not their business. And how can they hand off massive orders to IFS when IFS no longer have this strong backing from Intel? when they have no money to invest in next generation of EUV, process technologies, packaging technologies, retaining talent, building more shells. Your advice will have the same result as AMD spinning off Global Foundries, making the latter no longer can afford go into more advanced processes. And Intel will forever doomed

I don't agree with a single point in this post.

 

[nghanayem]

Intel IP is not necessarily a good fit for fabless companies. I remember the big hoohaaa about all of the Intel IP IFS can offer to customer. As it turns out there is not much IP that would work for the big fabless companies. Established fabless companies have a tried and true IP library that they will not give up. And if they did chose Intel IP they would be stuck at IFS which is not a good position to be in. Intel can, however, port customer IP over easy enough so IP is not a roadblock but the Intel IP catalog is not a big part of the IFS value proposition.

That is my view as well for large design houses. But I think the value prop of intel licencing out cores and IP to IFS customers would be for companies who don't have a strong design core competency rather than an Nvidia or MTK. Given Google's new phone SOC was co-designed by Samsung and Microsoft's apparent struggles designing their own ARM SOCs and server chips. As an example I could see these folks being interested in Seira Forest cores, with their own customization to the accelerators to best suit their workloads. All of this then being co-packaged with some AI ASIC that they either made at TSMC, or bought off the shelve from a smaller design house.

Completely agree. Intel's IP they are willing to license is probably not interesting, but if a big IFS customer is involved I suspect Intel will be willing to fund the port. When I say "willing to license", I suspect Intel will not be willing to license proprietary IP that is fundamental to the Intel product groups' competitiveness. For example, Intel's excellent state machine-based compression and encryption blocks.

They've said they want to licence out the x86 ISA and their newest cores. Time will tell though how much of that (if anything) comes to fruition. I hope it does though, it would be so cool to see this idea of a "systems foundry" be realized. It is certainly unique to have foundry services extend to IP licencing/semi-custom design/software rather than IFS only existing as a typical wafer foundry with advanced packaging capabilities. If this concept ends up working, I think that would also be an area where IDM can strengthen IFS's offerings in one aspect that I don't think would otherwise be possible from a pureplay operating model.

 

[blueone]

They've said they want to license out the x86 ISA and their newest cores. Time will tell though how much of that (if anything) comes to fruition. I hope it does though, it would be so cool to see this idea of a "systems foundry" be realized. It is certainly unique to have foundry services extend to IP licensing/semi-custom design/software rather than IFS only existing as a typical wafer foundry with advanced packaging capabilities. If this concept ends up working, I think that would also be an area where IDM can strengthen IFS's offerings in one aspect that I don't think would otherwise be possible from a pureplay operating model.

Yeah, I read that stuff about licensing x86 cores in The Register. (Link below for those who haven't seen this.)

Intel to license x86 CPU cores for use in custom processors

If you've wanted to create your own specialized Xeon, now's your chance

www.theregister.com

Before anyone gets all excited about it, if you read closely you will find that Intel is willing to license cores as hard IP for production purposes, and I would guess they're referring to Xeon D cores (not stated in the article), and this would be a very stiff lock-in strategy to IFS. (The soft IP is intended for simulation on FPGAs, as I would have guessed.) This is not anything like being able to build an AMD competitor, or getting an Arm Architecture License. I'm also not thinking Xeon Ds are competitive with the best of Arm and RISC-V cores becoming available either.

I'm not excited about x86 IP at all. If IFS wants to win it needs winning processes, an awesome external IP portfolio, PDKs customers want, great execution, and competitive pricing.

 

[ta152h]

That's like asking "how long is a piece of string?"...

Huge chips pushing the maximum reticle size like Nvidia and Intel and Xilinx used (700mm2 or so) have the advantage of being monolithic so not needing any inter-die links, but the yield is *terrible* so the cost per mm2 of silicon is high -- and this cost dwarfs any adder for multi-die chiplet packaging, even if silicon interposers are used. There's also the issue that if you want a family of parts you need a family of chips, and have to pay the development and mask costs for each one which are huge nowadays. With chiplets TTM can be quicker, and spinning out new family versions only needs new interposer/package which is *way* quicker and cheaper than a new chip design.

Depending on what the application is, the inter-die links with chiplets may have almost zero, a small, or a large penalty in performance -- but for big devices a single die will always be more expensive to manufacture, and harder to ship in volume because you throw many bad die away for every good one, some of the biggest chips have single-digit percentage yields.

So if your application values performance over everything else including design/manufacture cost, flexibility, TTM, and high volumes, and it will fit on one big die, and you can sell the end result for a high enough price, then single-die is the way to go. This is what Intel used to do, and why their market share is shrinking.

If cost and flexibility and TTM and big volumes (high yield) are more important, or if it's just too big for single-die, then chiplets are the way to go. This is what AMD have been doing, and why their market share is growing.

There are other advantages to chiplets such as being able to mix different technologies optimised for different functions (e.g. CPU/DSP core vs. SERDES/IO) and not having to make or buy all the (increasingly expensive) IP you need on one bleeding-edge process.

As the cost/effort of each chip development has gone up with each process node, the market is moving more and more to chiplets to keep cost and headcount under control, especially where you want a family of parts not just one.

Intel's market share is shrinking? I'm guessing you're referring to servers, because it's well over 80% in desktops and mobile.

But, to say it's because they didn't use chiplets, is a bit oversimplified, at best, and incorrect, at worse.

They had horrible performance because 10nm was delayed at first, so they were stuck with an architecture from 2015 using an obsolete fabrication technology. Then then went to Ice Lake, which was on 10nm before 10nm had any real performance and was renamed. It was slower, had fewer cores, and was less power efficient. Not sure how chiplets were going to help there. And Sapphire Rapids was delayed, and delayed, and delayed and by most people's considerations, are a node behind what AMD is selling on.

Chiplets, or a lack thereof, were not the primary cause. By no means am I suggesting it's a bad approach, and/or doesn't have advantages; somehow people often extrapolate from what is stated and make it a bigger brush than was intended. I am just saying their loss of market share in servers was not primarily because they were using larger dies, there were other fundamental problems with the efficiency and performance of their architectures that gave AMD a clear advantage that still has not been completely overcome. Even with chiplets.

Granite Rapids may gain back market share, with a new process node and new architecture. Sapphire Rapids is not, and it's doubtful Emerald Rapids will either. Intel 3 is going to be the most salient characteristic of that processor, in my opinion, and better offer enough efficiency and density to make it a compelling server chip.

 

[ta152h]

That is my view as well for large design houses. But I think the value prop of intel licencing out cores and IP to IFS customers would be for companies who don't have a strong design core competency rather than an Nvidia or MTK. Given Google's new phone SOC was co-designed by Samsung and Microsoft's apparent struggles designing their own ARM SOCs and server chips. As an example I could see these folks being interested in Seira Forest cores, with their own customization to the accelerators to best suit their workloads. All of this then being co-packaged with some AI ASIC that they either made at TSMC, or bought off the shelve from a smaller design house.


They've said they want to licence out the x86 ISA and their newest cores. Time will tell though how much of that (if anything) comes to fruition. I hope it does though, it would be so cool to see this idea of a "systems foundry" be realized. It is certainly unique to have foundry services extend to IP licencing/semi-custom design/software rather than IFS only existing as a typical wafer foundry with advanced packaging capabilities. If this concept ends up working, I think that would also be an area where IDM can strengthen IFS's offerings in one aspect that I don't think would otherwise be possible from a pureplay operating model.

This could also really play out in the game console business. I'd be very surprised to not see Intel go after this pretty aggressively at some point.

 

[Xebec]

What's the direct interpretation of Jensen's comment --

Is it that he saw data on SRAM, ARM core or other "generic" test chips from Intel,
or that Nvidia actually ran some test chips through an advanced Intel node?

 

[ta152h]

What's the direct interpretation of Jensen's comment --

Is it that he saw data on SRAM, ARM core or other "generic" test chips from Intel,
or that Nvidia actually ran some test chips through an advanced Intel node?

I think he was intentionally vague, and your guess is as good as anyone else's, except the people close to him,

But, my interpretation, be careful TSMC, you aren't the only game in town, and if you keep jacking up prices, we'll go to Intel. Does he intend to carry out this threat? I have doubts, at least short term, but clearly having another possible maker, and saying they are getting good results, improves his bargaining position with TSMC. Plus, the increased orders he's putting in don't hurt either.

But, maybe I'm just be snarky, and he's just saying what he thinks. But, he's a smart guy, he normally has a purpose in what he says.

 

[hist78]

Very unlikely.

Why?

 

[hist78]

Nvidia to contract TSMC to produce next-generation chips: CEO

Nvidia to contract TSMC to produce next-generation chips: CEO - Focus Taiwan

Taipei, June 2 (CNA) U.S.-based graphics processing unit (GPU) designer Nvidia Corp. will continue to use Taiwan Semiconductor Manufacturing Co. (TSMC), the world's largest contract chipmaker, to roll out its next-generation chips.

focustaiwan.tw

 

[soAsian]

regarding Intel's IP. that's a trap. i don't see how any CEO will agree to it. you have to use Intel fabs to use Intel's IP.

 

[Brady]

regarding Intel's IP. that's a trap. i don't see how any CEO will agree to it. you have to use Intel fabs to use Intel's IP.

Is their IP becoming more standard with Intel design using TSMC so much, and even rumored to be doing so for x86 cores for some Arrow Lake SKUs (N3)?

 

[Daniel Nenni]

Completely agree. Intel's IP they are willing to license is probably not interesting, but if a big IFS customer is involved I suspect Intel will be willing to fund the port. When I say "willing to license", I suspect Intel will not be willing to license proprietary IP that is fundamental to the Intel product groups' competitiveness. For example, Intel's excellent state machine-based compression and encryption blocks.

Samsung does this as well. If a foundry customer needs IP Samsung gets it, simple as that. TSMC on the other hand already has the IP because most IP is developed with TSMC. It is an unfair advantage but an advantage just the same. One of the benefits of being a trusted pure play foundry. IDM foundries will never be on a level playing field with pureplay, absolutely. The advantage of IDM foundries is they they have a ton of cash to burn from their other businesses. Unfortunately that cash may be limited as we now see with recent Samsung and Intel financials.