Mark Liu, AI chips to Displace server chips

[Arthur Hanson]

Intel's Revenue Forecast Tops Expecations

Wall Street has been worried about the risk that surging demand for artificial-intelligence chips could divert spending from traditional processors.

www.barrons.com

Mark Liu of TSM sees AI chips displacing server chips. This should benefit TSM, but cut into Intel's business. Any thoughts on how large this displacement could be would be appreciated.

 

[blueone]

Intel has an opposing view. It is likely that both views are correct in some circumstances, depending on the applications.

Intel Xeon is all you need for AI inference: Performance Leadership on Real World Applications

Sanchit Misra is a senior research scientist and leads the efforts in computational biology/HPC research at Intel Labs. Highlights: Intel is democratizing AI inference by delivering a better price and performance for real-world use cases on the 4th gen Intel® Xeon® Scalable Processors...

community.intel.com

 

[nghanayem]

Intel's Revenue Forecast Tops Expecations

Wall Street has been worried about the risk that surging demand for artificial-intelligence chips could divert spending from traditional processors.

www.barrons.com

Mark Liu of TSM sees AI chips displacing server chips. This should benefit TSM, but cut into Intel's business. Any thoughts on how large this displacement could be would be appreciated.

Considering how many folks said mobile would make the PC into the next mainframe, I don't put much stock in the AI chips making all prior chips irrelevant claim. I wouldn't be surprised if a proliferation of ML workloads causes some segments to stop or slow their growth, but a complete displacement is a very rare thing to see unless something is absolutely better (for example the vacuum tube and the transistor). Currently when it comes to latency bound compute, high throughput chips are nowhere close.

 

[Daniel Nenni]

An interesting comment on AI which does not bode well for Intel and AMD:

C. C. Wei "Our model is based on the data center structure. We assume a certain percentage of the data center processor are AI processors. And based on that, we calculate the AI processor demand. And this model is yet to be fitted to the practical data later on. But in general, I think the - our trend of a big portion of data center processor will be AI processor is a sure thing. And we can cannibalize the data center processors. In the short term, when the CapEx of the cloud service provider are fixed, yes, it will. It is. But as for the long term, when their data service, when the cloud services having the generative AI service revenue, I think they will increase the CapEx, that should be consistent with the long-term AI processor demand. And I mean, the CapEx will increase because of a generative AI services."

As I have said before, bespoke silicon will rule the semiconductor world and this is just one example, absolutely.

 

[blueone]

An interesting comment on AI which does not bode well for Intel and AMD:

C. C. Wei "Our model is based on the data center structure. We assume a certain percentage of the data center processor are AI processors. And based on that, we calculate the AI processor demand. And this model is yet to be fitted to the practical data later on. But in general, I think the - our trend of a big portion of data center processor will be AI processor is a sure thing. And we can cannibalize the data center processors. In the short term, when the CapEx of the cloud service provider are fixed, yes, it will. It is. But as for the long term, when their data service, when the cloud services having the generative AI service revenue, I think they will increase the CapEx, that should be consistent with the long-term AI processor demand. And I mean, the CapEx will increase because of a generative AI services."

As I have said before, bespoke silicon will rule the semiconductor world and this is just one example, absolutely.

Your prediction of customer-designed chips I completely agree with. I'm only in partial agreement, if that, with Wei's view that it's AI processors that will displace general purpose CPUs. That may be a factor, but I believe chips like Amazon's Nitro will be a far larger factor.

 

[Arthur Hanson]

I think the market cap of TSM about 750 billion vs Intel's at under 50 billion should settle this argument. TSM since it doesn't compete with their customers will continue to widen the gap. They have taken collaboration to an art form to produce superior leading edge products.

 

[blueone]

Intel's market cap is, as of this posting, $145B. Still pathetic, but not as bad as you say.

 

[nghanayem]

I think the market cap of TSM about 750 billion vs Intel's at under 50 billion should settle this argument. TSM since it doesn't compete with their customers will continue to widen the gap. They have taken collaboration to an art form to produce superior leading edge products.

Boeing has like 10x the market cap of Nissan, does this mean that the B2 stealth bomber will replace the whole market for automobiles? Saying TSMC bigger than intel doesn't mean that CPUs/GPGPUs/FPGAs/non AI ASICs are suddenly going to all be replaced by A100s and other random customer AI ASICs. TSMC makes all of these products, so any shift in what chips people are buying would have ZERO, NADA, ZILCH impact on TSMC. TSMC is making just as much dough from x wafers of A100, as they make on x wafers of Genoa, which is also the same revenue as x wafers of A12. In short the size of intel and TSMC has no relation to the fantasy investors have and that NVIDA peddles of AI super chips replacing all compute from the mobile/edge to the DC.

 

[count]

I don’t think it will displace overall market size for servers but AI will capture most of the growth, so market share will fall.

Bigger threat to Intel is AMD/Ampere. If the pie for server chips stops growing, then only way to grow is take share, which will lead to compressed margins.

 

[ta152h]

I think the market cap of TSM about 750 billion vs Intel's at under 50 billion should settle this argument. TSM since it doesn't compete with their customers will continue to widen the gap. They have taken collaboration to an art form to produce superior leading edge products.

TSMC is about $524 billion, so both Intel and TSMC numbers are quite a bit off.

This is a snapshot in time, it's not definitive. Let's see a few years from now. Or even tomorrow ;-P . Earnings should be interesting tonight.

 

[blueone]

I don’t think it will displace overall market size for servers but AI will capture most of the growth, so market share will fall.

Bigger threat to Intel is AMD/Ampere. If the pie for server chips stops growing, then only way to grow is take share, which will lead to compressed margins.

Don't forget AWS Nitro/Graviton. MSFT is on the server CPU path too, according to rumors. I'm surprised Google hasn't yet. I think these are bigger threats than AMD and Ampere.

 

[Tanj]

Bigger threat to Intel is AMD/Ampere. If the pie for server chips stops growing, then only way to grow is take share, which will lead to compressed margins.

AMD yes. But if the focus shifts to AI leaving classic servers adrift, the resources for porting (at both clouds and their customers) shrink as funds and engineers move over to catch the AI wave, so that could leave Ampere with slow adoption. Especially if both Intel and AMD deliver gains in cost effectiveness, footprint, and power in the next couple of years. It will be a scramble.

 

[count]

AMD yes. But if the focus shifts to AI leaving classic servers adrift, the resources for porting (at both clouds and their customers) shrink as funds and engineers move over to catch the AI wave, so that could leave Ampere with slow adoption. Especially if both Intel and AMD deliver gains in cost effectiveness, footprint, and power in the next couple of years. It will be a scramble.

I don’t think porting is an issue anymore. Most use cases are already ported and available on ARM.

 

[count]

Don't forget AWS Nitro/Graviton. MSFT is on the server CPU path too, according to rumors. I'm surprised Google hasn't yet. I think these are bigger threats than AMD and Ampere.

MSFT is partnered with Ampere on this. But yes Graviton is a big threat as well.

 

[blueone]

MSFT is partnered with Ampere on this.

Yeah, I saw that announcement, but I also heard more than once that they're working on something Graviton-like for Azure. We'll see.

 

[lilo777]

Intel just reported their earnings and, based on market reaction, they are fine. As far as AI chips replacing conventional processors, this notion was probably taken out of context. If anything, we do not need more AI chips without simultaneously increasing manufacturing of CPUs. All major AI applications currently are ML based. ML models get trained on the datasets captured, processed, stored and managed by conventional CPUs. Without this ecosystem (social media, sites like stackoverflow, countless databases, streaming infrastructure etc.) AI chips would be a useless pile of silicon. Take EDA as an example. The AI (ML) applications, I am familiar with, rely on different types of training the models on tons of data produced by gazillions of simulation/validation runs. Processing resources needed for AI in this case are just a small blip compared to the compute resources consumed by conventional EDA algorithms that produce the datasets (potentially as a byproduct) used for ML training.

 

[Arthur Hanson]

My mistake on market caps, as of this morning TSM 473B, INTC 144B, these numbers and their growth/profit numbers clearly shows TSM has a substantial lead financially and technically.

 

[ta152h]

My mistake on market caps, as of this morning TSM 473B, INTC 144B, these numbers and their growth/profit numbers clearly shows TSM has a substantial lead financially and technically.

I'm not sure why you say technically. Financially is unclear as well, but there's a much stronger correlation. Technically? Intel is in markets TSMC knows nothing about, with technology TSMC simply doesn't have. They still have the best fab process for performance, and always have. But, if you want to say they are ahead in process technology, I'm OK with that, because overall I think most people would agree they do. But, overall, it's hard to make comparisons, because Intel still has the best single-threaded performance of any processor made, and that's often considered the holy grail of CPU design. It's an awkward comparison at best.

You could also say TSMC has leveraged their technology better than Intel, and probably not too a huge argument. But, that's for right now. Everything changes. It has so much already.

I'll tell you this, I'm much happier that I'm a big holder of Intel, than I would be if I were stuck with TSMC. I made the right choice, looks like, as Intel is recovering, and TSMC is languishing. And they might be OK for a while, but 60% is just so high a market share, I don't think its sustainable. Especially with Intel and Rapidus becoming more prominent players. Yeah, I know Rapidus is a huge unknown, but they have so much support from major companies, and the Japanese government, they shouldn't be ignored. Yet. It could play out they become largely irrelevant, that's for sure.

 

[ai268]

Intel is in markets TSMC knows nothing about, with technology TSMC simply doesn't have.

Could you elaborate? I am a programmer. That's why I ask.

 

[ta152h]

Could you elaborate? I am a programmer. That's why I ask.

I could write a 200 page book on things Intel does that TSMC doesn't. OK, maybe not that big.

If you're a programmer, what compilers do you use? Hmmmm, Intel always perform the best, don't they? Even for AMD processors they normally do. How about operating systems? Who makes the fastest Linux? Yeah, Intel. Intel designs firmware. TSMC? I doubt it. And I'm just touching on software, there are so many other things Intel does that TSMC does not in software.

Hardware, does TSMC have the first idea how to design a high performance processor, or GPU? How about car driving software and hardware? I could go on and on, but on balance, I'd have to say Intel technology is among the best in the world, whereas TSMC is very good at a market segment, but simply doesn't have the breadth of technology companies like Intel have.

Does that make them worse? Of course not. I just object to overreaching statements like they have better technology. It's such a strange and difficult comparison, I don't know how you could possibly judge it. Based on finances? Please. Especially when Intel typically makes much more money than TSMC. It's a mistake to take a snapshot in time and expand it to become an almost universal or broad truth.

I'm not taking a slap at TSMC, just broad comparisons that are really difficult to make because they are so different.