Intel Alder Lake vs. Apple M1

[senttoschool]

Which is exactly as predicted. Note that the graph is not correct in respect to the Apple CPU. M1 Max CPU cluster only draws 30W (35W package power including RAM) to reach 12K in R23 multicore.

So what we have here is a benchmark that is pretty much optimal case for Intel (sales very well with multiple cores and SMT), is know to underuse Apple hardware (runs an Intel SIMD optimised library with a SSE2-to-NEON software layer, using a suboptimal SIMD width on M1), where Intel's 6+8 SKU at 40W shows the same performance as Apple's 8+2 SKU at 30W. That's a real-world difference in perf-per-watt of around 40%. Intel still has a long way to go until they can catch up with Apple in the mobile space.

Where Intel of course has an obvious edge is performance scaling with power. Which is again not surprising given their architecture and the focus on power-hungry desktop applications. If you are a desktop HPC user that benefits from multicore scaling, ADL is very good product. But we already knew that. It will also undoubtedly be very popular in laptops, just don't expect those level of performance inside the usual multimedia chassis.

Anyway, all of this confirms that Apple's decision to move on was the correct one. Personally, I am happy to have my 15+ hour battery life and 2x better performance than Intel while working away from my desk, and I have no problems conceding the R23 or stockfish scores to the x86 folks if that makes them happy

Cinebench is a terrible benchmark in general and it is not indicative of M1's performance. Basically, it sucks for M1 but great for AMD and Intel chips, like you said.

What I do find from this is that Intel is quite a bit ahead of AMD chips, despite, in my opinion, being on an inferior Node. Intel 7 is still worse than TSMC 7nm. This basically confirms that ADL does have better performance/watt than Ryzen when it's configured more optimally in its power curve.

Basically, Cinebench is useless for M1. But Intel vs AMD is a fairer fight.

There were a lot of people in this thread who denied that ADL was as efficient or more efficient than Zen3. @crazy dave @Andropov

 

[mi7chy]

Separate CPU and GPU is also better for thermal management. More integrated heat sink surface area and two warm spots versus one hot spot.

 

[cmaier]

Separate CPU and GPU is also better for thermal management. More integrated heat sink surface area and two warm spots versus one hot spot.

No, that is not true at all. Thermal spread on the actual silicon die is minimal (which is why local hot-spotting is a thing). In general, with air cooling, you can dissipate a certain number of watts per square cm of package size surface area. (When I was designing packages, 10W/cm^2 was common. With aerosol cooling and recent material advances, the number is probably higher now).

The point being that it doesn’t matter whether you have two packages with space between them or one big package.

 

[crazy dave]

It's just the basic benefit of modularity. If there are 10 CPU types and 10 GPU types and the CPU and the GPU are manufactured separately, there are 20 different products. If you integrate them into the same chip, you may need up to 100 different products. Having too many product types makes everything more expensive, and predicting the demand for each product in advance becomes more difficult. A modular design makes it possible to serve smaller niches (such as consumers that need multithreaded performance) economically than an integrated design.

Very true. However that only comes into play if Apple was trying to serve 100 different markets with 100 different SKUs. They aren’t. They can absolutely serve consumers with high multithreaded throughput CPU needs. But not consumers with high multithreaded CPU needs who also want a mere 7 core GPU needs or the inverse … consumers who want a 128 core AS GPU but only an 8 core 4+4 CPU. That’s when they would run into difficulties that you describe if they tried to cater to all possible customers perfectly. Now multidie chips may eventually increase modularity (not to the extent of what’s possible from mixing and matching CPU and dGPUs obviously) but for now it is being setup to serve high CPU & GPU core counts with multi max dies. They are not stopping with the Max in terms of CPU throughput.

Where the economics favor this is in not having to pay AMD or Intel for parts - ie not having to pay for their profit margins, just TSMC’s. Then Apple simply constrains the number of distinct dies to keep costs down. The smallest number of chips to serve the broadest range of customers does indeed mean some niches won’t be catered to perfectly, but Apple will have a product there. Again, Apple are an OEM building the final product, not a chip company trying to sell a hundred variants of a CPU to be mixed and matched with an equal number of dGPUs. Different business models, different economics, different definition of scale.

 

[crazy dave]

Cinebench is a terrible benchmark in general and it is not indicative of M1's performance. Basically, it sucks for M1 but great for AMD and Intel chips, like you said.

What I do find from this is that Intel is quite a bit ahead of AMD chips, despite, in my opinion, being on an inferior Node. Intel 7 is still worse than TSMC 7nm. This basically confirms that ADL does have better performance/watt than Ryzen when it's configured more optimally in its power curve.

Basically, Cinebench is useless for M1. But Intel vs AMD is a fairer fight.

There were a lot of people in this thread who denied that ADL was as efficient or more efficient than Zen3. @crazy dave @Andropov

View attachment 1955021

View attachment 1955023

No what we denied was that the ADL big cores, Golden Cove, were as efficient in either die area or power than Zen 3 big cores. Which this doesn’t refute because that’s not what is measured in either graph.

With respect to these graphs @Gerdi ’s posts still apply:

Intel Alder Lake vs. Apple M1

Crazy Dave already pointed out, that this is not true at all. The really big flaw in the drawing is, that architectures are not compared at iso voltage. The fundamental reason, that this happens is, that you compare CPUs with difference number of cores. CPUs with less cores are at an inherent...

forums.macrumors.com

Intel Alder Lake vs. Apple M1

Of course, if you strictly restrict your discussion to actual products, thats is a somewhat valid argument. And you can make an inefficient microarchitecture look good in actual products as long as you have sufficiently more cores compared to others. My drawing just illustrates how much the...

forums.macrumors.com

It’s also covered by @jeanlain and by @BigPotatoLobbyist and by @leman

When it has been measured against Zen (2 only alas) the efficiency of the big cores is still wanting:

Alder Lake’s Power Efficiency – A Complicated Picture

Reviews across the internet show Alder Lake getting very competitive performance with very high power consumption. For example, Anandtech measured 272 W of package power during a POV-Ray run. Our o…

chipsandcheese.com

That’s why the middle cores exist. To improve Intel’s multithreaded throughput against AMD’s per area per watt.

 

[jeanlain]

If Apple wanted to throw four more P cores onto the M1 Pro instead of a fat GPU and the video encoding crap, it would blow Intel out of the water

Except in video encoding and rendering...

 

[senttoschool]

No what we denied was that the ADL big cores, Golden Cove, were as efficient in either die area or power than Zen 3 big cores.

Hardware Unbox literally just showed that ADL under more optimal power curves trumps Zen 3.

Why does the die area matter? ADL still has AVX512 silicon and it has a GPU (desktop zen3 does not have a GPU), both of which take up huge die space. We don't have official die area numbers for CPU anyways.

Are consumers buying chips based on die size? Why does dies size matter if the results are what matter the most?

The argument that Intel's big.Little is actually big.Medium is pointless. Just look at the results. Intel right now pretty much takes a dump on AMD's Zen3. Zen4 laptop isn't expected to ship until a year from now and the 6nm refresh of Zen3 isn't going to help much.

 

[crazy dave]

Hardware Unbox literally just showed that ADL under more optimal power curves trumps Zen 3.

Why does the die area matter? ADL still has AVX512 silicon and it has a GPU (on desktop while Zen 3 does not on desktop), both of which take up huge die space. We don't have official die area numbers for CPU anyways.

Are consumers buying chips based on die size? Why does dies size matter if the results are what matter the most?

The argument that Intel's big.Little is actually big.Medium is pointless. Just look at the results. Intel right now pretty much takes a dump on AMD's Zen3. Zen4 laptop isn't expected to ship until a year from now and the 6nm refresh of Zen3 isn't going to help much.

They made the measurements on multithreaded output. That’s not nothing but the whole point of @Gerdi ’s posts is that can belie the differences in core micro architecture and even chip design. It’s why I linked to the chips and cheese article which did the power curves on each core.

The point of our earlier discussion was why did Intel go to big.Medium? The reason why they did so was to make up for the deficiencies in their big cores in multithreaded output. The reason die area matters is because it constrains both chip design and economics. I’m speaking from the perspective of the people building the chips. The big.Medium tactic was a way for Intel to compete. And it works! You seem convinced that I think ADL is bad. I don’t. I think the hybrid core design is a viable solution to the problem Intel found itself in and it worked really well. It allowed them to put far more cores onto the same die without busting power thus increasing their multithreaded throughput per die per area over AMD. If you have workloads that can take advantage of all those extra threads, that’s great!

But you’re attempting to extend your conclusions too far. That’s the issue. Let’s take an example:

Two chips with the same architecture and all the same cores, let’s say Intel 10th gen it doesn’t really matter, one with 24 cores, one with 4 cores. At the same wattage which chip will out perform the other in a multithreaded benchmark? Almost always across the power curve the 24 core chip will win with increasing wins as power goes up. Does that mean that the 24 core chip has a better microachitectural design? Nope they’re the same. That’s what @Gerdi ‘s post was about.

 

[senttoschool]

When it has been measured against Zen (2 only alas) the efficiency of the big cores is still wanting:

That's not the conclusion of the article. It's quite clear that Intel pushes desktop ADL hard because efficiency matters less on desktop.

Let's not cling to one article's conclusion which is interpreted wrongly by you anyways.

PS. Performance/watt is not all that matters. What actually matters is "desired performance by user / watt". If you try to push Zen2 to have equivalent performance to ADL, I'm 100% sure that you'd need far more power and probably LN2 to achieve that goal.

If you go by performance/watt as the only metric, then some little tiny CPU would beat the pants off the M1, A15, Intel, AMD, etc. Performance does not scale linearly with power.

 

[crazy dave]

That's not the conclusion of the article. It's quite clear that Intel pushes desktop ADL hard because efficiency matters less on desktop.

Let's not cling to one article's conclusion which is interpreted wrongly by you anyways.

PS. Performance/watt is not all that matters. What actually matters is "desired performance by user / watt". If you try to push Zen2 to have equivalent performance to ADL, I'm 100% sure that you'd need far more power and probably LN2 to achieve that goal.

If you go by performance/watt as the only metric, then some little tiny CPU would beat the pants off the M1, A15, Intel, AMD, etc. Performance does not scale linearly with power.

Yes I’m aware of that and no I’m not misinterpreting the article. I’m aware of what it’s saying and what it isn’t. I’m aware of how Zen 3 scales and how Golden Cove scales and I know how ADL works and how it doesn’t. I know how to read a graph about multithreaded performance per watt and what that means and what it doesn’t mean. If you don’t take my word for it, please read @Gerdi ‘s post.

 

[JouniS]

They can absolutely serve consumers with high multithreaded throughput CPU needs.

They can serve high-end customers who want multithreaded performance, but they can't sell those chips to consumers. Because Apple makes balanced SoCs, the M1 Max is already huge, and the hypothetical M1 Max Duo would be twice as big. For every M1 Max Duo Mac Apple sells, they can't sell 8 M1 Macs (or iPhones), which will be reflected in the price.

Intel's Alder Lake chips are all smaller than the M1 Pro, which helps keeping the prices of chips with a large number of CPU cores at the level consumers are willing to pay.

 

[senttoschool]

The reason why they did so was to make up for the deficiencies in their big cores in multithreaded output. The reason die area matters is because it constrains both chip design and economics. I’m speaking from the perspective of the people building the chips. The big.Medium tactic was a way for Intel to compete. And it works! You seem convinced that I think ADL is bad. I don’t. I think the hybrid core design is a viable solution to the problem Intel found itself in and it worked really well. It allowed them to put far more cores onto the same die without busting power thus increasing their multithreaded throughput per die per area over AMD. If you have workloads that can take advantage of all those extra threads, that’s great!

But you’re attempting to extend your conclusions too far. That’s the issue. Let’s take an example:

Two chips with the same architecture and all the same cores, let’s say Intel 10th gen it doesn’t really matter, one with 24 cores, one with 4 cores. At the same wattage which chip will out perform the other in a multithreaded benchmark? Almost always across the power curve the 24 core chip will win with increasing wins as power goes up. Does that mean that the 24 core chip has a better microachitectural design? Nope they’re the same. That’s what @Gerdi ‘s post was about.

Trying to say Zen2/Zen3's architecture is better than ADL because the big core's die area is big is non-sense.

First, the big core in ADL is significantly faster than Zen3's. AMD isn't likely to catch ADL's ST until Zen5. In addition, the big cores have AVX-512 which takes up a ton of dead diespace. This suggests that Golden Cove is actually a superior design to Zen3.

Second, the advantage in big.Little or "big.Medium" is that you can lend more die space to the big cores in order to significantly boost ST while the little cores take care MT boost. And it shows. ADL's ST is roughly 33 - 35% faster than Zen3.

No one buys a chip because the chip has a big die space. If they did, no one would be buying M1 computers.

 

[crazy dave]

They can serve high-end customers who want multithreaded performance, but they can't sell those chips to consumers. Because Apple makes balanced SoCs, the M1 Max is already huge, and the hypothetical M1 Max Duo would be twice as big. For every M1 Max Duo Mac Apple sells, they can't sell 8 M1 Macs (or iPhones), which will be reflected in the price.

Intel's Alder Lake chips are all smaller than the M1 Pro, which helps keeping the prices of chips with a large number of CPU cores at the level consumers are willing to pay.

Sure until you pair it with a GPU to match? And then tack on the price of an OEM’s margins.

 

[senttoschool]

Yes I’m aware of that and no I’m not misinterpreting the article. I’m aware of what it’s saying and what it isn’t. I’m aware of how Zen 3 scales and how Golden Cove scales and I know how ADL works and how it doesn’t. I know how to read a graph about multithreaded performance per watt and what that means and what it doesn’t mean. If you don’t take my word for it, please read @Gerdi ‘s post.

Why? We have direct results in front of our eyes.

If you're buying an x86 laptop today, and all things are equal except the CPU, would you buy an ADL laptop or a Zen3 laptop?

I know my answer. There is no way in hell I'm buying a Zen3 laptop over ADL.

/typing from my M1 Pro 16"

 

[crazy dave]

Trying to say Zen2/Zen3's architecture is better than ADL because the big core's die area is big is non-sense.

First, the big core in ADL is significantly faster than Zen3's. AMD isn't likely to catch ADL's ST until Zen5. In addition, the big cores have AVX-512 which takes up a ton of dead diespace. This suggests that Golden Cove is actually a superior design to Zen3.

It’s not nonsense because it’s what stopped Tiger Lake, Ice Lake, Comet Lake, and especially Rocket Lake from competing with AMD at the high end of core counts. Golden Cove still has issues here. AVX-512 yes takes up a lot of dead space but die space was a problem even with Comet lake which didn’t have it. I’m not sure where you are getting this AMD won’t compete with ADL in ST until Zen 5 since Zen 4 isn’t out?

Second, the advantage in big.Little or "big.Medium" is that you can lend more die space to the big cores in order to significantly boost ST while the little cores take care MT boost. And it shows. ADL's ST is roughly 33 - 35% faster than Zen3.

This is quite literally the point I’ve been trying to make to you the entire conversation. That’s why the Goldmont cores exist.

No one buys a chip because the chip has a big die space. If they did, no one would be buying M1 computers.

I’m stuck in two conversations one where I’m arguing with you and @JouniS about die sizes and I’m finding this very amusing. Because unless I’m misreading your post here the two of you also disagree with each other.

 

[crazy dave]

Why? We have direct results in front of our eyes.

If you're buying an x86 laptop today, and all things are equal except the CPU, would you buy an ADL laptop or a Zen3 laptop?

I know my answer. There is no way in hell I'm buying a Zen3 laptop over ADL.

/typing from my M1 Pro 16"

Again you seem to be under the impression that I think ADL is a bad x86-64 product. I don’t. In fact I quite like their Gracemont cores. The dual decoders (brought in for Tremont) are a nifty solution to Intel’s x86-64 variable with decode issue and it’s great to see atom cores being used so effectively.

I’m just more restrained when it comes to my conclusions about microarchitecture and the future of each company’s designs and prospects in the market. I understand what these results *mean* and what they *don’t*. If you want to constrain the topic to which chip would I currently buy? Sure I’d probably buy ADL over Zen3 in most product tiers. But if you’re asking is AMD still competitive and can they respond in the near future? Yes and yes, quite well. You’re trying to draw conclusions unsupported by the data at hand.

 

[senttoschool]

I’m not sure where you are getting this AMD won’t compete with ADL in ST until Zen 5 since Zen 4 isn’t out?

I'm not saying AMD "won't". I'm saying it's unlikely. Because ADL leads in ST by 30-35%. New gens usually don't usually increase ST by that much.

Again you seem to be under the impression that I think ADL is a bad x86-64 product. I don’t. In fact I quite like their Gracemont cores. The dual decoders (brought in for Tremont) are a nifty solution to Intel’s x86-64 variable with decode issue and it’s great to see atom cores being used so effectively.

I’m just more restrained when it comes to my conclusions about microarchitecture and the future of each company’s designs and prospects in the market. I understand what these results *mean* and what they *don’t*.

Cool. I'm glad you're giving ADL its dues. It's only taken a few days and a ton of data to convince you to go from "Zen2 is more efficient than ADL" to ADL is better than Zen3.

Let Intel worry about the future of its micro architectures. I don't think you can make any conclusions about Intel not being able to scale their designs when Raptor Lake or Meteor Lake arrives.

 

[crazy dave]

I'm not saying AMD "won't". I'm saying it's unlikely. Because ADL leads in ST by 30-35%. New gens usually don't usually increase ST by that much.


Cool. I'm glad you're giving ADL its dues. It's only taken a few days and a ton of data to convince you to go from "Zen2 is more efficient than ADL" to ADL is better than Zen3.

Let Intel worry about the future of its micro architectures. I don't think you can make any conclusions about Intel not being able to scale their designs when Raptor Lake or Meteor Lake arrives.

That’s never been my position. I’ve never changed my position. You created a straw man. The point about Zen 2 still stands. The point about Golden Cove still stands. I never said ADL sucked. I said it’s hybrid design was to cover for Intel’s deficiencies in big core design when it comes to multithreaded throughput. That’s true and the hybrid design works.

AMD don’t have to increase their IPC by 30%. Run at ISO power Intel’s cores are not 30% ahead. That’s the point of the chips and cheese article. When people talk about standard performance increases between generations that’s what they are measuring.

 

[JouniS]

Sure until you pair it with a GPU to match? And then tack on the price of an OEM’s margins.

The point was that the buyer may not want a GPU to match. Selling only balanced configurations is obviously cost-effective if the customer wants one. The benefits of modularity become apparent when you want to emphasize one aspect of the system over others.

 

[senttoschool]

That’s never been my position. I’ve never changed my position. You created a straw man. The point about Zen 2 still stands. The point about Golden Cove still stands. I never said ADL sucked. I said it’s hybrid design was to cover for Intel’s deficiencies in big core design when it comes to multithreaded throughput. That’s true and the hybrid design works.

AMD don’t have to increase their IPC by 30%. Run at ISO power Intel’s cores are not 30% ahead. That’s the point of the chips and cheese article. When people talk about standard performance increases between generations that’s what they are measuring.

If you haven't changed your position, then your position is simply wrong.

Zen2 is far behind ADL in both performance and perf/watt. Scale Zen2 to ADL's performance level and Zen2 will draw significantly more power than ADL, and even then, it might even be impossible to get Zen2 to match ADL.

Intel no longer has a deficiency in MT. It matches the 5950x on the consumer level at a cheaper price. It's 30-35% faster than the 5950x in ST. It beats mobile Zen3 in ST and MT significantly.

It works. And it's a better architecture than both Zen2 and Zen3 at the consumer level. No matter how you twist it, it is true.

 

[EntropyQ3]

Not absolutely sure about this. Probably true since TSMC 5nm is probably more expensive than Intel's 7 and obviously more than TSMC 7nm (AMD). Also Apple has big caches which could mitigate their physical core size advantage. Someone adept at looking at die shots could probably estimate how big Apple's CPU complex is versus a similar AMD/Intel 8-10 core processor. Of course things like the system level cache for Apple is not being directly analogous to AMD/Intel 3rd level caches would complicate such a calculation. And we don't know exactly the costs (I think Sophie Wilson gave a talk about this where she said more than she probably should've, but I don't know the numbers off the top of my head and I think it might've stopped at TSMC 7nm). Having said that, since Apple is an OEM, no longer having to pay Intel or AMD's margins probably means the cost is still less for *them*. People seem convinced that Apple is probably paying less per chip than they were with Intel, but I'm not sure how concretely industry watchers know the relevant numbers especially the full SOC is hard to compare with what they were offering earlier.

There is little doubt Apple is paying less for the "computing part" per device than before. Solid numbers are basically impossible to nail down however, as the contracts aren’t in the public domain. Industry watchers (good ones) piece together snippets of information from customers, and then make a rough estimate. Which is typically good enough for their purposes, since such costs are only a small part of the bigger profitability picture analysts are typically interested in. Thus you will often see their assumed design costs and assumed production volume "baked in" when analysts say things about silicon costs on different nodes, which causes a lot of contradictory numbers to float around, all off which are based on better or worse assumptions.
Furthermore, when we are discussing Apple, they are a large enough customer that they can finance a whole production line for their supplier on their own - they are going to negotiate pricing on their own terms. What "the industry" pays, or what for instance the "contract pricing" tracked on DRAMXchange is doesn’t apply to them. They pay less. How much less? God knows.

 

[crazy dave]

The point was that the buyer may not want a GPU to match. Selling only balanced configurations is obviously cost-effective if the customer wants one. The benefits of modularity become apparent when you want to emphasize one aspect of the system over others.

Sure but Apple is the OEM. Apple has generally always offered the balanced build. This is true for Apple’s direct competitors because tautologically they are direct competitors. So from the *perspective of Apple* it’s no different except they don’t have to pay Intel and AMD’s margins. Apple doesn’t sell its chips direct to customers or businesses so the lack of modularity isn’t an issue. It’s not the same business. So yes they can increase core counts and yes those will come with increased other things too. That’s also true for buying most Dell, HP, Lenovo, etc … computers. Can you find the machines that cater to a specialized market? Yes of course but that’s not a standard consumer build either. The PC market is massive. But Apple doesn’t compete in all of it. For the segments it competes in it's reckoned to be paying less than it was before (@EntropyQ3 's post above).

I think we may be talking past each other here. The original point of @BigPotatoLobbyist ‘s post was about architecture. What would Apple’s dot on that graph look like with an increased number of cores? Far better. Could they do it? Yes. Will they do it? Yes but probably not for laptops because Apple so far has prioritized noise and battery life.

 

[leman]

Not absolutely sure about this. Probably true since TSMC 5nm is probably more expensive than Intel's 7 and obviously more than TSMC 7nm (AMD). Also Apple has big caches which could mitigate their physical core size advantage. Someone adept at looking at die shots could probably estimate how big Apple's CPU complex is versus a similar AMD/Intel 8-10 core processor. Of course things like the system level cache for Apple is not being directly analogous to AMD/Intel 3rd level caches would complicate such a calculation. And we don't know exactly the costs (I think Sophie Wilson gave a talk about this where she said more than she probably should've, but I don't know the numbers off the top of my head and I think it might've stopped at TSMC 7nm). Having said that, since Apple is an OEM, no longer having to pay Intel or AMD's margins probably means the cost is still less for *them*. People seem convinced that Apple is probably paying less per chip than they were with Intel, but I'm not sure how concretely industry watchers know the relevant numbers especially the full SOC is hard to compare with what they were offering earlier.

What I meant more were the basic system design choices, like UMA or large integrated GPUs which inevitably push the cost up. Also, M1 Pro is around 20% larger die-wise than ADL…

There were a lot of people in this thread who denied that ADL was as efficient or more efficient than Zen3. @crazy dave @Andropov

Hardware Unbox literally just showed that ADL under more optimal power curves trumps Zen 3.

Because ADL has more cores. You are comparing a 6+8 design to a 8-core design on tests that are pretty much optimal for multiple cores. You can always improve efficiency by throwing more lower clocked cores at the problem. But that does not make your cores more efficient. Basically, this is exactly the same situation thst we had in AMD vs. Intel but reversed. Folks were praising AMD in 15 and 30W bracket because it did so well in throughput benchmarks compared to Intel, but that was 8 cores vs. 4. Now Intel did exactly the same thing.

But at the end of the day, we do have single-core results and we do know that Golden Cove is significantly less power efficient than zen3. At the same time though, ADL is faster. That aspect of Intel vs. AMD has not changed in years. AMD Zen was always slightly slower, but 30% or some more efficient than Intel cores, and AMD had the tech and yields to ship 8 cores even at the low end. Now Intel put more low-power cores to close the gap, and that’s exactly what we see.

There is no innovation here. No progress. No interesting technical solutions. Just “oh yeah? then I bring more dudes and we’ll mess you up”. Frankly, Apples refusal to play on this level is one of the main reasons why I like their architecture so much. They focus on things that actually matter for real users.

 

[crazy dave]

If you haven't changed your position, then your position is simply wrong.

Zen2 is far behind ADL in both performance and perf/watt. Scale Zen2 to ADL's performance level and Zen2 will draw significantly more power than ADL, and even then, it might even be impossible to get Zen2 to match ADL.

Intel no longer has a deficiency in MT. It matches the 5950x on the consumer level at a cheaper price. It's 30-35% faster than the 5950x in ST. It beats both mobile Zen3 in ST and MT significantly.

It works. And it's a better architecture than both Zen2 and Zen3 at the consumer level. No matter how you twist it, it is true.

You’re not understanding what I’m trying to say. Maybe if this were face to face with a whiteboard I could get it through. I’ll try one more time then call it quits. I think the issue here is that there are three separate arguments which are being conflated together:

1) Holistic analyses of the products themselves and the cost/benefit of buying a system vs another.

2) What conclusions about the underlying architectures we can draw from those analyses - why have they been built the way they have been built.

3) What predictions we can make about the strengths and weaknesses of each chip design moving forwards and what it means for each company.

Let's go back to the beginning. You were arguing with @Andropov and I about the nature of Intel's chip design - i.e. number 2. We were trying to explain to you the purpose behind the hybrid design - that Intel's P-core design is bigger and more power hungry than its Zen 3 counterpart. It achieves its ST performance by huge die are and running more watts. This has meant that since Comet Lake (no AVX-512) vs Zen 2, the core count of Intel's chips have been its achilles heel in multithreaded throughput. AMD has been consistently able to either offer more cores at better efficiency than Intel for both laptop and desktop. This culminated in the infamous "waste of sand" launch known as Rocket Lake which had to both lower its core counts at the high end and blew up thermals.

To combat this trend, Intel came up with a new strategy: Don't compete head-to-head on performance cores in multithreaded throughput. Instead, keep a smaller number of performance cores on die but add Atom cores as a medium core to make up the difference. The downside is the complexity this adds to thread scheduling and core co-design - problems which, with a couple of notable exceptions, Intel and MS solved before launch successfully. The result has meant that for 1) the products as they stand today: Intel can keep its power hungry performance cores big and toasty but also dramatically improve its multithreaded throughput over most Zen 3 products. This was and always has been my position. I never said ADL as an idea sucked. It was a narrow argument that Zen 3 cores maintain an uarch advantage over Golden Cove when it comes to perf efficiency and die area which is crucial to multithreaded output. It determines how many and how fast cores you can add. So to counter that Intel complements Golden Cove with Atom cores to overcome this deficit. I fully acknowledge that this was successful and agree that especially in the enthusiast end of the spectrum, if you are buying a chip in order to run large numbers of threads, ADL has clear advantages over Zen 3 in mobile and desktop.

However! Now we're on the argument 3: the future. Does this spell the doom of Zen and AMD? No. AMD still has exploitable advantages in uarch design - while Atom puts some armor over it, Intel's deficiencies in P-core remain their Achilles' heel. (I should add that yes the P-core is better than it used to be - while not as nifty as their E-cores, Intel finally made improvements to it) AMD's disadvantages are that while Intel's missteps have allowed it to grow far bigger than it had been, it's is still much smaller. Can it get the fab space it needs? Can they (economically) increase their own die area/number of dies to combat Intel in the high core/thread count enthusiast market? Will Zen 4 be as successful as Zen 3 over Zen 2 or Zen 2 over Zen 1? Basically you're taking correct analyses from 1 and overdrawing conclusions about 2 thus making overwrought predictions in 3.

 

[crazy dave]

What I meant more were the basic system design choices, like UMA or large integrated GPUs which inevitably push the cost up. Also, M1 Pro is around 20% larger die-wise than ADL…

Absolutely, but then you have to consider adding a GPU's die area to ADL ... i.e. if we're going to consider the whole package, then we have to consider the whole package. Basically is Apple paying more for the similar solution that they would have to get if they went out and bought AMD GPUs and Intel CPUs? Probably not. That's where I was coming from in my response - also see @EntropyQ3 's post.

Because ADL has more cores. You are comparing a 6+8 design to a 8-core design on tests that are pretty much optimal for multiple cores. You can always improve efficiency by throwing more lower clocked cores at the problem. But that does not make your cores more efficient. Basically, this is exactly the same situation thst we had in AMD vs. Intel but reversed. Folks were praising AMD in 15 and 30W bracket because it did so well in throughput benchmarks compared to Intel, but that was 8 cores vs. 4. Now Intel did exactly the same thing.

But at the end of the day, we do have single-core results and we do know that Golden Cove is significantly less power efficient than zen3. At the same time though, ADL is faster. That aspect of Intel vs. AMD has not changed in years. AMD Zen was always slightly slower, but 30% or some more efficient than Intel cores, and AMD had the tech and yields to ship 8 cores even at the low end. Now Intel put more low-power cores to close the gap, and that’s exactly what we see.

There is no innovation here. No progress. No interesting technical solutions. Just “oh yeah? then I bring more dudes and we’ll mess you up”. Frankly, Apples refusal to play on this level is one of the main reasons why I like their architecture so much. They focus on things that actually matter for real users.

Just forewarning, I've been trying to explain this to him ... you're basically retreading the last 3 hours of my conversation with him. Although you are far harsher on both AMD and Intel than I.?