Question x86 and ARM architectures comparison thread.

[poke01]

Good comparison but Apple had to go through a lot more trouble for that performance at 60W whereas AMD is doing it AFTER they got done with their main market, i.e. server. It's pretty close and Apple seems to be at a disadvantage because their design cost a lot more money and man-hours than AMD's.

How do you know Apple had to go to more trouble? You got proof otherwise these are just assumptions.

I could also say it’s more likely Apple didn’t go to lot of trouble as they just scaled up the CPUs they use for the iPhones, their A18 P cores to make it suitable for laptops ie higher clocks and more cache and more I/O.

As for cost yes N3E is more expensive than N4P but guess what Apple can afford it. AMD is also going to use N2P for Zen6 as is Apple for M6. This is a bad argument. For Apple the use of N3E is negligible as they sell the product not the chip.

it's just these two benchmarks that we use to compare x86 and ARM, yes, Apple has an edge at a much higher cost (design, manufacturing and end-user cost). There's also the question of whether Apple can continue getting IPC gains. Things will be pretty interesting to compare between Zen 6, M5 and Snapdragon Elite X Gen 2

Nah, just not these two benchmarks I got more. Anyway, SPEC2017 has all the answers. That David Huang SPEC chart is all very much correct.

 

[igor_kavinski]

AMD is also going to use N2P for Zen6 as is Apple for M6.

Oh no. Apple M5 is going to get compared to Zen 6. Let's not jump ahead to M6 which should be a Zen 7 competitor.

 

[511]

Apple M5 is N3P Apple M6 is Zen 6/NVL Competitor on N2 all 3 design except for NVL-Mobile.

 

[poke01]

How do you know Apple had to go to more trouble? You got proof otherwise these are just assumptions

Adding onto this, Apple spends less time on their CPU architectures than AMD does cause they release them yearly, they kind of have too cause of the iPhone. So Apples Sillicon team got a strict schedule.

So does Qualcomm and ARM for Android flagships. Intel also releases yearly architectures because the laptop space commands huge volumes.

 

[igor_kavinski]

Adding onto this, Apple spends less time on their CPU architectures than AMD does cause they release them yearly, they kind of have too cause of the iPhone. So Apples Sillicon team got a strict schedule.

That doesn't mean that Apple doesn't have a much bigger silicon engineering team.

If AMD had the kind of R&D money to throw at the problem that Intel used to have, we would see even more progress from them.

 

[511]

If AMD had the kind of R&D money to throw at the problem that Intel used to have, we would see even more progress from them.

Intel throws Money at Fab and Process don't forget that something none of other companies has to do cause TSMC does it for them not to mention Intel is multi sourcing this adds to cost as well

 

[igor_kavinski]

Intel throws Money at Fab and Process don't forget that something none of other companies has to do cause TSMC does it for them not to mention Intel is multi sourcing this adds to cost as well

Pretty sure Intel appreciates you whiteknighting them. Hope you receive a thank you email from them soon

 

[511]

Pretty sure Intel appreciates you whiteknighting them. Hope you receive a thank you email from them soon

I will DM you the first mail i get from them

 

[igor_kavinski]

@poke01 Thank you!

50 minutes of x264 AVC and only 10% battery utilized.

Fascinating!

 

[Geddagod]

That doesn't mean that Apple doesn't have a much bigger silicon engineering team.

If AMD had the kind of R&D money to throw at the problem that Intel used to have, we would see even more progress from them.

I doubt AMD's core design team is funded well enough. There's bound to a point of diminishing returns, and there's a bunch of relatively small 'startups' working on new core IP as well, showing you don't need billions of dollars to design new CPU archs.

whereas AMD is doing it AFTER they got done with their main market, i.e. server.

I think people vastly underestimate the strength of the client market. The only reason these cores even hit close to 6GHz, and the bother to develop the classic cores rather than just dense cores at all, is because of client. One just has to look at Intel's CCG revenues and operating income to see how big and profitable of a market it is.
AMD undoubtedly strikes a pretty good balance between client and server for their new architectures. I don't think it's mainly server.

 

[igor_kavinski]

I don't think it's mainly server.

Maybe but their focus on server makes them design their cores carefully to not be crazy power guzzlers and this helps them design better client cores.

 

[Schmide]

M4 Max has 546GB/s of bandwidth, which is above the 410GBs threadripper wrx90; (maybe equal if Wendell can get the right timings), Comparing M4 MAX to AM5 sockets is really punching down from its weight class. It's basically double strix halo as well (256GBs).

If you melded 2 strix halos as in the same way as the M4 Max that would be similar.

 

[gdansk]

M4 Max has 546GB/s of bandwidth, which is above the 410GBs threadripper wrx90; (maybe equal if Wendell can get the right timings), Comparing M4 MAX to AM5 sockets is really punching down from it's weight class. It's basically double strix halo as well (256GBs).

If you melded 2 strix halos as in the same way as the M4 Max that would be similar.

It's also N3 with an N3 memory controller. They're comparing it to Granite Ridge at 80W, which is a N4 with a N6 memory controller. Basically I see it as "Apple has more money" comparison, again. Which would probably make AMD panic if they felt they competed with one another. Ultimately it seems they don't which means AMD can lazily move along to produce an M4 Pro for the x86 world in 2027.

 

[Geddagod]

Maybe but their focus on server makes them design their cores carefully to not be crazy power guzzlers and this helps them design better client cores.

AMD arguably looks better in client than they do in server in comparison to Apple silicon, from the CPU side.
The trademark of Apple is much better power. The extremely high ST perf is just the cherry on top.

 

[gdansk]

AMD arguably looks better in client than they do in server in comparison to Apple silicon, from the CPU side.
The trademark of Apple is much better power. The extremely high ST perf is just the cherry on top.

Really? AMD's most power efficient part this generation is only available in server.
And no one has ever done a comparison of Turin D vs an M4 part that doesn't exist at fixed power because it's impossible.

 

[johnsonwax]

Someone or some company that has written a custom Apple Silicon application that scales very well with more cores.

But that's just my guess. You give more cores. More use cases get invented. This is what developers do. Use hardware if it's available.

Ok, and what's the addressable market by that application? The cost just to tape out is $100M+. So can that application generate a billion in revenue for Apple?

But that's also pretty much not how it works. When I was working and needed something that scaled like that I was renting DC. Almost everything that scales like that is server. And since Apple doesn't sell silicon, what DC is going to going to buy Apple racks rather than their own designs? There's no market there for Apple.

 

[johnsonwax]

Good comparison but Apple had to go through a lot more trouble for that performance at 60W whereas AMD is doing it AFTER they got done with their main market, i.e. server. It's pretty close and Apple seems to be at a disadvantage because their design cost a lot more money and man-hours than AMD's.

First, I don't know why you have such a boner for MT when 95% of consumer use is not benefitted much by MT. The reason everyone feels the Apple Silicon is faster is because most applications, basically all games that are CPU bounded, are single-core constrained. Even Excel doesn't benefit often from MT. Note: Apple is a consumer facing company, not an enterprise facing one.

If it's just these two benchmarks that we use to compare x86 and ARM, yes, Apple has an edge at a much higher cost (design, manufacturing and end-user cost). There's also the question of whether Apple can continue getting IPC gains. Things will be pretty interesting to compare between Zen 6, M5 and Snapdragon Elite X Gen 2.

Apple didn't get there at a higher cost though. Almost certainly they got there at a lower cost. Why? Because they have what, at most 6 variants of M4 plus M3 Ultra? Currently listed on AMDs website, so presumably all current offerings: 8 variants of the 9000 series, 6 of the 8000 series, 13 of the 7000 series, 26 of the 5000 series, and 5 of the 4000 series. That's 58 different variations to cover the space Apple is doing with 7? And Apple is getting half their tablet line covered as well. And the Mac makes about ⅓ more revenue annually than all of AMD does.

Hey, AMD why you working so hard for such garbage financial results? You are so determined to see Apple in a bad light that you make these arguments before you've even thought them through. That's a LOT of design work and supply chain cost and management that AMD is doing just to make their business work that Apple doesn't need to do.

And you say that Apple is working harder even on the design, but I don't think they are, mainly because they aren't fighting against other parts of their own business. I guarantee there's a laundry list of things AMD wishes they could do to make their lives much easier if they could force OEMs to change how they work, force Microsoft to change what they do and so on. Those are the easy technical paths that they don't get to take because some business decision is more important so they have to jump through all kinds of technical hoops, hoops that Apple doesn't need to jump through - they can just go directly at the problem in the most straightforward way all the way to 'hey, we need to change the ISA'. Apple can straight up do that. They can take Apple Silicon to RISC-V if that would be the performant path. There's no universe AMD could do that - look at how hard it is for Qualcomm/Microsoft to get ARM working in that market.

 

[dacostafilipe]

Just watched the latest video from STH and had to think about this thread

Somehow he went over multiple points discussed here, is he reading us?

 

[johnsonwax]

That doesn't mean that Apple doesn't have a much bigger silicon engineering team.

Estimates I've seen is that AMD after they bought Xilinx had about 5000 silicon engineers to Apple's 1000 and that AMD had about 2000 before the acquisition. That headcount was likely going to get reduced (it usually is after acquisitions) but nowhere near 1000. Again, Apple has a VASTLY smaller portfolio of silicon components, but they are similarly diverse at least to pre-Xilinx AMD. Apple is making state of the art cellular modems which are a particularly tricky product, they both make CPU, GPU, NPU, etc. Apple has their wearable silicon, U1, R1 their real time low latency video processing unit, etc. Apparently they're now making a server AI chip for in-house use.

 

[Doug S]

I'm not convinced. Why not core spam and offer even higher priced products?

Because you have to be selling into a market large enough to make that worth the effort. Sure Apple could make something with 128 cores if they wanted, but if they only sell 1000 of them that profit from selling it doesn't pay back the product development cost.

They need to stitch two Max dies to get higher than 16 cores and that's expensive for them. That's why they are forced to offer previous gen Ultra chip. And then they price it so high that people with even great interest are forced to think twice about plopping down so much hard cash.

Apple is good at pretending to offer more value than they actually do. That's their secret.

Because the Max is their largest die, and that size is dominated more by GPU cores than CPU cores. They aren't going to tape out a die without GPU like Intel/AMD do because they don't sell servers and they don't support dGPU. If they took out the GPU cores they'd have room for nearly 80 P cores at the same size die. That is "more value" in your mind I suppose, but it would cost tens of millions to tape that out. If they did, what would they put it in and who would buy what they put it in? I mean sure you could build an "Ultra" that's got only Max level GPU but 3x current Ultra CPU. There are some people who might want it, but likely not enough - especially when Apple seems pretty lukewarm about Ultra in the first place.

Now with Apple making their own servers they are going to need custom dies for that, but likely we won't ever see them. Maybe we'll get hints from stuff that shows up in macOS to support those servers/chips, but they aren't going to sell servers or sell server like workstations that support a dGPU.

 

[OneEng2]

Busy thread since last night!

First, the benchmarks using blender are not only using the CPU cores IIRC. They are also using the GPU.

Secondly, M4 is within spitting distance of Zen 5 ..... from a full die shrink ahead on the CPU and two die shrinks ahead on the IOD.... and only at the same power level.

Not impressive IMO if you are attempting to highlight the architectural merits of M4 over x86 Zen 5.

Things may heat up really quick if AMD also debuts their own ARM effort (Sound Wave APU).

But still, I think it won't win in volume because x86 chips hit a lot of the price points that actually matter to most people. I can't still see Snapdragon Plus laptops selling for $500 yet even though it's supposed to be a trash SKU.

I also find it tough to believe that AMD's first ARM will not be a learning experience for them and will likely not measure up... but that doesn't mean they wont get there within a couple of generations though.... and possibly then some.

I don't agree with this cause we can use x86 core running Linux via SPEC2017. AMD is able to optmise it cores as well with Linux as Apple does with macOS

I am having a tough time finding comparisons between M4 and Zen 5. Do you have any?

Wow you compare something with 10 cores (many of them E cores) with something that has 16 cores and we're supposed to be impressed that the one with 16 cores beats the one with half as many P equivalent cores? That "bunch of other stuff" is doing a LOT of heavy lifting in your comparison. Zen 5c is useless without a "chipset", which the M4 doesn't need. So if you want to do a 10 core vs 16 core comparison I say let's do just the 80 mm^2 chip versus the 73 mm^2 chip comparison next. I predict Zen 5c's performance will be 0, because it can't do anything at all on its own without a whole other chip you decided not to include when comparing mm^2.

Wow, you compare something on N3E with something on N4P and N6?

Show me some DC benchmarks on that M4. Just find the highest core count DC M4 you can locate and compare it to the highest core count Zen 5 you can find.

I predict that the M4's performance will LOOK close to 0 compared to a 192c Zen 5c in a dual socket setup.

I think people vastly underestimate the strength of the client market. The only reason these cores even hit close to 6GHz, and the bother to develop the classic cores rather than just dense cores at all, is because of client. One just has to look at Intel's CCG revenues and operating income to see how big and profitable of a market it is.
AMD undoubtedly strikes a pretty good balance between client and server for their new architectures. I don't think it's mainly server.

AMD has stated on many occasions that their designs are "Server First". Desktop and client get hand-me-downs .

It's also N3 with an N3 memory controller. They're comparing it to Granite Ridge at 80W, which is a N4 with a N6 memory controller. Basically I see it as "Apple has more money" comparison, again. Which would probably make AMD panic if they felt they competed with one another. Ultimately it seems they don't which means AMD can lazily move along to produce an M4 Pro for the x86 world in 2027.

I agree.

I am wondering how the comparison will look when AMD is a full node ahead? Lets compare the Zen 6 on N2 to the M4 on N3E.

Apple does amazing things with their ARM processors, but to dream that this somehow makes ARM fundamentally superior to x86 is just silly IMO.

 

[S'renne]

You guys using M4 to compare with a DC CPU in MT performance when all known benchmarks from Geekerwan and David Huang and others already confirms that it is a client focused architecture, where its extremely high ST and low ST power draw with high bandwidth on that metric is just...might as well use the Strix Point to compare against an EPYC for the same points to argue instead of an M4

 

[poke01]

First, the benchmarks using blender are not only using the CPU cores IIRC. They are also using the GPU

The tests I referenced are all CPU blender tests.

Secondly, M4 is within spitting distance of Zen 5 ..... from a full die shrink ahead on the CPU and two die shrinks ahead on the IOD.... and only at the same power level

What? It’s not the same power level. M4 Max is using 60 watts meanwhile the rest of the Zen5 mobile processors are above or at 80 watts.

The initial point you made was that Zen5 can beat the M4 with SMT by a large margin. Uh, no it can’t unless you feed ton of power but then the x86 cores lose laughably in perf/w.
The node isn’t the deciding factor here, it’s the microarchitecture being a LOT better.

It’s funny isn’t when AMD beats Intel using a older node, it goes “wow, AMD is beating Intel using an older node”, but when other companies have better microarchitectures on a better node and mind you N3E isn’t all that different from N3B, the argument becomes, “well AMD is on a an older node”.

Show me some DC benchmarks on that M4. Just find the highest core count DC M4 you can locate and compare it to the highest core count Zen 5 you can find.

I predict that the M4's performance will LOOK close to 0 compared to a 192c Zen 5c in a dual socket setup

I don’t need to, a Zen5 core is losing pathetically in a laptop environment.

There are no M4 DC products because Apple doesn’t care for that market just as there are no Zen5 powered smartphones because AMD doesn’t care. But where these architectures converge is laptop so we can do some comparisons here.

Wow, you compare something on N3E with something on N4P and N6?

Because you can. Why can’t you compare client products to Apple's N3E but can for Intel's N3B?

I am wondering how the comparison will look when AMD is a full node ahead? Lets compare the Zen 6 on N2 to the M4 on N3E.

Apple does amazing things with their ARM processors, but to dream that this somehow makes ARM fundamentally superior to x86 is just silly IMO.

Sure go ahead, but Apple will be on M6 around the time Zen6 is released.

No, this doesn't make ARM better than x86. It makes Apple's microarch M4 P core better than AMD's Zen5 core.

 

[S'renne]

The tests I referenced are all CPU blender tests.


What? It’s not the same power level. M4 Max is using 60 watts meanwhile the rest of the Zen5 mobile processors are above or at 80 watts.

The initial point you made was that Zen5 can beat the M4 with SMT by a large margin. Uh, no it can’t unless you feed ton of power but then the x86 cores lose laughably in perf/w.
The node isn’t the deciding factor here, it’s the microarchitecture being a LOT better.

It’s funny isn’t when AMD beats Intel using a older node, it goes “wow, AMD is beating Intel using an older node”, but when other companies have better microarchitectures on a better node and mind you N3E isn’t all that different from N3B, the argument becomes, “well AMD is on a an older node”.

I don’t need to, a Zen5 core is losing pathetically in a laptop environment.

There are no M4 DC products because Apple doesn’t care for that market just as there are no Zen5 powered smartphones because AMD doesn’t care. But where these architectures converge is laptop so we can do some comparisons here.


Because you can. Why can’t you compare client products to Apple's N3E but can for Intel's N3B?



Sure go ahead, but Apple will be on M6 around the time Zen6 is released.

No, this doesn't make ARM better than x86. It makes Apple's microarch M4 P core better than AMD's Zen5 core.

Exactly, but all they are comparing is DC MT which is irrelevant to individual core performance for architecture comparisons

 

[gdansk]

And no one is comparing N3E Zen 5, which does exist in compromised form, to M4 either.

At the end of the day no one else in the ARM world is doing anything close to Apple yet either. These threads always become useless because it's just Apple fangirls running in circles about how good their architecture is when made on always the most bleeding edge process and extrapolating that to ARM in general.

And then extrapolating without foundation that it'd do well in servers too. I'm sure it would, but would it do better than other mythical parts which don't exist? Who knows.