When ARM is the PC CPU.

[Edward78]

I wonder when the PC move to AEM will really happen? Would love to put a ARM equal to my Ryzen 7 5800x in a cas like this & it keep cool. Sorrt of topicc, wish it was deep enough for a 2 fan GPU though.

[Eigenvektor]

When legacy software that requires x86 has been replaced by legacy ARM based software.

 

What makes you think there will be or has to be such a move in the first place?

[danalog]

ARM will replace x86 when Linux replaces Windows (it's gonna happen any day now guys trust me I promise) 

When/if it ever happens, an ARM CPU as fast as a 5800X will be fully irrelevant. Like, we have ARM chips right now that perform like midrange x86 CPUs and still no one really cares 

[Needfuldoer]

1 hour ago, Eigenvektor said:

When legacy software that requires x86 has been replaced by legacy ARM based software.

Or once emulation gets good enough to run legacy x86 software seamlessly on ARM (which also necessitates ARM being fast enough to run x86 code as fast as said code would run on an x86 of its time).

[Eigenvektor]

43 minutes ago, danalog said:

Like, we have ARM chips right now that perform like midrange x86 CPUs and still no one really cares 

That's always been the issue of the second one to market. You don't switch to get the same thing, just with a different color. You switch if it gives you significantly more than you have right now, enough to make it worth the pain of switching

[Jurrunio]

By the time an ARM CPU matches your 5800x, x86 would have made progress over what they are today too. It's never only about ARM getting better, but the gap between x86 and ARM. 

[manikyath]

ARM will have to beat x86's raw single core performance before x86 figures out how to beat ARM's efficiency for them to have any chance of winning.

 

oh, and they'll have to do so before RISC-V becomes a much more interesting opportunity. ARM licensing is as troubled as x86 licensing is, so essentially other than efficiency ARM has nothing to offer.

[TudorFinalBosz]

ARM will lead to more fragmentation.

I don't fully understand why the megacorps are pushing for ARM so aggressively, could be multiple reasons: to make mobile devices even more appealing in terms of performance, to make battery last longer, maybe to compete with Apple's M series in terms of performance/powerdraw, to simplify compilation and builds cost, what else. ARM CPUs might be cheaper and have lower licensing costs too.

 

They have no chance in hell of beating x86 on sheer performance, though, if power is no object and devices are plugged.

 

[Electronics Wizardy]

3 hours ago, manikyath said:

ARM will have to beat x86's raw single core performance before x86 figures out how to beat ARM's efficiency for them to have any chance of winning.

 

Hasn't that already happened? Apple's Big cores are on par or better than the current gen intel or AMD in single threaded perf. Depends on application and instructions used, but the x86 performance advantage is gone I'd say.

[Kilrah]

4 minutes ago, Electronics Wizardy said:

Apple's Big cores are on par or better than the current gen intel or AMD in single threaded perf.

Problem is that's on Mac OS and there's nothing else that can "correctly" be run on these machines to determine whether it's really due to the CPUs or software optimisation shenanigans. 

[Electronics Wizardy]

1 minute ago, Kilrah said:

Problem is that's on Mac OS and there's nothing else that can "correctly" be run on these machines to determine whether it's really due to the CPUs or software optimisations. 

Can you show more info about these optimization? And from the looks of things Linux is often faster M systems than macOS, showing that there likely isn't some secret sauce that makes the M chips fast, there a well designed core given lots of die space. 

 

https://www.phoronix.com/review/apple-m1-linux-perf/3

 

I think its hard to say its software optimizations when you can compile code on macOS and it runs faster than a Windows/Linux system  on x86 hardware. 

 

 

[manikyath]

35 minutes ago, Electronics Wizardy said:

Hasn't that already happened? Apple's Big cores are on par or better than the current gen intel or AMD in single threaded perf. Depends on application and instructions used, but the x86 performance advantage is gone I'd say.

i'm not sure how much of a benchmark apple's big cores are for what can be expected for the rest of ARM's future in computers. it's probably more of an advertisement for unified memory and fully integrated CPU/GPU platforms, than it is for ARM specificly.

[Electronics Wizardy]

22 minutes ago, manikyath said:

i'm not sure how much of a benchmark apple's big cores are for what can be expected for the rest of ARM's future in computers. it's probably more of an advertisement for unified memory and fully integrated CPU/GPU platforms, than it is for ARM specificly.

Looking at CPU performance alone, there typically on par or better than the top x86 desktop/server chips. Unified memory or the iGPU isn't helping in those tests. Apple has been putting out very impressive CPU cores.

 

 

[manikyath]

14 minutes ago, Electronics Wizardy said:

Looking at CPU performance alone, there typically on par or better than the top x86 desktop/server chips. Unified memory or the iGPU isn't helping in those tests. Apple has been putting out very impressive CPU cores.

it is because memory bandwidth is VERY high in highly integrated chips like apple's. that's the advantage of a design that's so vertically integrated, but unless we give up a lot of customizability (see: the ryzen 'AI' processors that even in the framework have soldered memory) we're not going to get apple levels of performance out of ARM on the desktop.

 

i'm not denying that apple's performance numbers are impressive, what i'm saying is that it's not a direct comparison because there's other design choices that give them quite a leg up as well.

[Electronics Wizardy]

19 minutes ago, manikyath said:

it is because memory bandwidth is VERY high in highly integrated chips like apple's. that's the advantage of a design that's so vertically integrated, but unless we give up a lot of customizability (see: the ryzen 'AI' processors that even in the framework have soldered memory) we're not going to get apple levels of performance out of ARM on the desktop.

 

i'm not denying that apple's performance numbers are impressive, what i'm saying is that it's not a direct comparison because there's other design choices that give them quite a leg up as well.

Aren't both M4 and Ryzen AI 300 using LPDDR5x memory? I think clocks are a bit higher on Ryzen AI 300(8000 vs 7500) I think the memory system of those chips are fairly similar so memory can't be used to explain why the m4 is faster. 

 

I also think a single core can't use all the memory bandwidth, so single threaded tasks should benefit from the higher memory bandwidth. The bigger M chips have much more memory bandwidth, but they also have more cores and a bigger GPU to use them.

[starsmine]

Why do you want ARM?
Arm isnt faster.
arm isnt more power efficient. 

The general consumer, hell the general power user should not care about what ISA the chip is using, it means nothing outside of who designed the chip. 
  

4 hours ago, Electronics Wizardy said:

Aren't both M4 and Ryzen AI 300 using LPDDR5x memory? I think clocks are a bit higher on Ryzen AI 300(8000 vs 7500) I think the memory system of those chips are fairly similar so memory can't be used to explain why the m4 is faster. 

 

I also think a single core can't use all the memory bandwidth, so single threaded tasks should benefit from the higher memory bandwidth. The bigger M chips have much more memory bandwidth, but they also have more cores and a bigger GPU to use them.

Yea and strix halo is quad channel. 

And I wouldnt say a single core cant use all that bandwidth. it depends on the work load. Like I can make a work load that is bandwidth bound on just a single core as soon as you escape L3 to 1/8th of the performance. 


What is going on when it half fills up L2?
NO idea. Literally none. 

Anyways, thats single core, duel channel 6000MT/s zen 4. More memory bandwidth for data sets that cross the LLC line there would be a 1 to 1 speed up in performance. its 100% bandwidth limited there. If I could run quad channel ddr5 I could get close to 20GFLOPS in Single precision on that right tail. 

[twin_savage]

WRT ARM vs x86 core performance:

I've got a cross platform finite element analysis workload benchmark I've been having people run that uses the most optimized math libraries for Apple's arm64, Intel/AMD's x86_64 or ARM's aarch64 so that it shows how the different cores stack up against each other using the more efficient code on a given platform:

 

 

It would be reasonable to compare the Apple M1 Ultra with the Intel Xeon W5-3435X or Ryzen 7950X3D, they all have 16 P-cores and came out within a year of each other. The CFD-only benchmark is not memory bound, but the other two benchmark columns are.