Why Intel and AMD don't make chips like the M2 Max and M2 Ultra

[bobcomer]

The fact that some can’t keep to a topic, is not a reason for you to do so.


With respect, you haven’t shown you comprehend the point.

Is your mind capable of changing?

Given the efficiencies in dividing work, dependencies and other issues. I can assure you, for anything you use a computer for, one 5ghz core will be faster than two 2.5ghz cores.

You're getting WAY too personal and attacking. Stick to the argument, not the people.

And yes, my mind can be changed, quite easily -- with a good counterargument. I'm nothing if not flexible that's part of the job. I have seen nothing in this thread to convince me that single core benchmarks are the greatest measure of a processor. Quite the opposite actually. Even by you own words.

And I still disagree with the last statement. You just saying something is true, doesn't prove anything either.

 

[bobcomer]

Two 2.5Ghz cores are not going to get more work done than a single 5Ghz core. In an ideal scenario, you’d get the same amount of work done. At worst, you’d get half that.

You haven't proven that to me, and I know that's not the case in my experience either. But then I'm not singletasking.

 

[APCX]

You're getting WAY too personal and attacking. Stick to the argument, not the people.

Nothing personal. Job observations from your responses.

And yes, my mind can be changed, quite easily -- with a good counterargument.

They’ve been shown to you, with reason and logic. You just don’t understand or like the answers.

I'm nothing if not flexible that's part of the job. I have seen nothing in this thread to convince me that single core benchmarks are the greatest measure of a processor. Quite the opposite actually. Even by you own words.

Then you misunderstood my words.

And I still disagree with the last statement. You just saying something is true, doesn't prove anything either.

Indeed, but showing some reasoning should show you that. Alas I fear it’s futile.

 

[APCX]

You haven't proven that to me, and I know that's not the case in my experience either. But then I'm not singletasking.

Whether you’re “singletasking” ( a bizarre statement) has no relevance to whether single core speeds are more beneficial than multi core processing.

Your bizarre phrasing betrays a complete lack of understanding.

 

[leman]

You haven't proven that to me, and I know that's not the case in my experience either. But then I'm not singletasking.

It’s a basic fact of computing. Since you mentioned that you develop software I’m a bit puzzled that this is something you require proof for.

 

[bobcomer]

Nothing personal. Job observations from your responses.

What you are saying is very personal toned. Compare to what I say, when I have even mentioned you at all? See the difference in argument? Please stop it and stick to the argument only. At least if you want me to read what you write in the future. If you don't, fine, I'll just block you and not read anything you ight say, good or bad.

Whether you’re “singletasking” ( a bizarre statement) has no relevance to whether single core speeds are more beneficial than multi core processing.

Your bizarre phrasing betrays a complete lack of understanding.

Well, that does it, bye.

 

[APCX]

What you are saying is very personal toned. Compare to what I say, when I have even mentioned you at all? See the difference in argument? Please stop it and stick to the argument only. At least if you want me to read what you write in the future. If you don't, fine, I'll just block you and not read anything you ight say, good or bad.

Well, that does it, bye.

Ok have a good day.

 

[R2DHue]

Like I said before, since when does a thread here stick to the original topic. That's an irrelevant complaint, don't like it, don't argue it. I got in this thread because I didn't agree with that point *at all*, and I haven't seen anything to change my mind.

And to prove me wrong about single core, you would have to be able to prove that a single core processor can do more work than a multicore processor

I admit I’ve never understood why single-core benchmark scores are given so much consideration.

When does a 4 or 6 core CPU use only one core? (I seriously don’t know the answer.)

I can see how it might be true in the case of computational tasks that don’t need very much horsepower, or in a case where the OS or controller is trying to reduce power consumption and heat, but, in Apple Silicon, tasks like this would probably instead be handed to the Efficiency CoreS — which, as the name implies, are multiple cores, too.

If there are times when software is executed on a single core, I’m interested in learning more about it and why.

And the most I know about CPU vs. GPU is that multi-core CPUs and multi-core GPUs are very different in terms of symmetric vs. parallel processing, respectively.

I don’t know if the symmetric multiprocessing Grand Central Dispatch/libdispatch affords multicore CPUs is in any way the same (I’m under the impression that SMP has more to do with how RAM is orderly allocated to cores), but I sure would like to learn more about it.

GCD used to also organize multiple identical CPUs in one machine. PowerMacs used to be dual processor computers.

It would be cool if Apple made a Mac Pro with two M[#] Ultras — but Apple seems more interested in fusing two M[#] chips together and calling it a new processor — and I’m not saying it’s for the wrong reasons, either, btw.

 

[bobcomer]

It’s a basic fact of computing. Since you mentioned that you develop software I’m a bit puzzled that this is something you require proof for.

Maybe because, like I said before, there's *never* just one thing running on a modern computer, and I don't think that way either.

Sure, I can run a single task faster on a twice as fast core over 2 half speed cores, but that's never how I work or how I expect a computer to be behaving in my user's usage.

Their PC may be running many things at once and they expect switching to any of them without delay, or things that don't even concern themselves directly like backups and sync's to stay in the background and not delay what they're doing up front.

The difference between single job and general usage is why. I have 2 email programs, two browsers with multiple tabs open (edge and Chrome, about 10 tabs each), a powershell command line, 6 terminal sessions to our midrange machine, 2 excel speadsheets, and it's a friday, as light as it gets. A single core machine wont be able to do it as well. 346 processes, 5500 threads. 8 cores. (Ryzen7)

 

[R2DHue]

Okay, assuming the same kind of cores, just different clocks, I'd take the 2 2.5Ghz cores.

For tasks being handled that don’t need 5GHz, and if two 2.5GHz cores can process two different instructions simultaneously, 2x2.5GHz seems like the obvious answer.

But if tasks can be handled faster serially on a 5GHz processor, that changes things.

 

[APCX]

For tasks being handled that don’t need 5GHz, and if two 2.5GHz cores can process two different instructions simultaneously, 2x2.5GHz seems like the obvious answer.

the 5ghz processor will process those two instructions in the same time (theoreticallyj as the dual core machine. In reality, most of the time the single core will be faster due the fact that the work doesn’t need to divided etc.

But if tasks can be handled faster serially on a 5GHz processor, that changes things.

 

[bobcomer]

I admit I’ve never understood why single-core benchmark scores are given so much consideration.

When does a 4 or 6 core CPU use only one core? (I seriously don’t know the answer.)

I can see how it might be true in the case of computational tasks that don’t need very much horsepower, or in a case where the OS or controller is trying to reduce power consumption and heat, but, in Apple Silicon, tasks like this would probably instead be handed to the Efficiency CoreS — which, as the name implies, are multiple cores, too.

If there are times when software is executed on a single core, I’m interested in learning more about it and why.

And the most I know about CPU vs. GPU is that multi-core CPUs and multi-core GPUs are very different in terms of symmetric vs. parallel processing, respectively.

I don’t know if the symmetric multiprocessing Grand Central Dispatch/libdispatch affords multicore CPUs is in any way the same (I’m under the impression that SMP has more to do with how RAM is orderly allocated to cores), but I sure would like to learn more about it.

GCD use to also organize multiple CPUs in one machine. PowerMacs used to be dual processor computers.

It would be cool if Apple made a Mac Pro with two M[#] Ultras — but Apple seems more interested in fusing two M[#] chips together and calling it a new processor — and I’m not saying it’s for the wrong reasons, either, btw.

I like the curiosity!

I'm more of just a software guy and I've been out of college for a long time.

leman is who I'd listen to, I respect his opinion a great deal and know he knows more than I do, especially about hardware. (I just disagree with something like the single core ideal in this thread. <g>)

 

[R2DHue]

the 5ghz processor will process those two instructions in the same time (theoreticallyj as the dual core machine. In reality, most of the time the single core will be faster due the fact that the work doesn’t need to divided etc.

The cost of “thread context switching” in dividing tasks?

 

[R2DHue]

I like the curiosity!

I'm more of just a software guy and I've been out of college for a long time.

leman is who I'd listen to, I respect his opinion a great deal and know he knows more than I do, especially about hardware. (I just disagree with something like the single core ideal in this thread. <g>)

Thanks for the help.

I suppose, to be charitable, it could be argued that single-core scores are important to show off in isolation the quality and the power of the generational core itself — that is replicated 4 or 6 or more times on a multi-core die.

(But it is less meaningful when you see that multi-core scores far exceed the number of cores multiplied by the individual single-core score.)

 

[APCX]

The cost of “thread context switching” in dividing tasks?

I’m referring mainly to the fact that much work can’t be easily divided, or may depend on work that is taking place or hasn’t yet taken place.

Many tasks are stubborn like that. Some aren’t but with single core perf, you don’t need to worry.

 

[Analog Kid]

Maybe because, like I said before, there's *never* just one thing running on a modern computer, and I don't think that way either.

Sure, I can run a single task faster on a twice as fast core over 2 half speed cores, but that's never how I work or how I expect a computer to be behaving in my user's usage.

Their PC may be running many things at once and they expect switching to any of them without delay, or things that don't even concern themselves directly like backups and sync's to stay in the background and not delay what they're doing up front.

The difference between single job and general usage is why. I have 2 email programs, two browsers with multiple tabs open (edge and Chrome, about 10 tabs each), a powershell command line, 6 terminal sessions to our midrange machine, 2 excel speadsheets, and it's a friday, as light as it gets. A single core machine wont be able to do it as well. 346 processes, 5500 threads. 8 cores. (Ryzen7)

For that kind of workload, it's probably close to a wash, frankly. There's overhead in context switching, but there's also overhead in load balancing between two cores. If you had two fully independent computers processing fully independent data they could probably perform faster than one machine that was twice as fast. But a desktop scenarios where you have multiple windows open doing things in parallel all under one OS instance isn't independent. So if you have two cores sharing work, some level of coordination needs to happen between the cores and that introduces overheads just like context switching on a single core does. And typically the overhead of context switching will get hidden by the idle time between operations in a desktop system anyway.

One exception to the 1 fast is better than 2 slow is real time systems. If you have two real time tasks then the latencies and interrupt contention of running two processes on one CPU can give the two-processor model an advantage. I remember you saying you do a lot of work with industrial controls, so went back to that thread to see if that's what was leading you to think this way-- but it sounds like you're running everything under Windows, so if you're having problems handling real time processing the OS is a bigger problem than the hardware.

 

[bobcomer]

One exception to the 1 fast is better than 2 slow is real time systems. If you have two real time tasks then the latencies and interrupt contention of running two processes on one CPU can give the two-processor model an advantage. I remember you saying you do a lot of work with industrial controls, so went back to that thread to see if that's what was leading you to think this way-- but it sounds like you're running everything under Windows, so if you're having problems handling real time processing the OS is a bigger problem than the hardware.

Real-time is something else, I agree. Those testing systems I talked about don't run modern windows and the only more modern real time I've dealt with are specialized linux systems. I don't personally do anything more than setup with that kind of thing.

As for if it's a wash, not in my experience.

 

[JouniS]

Two 2.5Ghz cores are not going to get more work done than a single 5Ghz core. In an ideal scenario, you’d get the same amount of work done. At worst, you’d get half that.

You are applying the pattern out of context instead of thinking about cause and effect. As @APCX explained, we only have multi-core CPUs because our ability to build fast single-core CPUs has hit a wall.

One fast core is better than multiple slow ones, as long as you are running a single task.

The situation changes with multi-tasking, where private resources are better than shared resources. And the reason is not performance but robustness. Outside some limited circumstances, humans are too lazy and irresponsible to write software that uses shared resources properly. When one piece of software starts misbehaving or otherwise consumes all the resources it can get, private resouces help to control the damage. When the other tasks have their own CPU cores (and other resources), they can probably continue running without too much disruption.

 

[R2DHue]

I like the curiosity!

I'm more of just a software guy and I've been out of college for a long time.

leman is who I'd listen to, I respect his opinion a great deal and know he knows more than I do, especially about hardware. (I just disagree with something like the single core ideal in this thread. <g>)

A wise person once said,

“If your expertise is in software engineering, never gain too in-depth a knowledge of hardware engineering. If you do, you’ll learn all about the awesome capabilities of the most modern hardware, but you’ll also be infected with knowledge of the inescapable limitations of the hardware of any given era, and it will inexorably restrict your outlook on what’s possible to achieve in software. The reverse applies if your expertise is in hardware engineering.”

I remember reading that Steve Jobs was always careful not to gain too in-depth a knowledge of either. He once said something like, “Engineers are always happy to tell you what’s not possible.”

It’s oft said that, “Knowledge is power.” But for his purposes, too much knowledge would prove disempowering.

Ever the visionary, Steve Jobs would task a hardware or software engineer with something, and after they told him it was impossible, he’d say, “Do it anyway.”

They’d always grumble, but often — wouldn’t you know it — they’d achieve “the impossible” after all!

See, they lacked his imagination to see creative solutions to problems they very technically deemed “impossible.” (Not true of all engineers; many are artists, but…)

Steve Jobs always remained floating above in the upper regions of the “visionary ether,” never down “in the weeds.”

He was truly one in eight billion…

That could explain why “brainy” Bill Gates sucked so bad — I mean, objectively speaking, that is.

I’m not in either field career-wise, so I learn stuff for fun, entertainment — for nourishing my insatiable curiosity.

Now…

When is iFixit or someone going to reveal whether or not there are more AMX coprocessors on the A17 than on the 16?

 

[leman]

One fast core is better than multiple slow ones, as long as you are running a single task.

The situation changes with multi-tasking, where private resources are better than shared resources. And the reason is not performance but robustness. Outside some limited circumstances, humans are too lazy and irresponsible to write software that uses shared resources properly. When one piece of software starts misbehaving or otherwise consumes all the resources it can get, private resouces help to control the damage. When the other tasks have their own CPU cores (and other resources), they can probably continue running without too much disruption.

I don’t understand what you are saying. What do you mean by private resources ? Running different tasks on their own thread? A single core can run multiple threads with no problems at all. Multi-threading is much older than multi-CPU systems.

 

[leman]

The cost of “thread context switching” in dividing tasks?

Context switch cost will only become a problem if the number of threads is very high. And then it doesn’t matter much if you are running one or two cores. Unless of course you are running thousands of demanding threads on thousands of cores (hello GPUs!) but I hope we are not taking the argument that far

One exception to the 1 fast is better than 2 slow is real time systems.

If you have hard real time demands you need a special real time CPU and OS anyway. That’s a different discussion.

 

[Analog Kid]

If you have hard real time demands you need a special real time CPU and OS anyway. That’s a different discussion.

The definition of real time varies with the application, but you often don't need a special real time CPU (Arm gets used for a lot of real time applications and can be helped by locking the cache and other tricks). Depending on your requirements you may not even need a specialized RTOS, but Windows is notoriously bad at real time applications, as I mentioned:

you're running everything under Windows, so if you're having problems handling real time processing the OS is a bigger problem than the hardware.

 

[Analog Kid]

One fast core is better than multiple slow ones, as long as you are running a single task.

The situation changes with multi-tasking, where private resources are better than shared resources. And the reason is not performance but robustness. Outside some limited circumstances, humans are too lazy and irresponsible to write software that uses shared resources properly. When one piece of software starts misbehaving or otherwise consumes all the resources it can get, private resouces help to control the damage. When the other tasks have their own CPU cores (and other resources), they can probably continue running without too much disruption.

346 processes, 5500 threads. 8 cores.

In this kind of situation, the overheads are much the same regardless of how many cores you have. So, divide those numbers all by 8? You're still context switching and quite possibly moving threads and processes from core to core, trying to maintain cache coherency, etc.

If you're worried about bad citizens, limit their resource usage at the OS level.

 

[Basic75]

I admit I’ve never understood why single-core benchmark scores are given so much consideration.

One woman can make a baby in 9 months. That does NOT mean that 9 women can do the same thing in 1 month. Because some things just can't be done in parallel, or it's a lot of hard work to make use of more processors. See games for example. It took them years to make use of 2 then 4 then 8 cores. You'll be hard pressed to find a game that runs faster on an M2 Ultra than an M2 Max because of the higher number of CPU cores.

 

[JouniS]

I don’t understand what you are saying. What do you mean by private resources ? Running different tasks on their own thread? A single core can run multiple threads with no problems at all. Multi-threading is much older than multi-CPU systems.

Separate CPU cores. Separate SSDs. Separate GPUs. And so on. When a task starts misbehaving or otherwise consumes more resources than expected, it doesn't consume all the resources in the system. Just the ones that have been allocated to it.

A CPU core is a convenient unit of resource allocation. A scheduler can just let a task use 100% of a core for extended periods of time without any special permissions, because many tasks have a legitimate need for that. If a task misbehaves, it doesn't matter much, as other cores remain for other purposes. An equally fast single-core system would be more fragile, as tasks that need CPU time for legitimate reasons would end up competing against ones that misbehave.