Why not one really big core?

[hat]

I've seen many die shots detailing sections of a processor, and I'm sure many of us here have seen the same shots. All these chips with 4 or more cores... what if they just designed all that to be one big single core?

 

[newtekie1]

Eventually, now matter how big you make a single core, you reach the point where making it bigger and more complex doesn't really improve performance. So we need muti-core to increase processing power.

The reason all the cores aren't just clustered together into one big core is for easy of reconfiguration. It make it easy to disable parts of the CPU to make lower CPUs, or even just cutting out part of the CPU to make a low CPU. They design the big CPU first, with the knowledge they are going to carve out lower models from the die. So not having all the cores merged into one makes doing this easier.

 

[RejZoR]

That's not how chips or more specifically processors work.

Imagine that one core is an 8 lane highway. You can tweak it a bit to make traffic on it go faster back and forth (frequency and compute stages aka pipeline length) and maybe you can add 2 more lanes, but then it just becomes too impractical because it's too big. So, adding more lanes to it is not possible anymore and you can't make cars on it go at the speed of sound either. So, what do you do? You route another highway between places. And you make a second highway from some other direction. And a third one. And forth. Etc. That's basically multicore processors. A stack of fast highways because making one massive highway is impractical, but routing many of smaller ones is.

Same reason why workstation and compute cluster processors run at relatively low frequencies, but they have shit tons of cores. Some of it is reliability because high clocks are more difficult to maintain stable and because stupid high clocks only give you as much gain. Where every added core, even at slower clocks gives huge gains. So, thy just stack more and more cores together to get huge amounts of traffic through without all the inconveniences and impracticalities. It's similar with GPU's. Technically, with GPU's, every shader is a core. And modern ones have them in 4 digit figures. It just works better than having 8 shaders that have to run at the speed of light to be as efficient as those 4096 shaders running at just 1.6 GHz.

What is limiting us is the frequency and how electrons behave at such frequencies. The higher you go, the more problems you start encountering. It's why we aren't seeing any 20 GHz processors, instead we are somewhat capped at 5GHz. Anything beyond that requires extreme cooling to ease off the electrical issues we start facing in those scenarios. Just like you can't have cars going at the speed of sound on only 8 lane highways, just the same you can't make processors go at 20 GHz. Natural course to overcome that is to add more of slower parts and stack them up.

That's about as far as I could dumb it down so it should be easy to understand even for non techy people. Hope it helps.

 

[qubit]

One big reason is the end of Moore's Law, which prevents clock speed from going ever higher which we used to see, until around 2003.

We should have been running at something like 15-20GHz if it had continued to scale.

 

[BiggieShady]

CPUs are traditionally made for sequential algorithms that can be translated into ordered stream of instructions.
Generally speaking execution of any sequential algorithm always depend on a previous step because of the nature of a sequence, but there are also always many parts inside those steps that are mutually independent. More so, when compiled, depending how complex the instruction set is, there is much instruction level parallelism to be extracted from the machine code.
Super scalar processors (every cpu from this millennium) execute instructions in parallel on a single core in a single thread. Add to that thread level parallelism via SMT/HT.
How much instruction/thread level parallelism can be extracted from a code on a architecture, depends on code and architecture. It's always limited and not constant.
How many full cores can be added for true parallelism is much less limiting and scaling is constant.
Having one huge core would mean being able to exploit only instruction/thread level parallelism and not being able to run multiple independent hardware threads concurrently. Single thread performance may even be ok but it would be one power hungry cpu with so much unused powered die areas, and it would never see 100% usage.

 

[silentbogo]

what if they just designed all that to be one big single core?

One big drawback - context switching.
Each CPU core has registers, ALU(s), FPU(s) which are available to a single program(thread) running at a time. Every time you switch to another thread, you need to dump the contents of all registers and the current CPU state (status register) onto the stack (either memory, or cache if used as memory), load all of the above for thread #2 and keep on going with execution. If you did any assembly programming or at least familiar with basic CPU inner workings, you should know that anything involving reading or writing to memory is slo-o-o-ow comparing to any other instruction execution. This is a significant delay, considering that modern PCs run hundreds or even thousands of threads. Hyperthreading just gives you the ability to kind-of run two threads simultaneously, given that one thread won't use all of the core resources at once, but it still does not eliminate the problem of context switching.
Having a quad-core CPU reduces the switching time by a factor of 4 and you don't want to sacrifice that in favor of some mythical mega-single-core performance.

 

[Vya Domus]

There are two way to make a CPU core faster , make it execute instructions faster or make it execute more instructions concurrently. The limitations involving the first aspect of it are pretty obvious , clock speed is limited and at this point pretty much all instruction are as efficient as they can possibly get.

Making a CPU execute things concurrently is more complicated because there is so much instruction level parallelism you can achieve at any given moment.

Modern CPUs have something called an "instruction window" , which basically says how many instructions can it look at ahead of time to figure out which can be executed in parallel. Problem is the larger the instruction window the less effective it is because programs branch out very often. You do have something called a "branch predictor" to deal with that , but again , this has it's limitations as well. Essentially you can't realistically look ahead of time throughout all the programs you have to run and figure out the right decisions so that you can execute everything at once and get it right enough of the time to make it work.

So to wrap it up , making one big core would involve a lot of diminishing returns therefore it's not really worth it.

We should have been running at something like 15-20GHz if it had continued to scale.

That's a common misconception. Clock rate did scale pretty much how everyone expected it to , the problem is core width and count did as well , that's why you don't see 20 Ghz processors. I am pretty sure you can have a 20 Ghz 80186 if you wanted to.

 

[BiggieShady]

I am pretty sure you can have a 20 Ghz 80186 if you wanted to.

Nope, there's this thing called transistor switching speed and even if silicon fin fets were that fast, pipeline would have to be unmanageable level deep

 

[Vya Domus]

Nope, there's this thing called transistor switching speed and even if silicon fin fets were that fast, pipeline would have to be unmanageable level deep

I know but what I was trying to say was that clock speed would have increased by a lot more if there weren't any other additions throughout time that introduced more concerns about power consumption , on-die-latency , etc.

 

[phanbuey]

token inaccurate and unrelated car analogy: It would be like having a one cylinder engine, bro.

 

[Aquinus]

token inaccurate and unrelated car analogy: It would be like having a one cylinder engine, bro.

This. It's like saying a 1-cylinder engine that's 2.0L large is just as good as a 4-cylinder engine with the same displacement. In theory, they both can breathe the same amount of air which would means it can produce the same amount of power but, you don't consider things like, how balanced is the engine and how smooth is power delivery? The actually similar part of this is actually with engine balance. A big single cylinder engine would need huge balancing shafts to counteract the movement of the single piston and connecting rod. This limits the RPM, so while it can breathe the same, you're limited by the the speed at which the engine can run. CPUs will be the same kind of way in the sense that a larger monolithic circuit is less likely to operate at higher frequencies because of losses as circuit length increases as well as latency that needs to be considered because electricity can't travel faster than the speed of light (and actually typically goes slower in most circumstances.)

So yes, you could build a huge monolithic core. You would get something like what SPARC has with a crap ton of integer cores paired with a single FPU. The result in some operations being able to be effectively executed in parallel but, it does absolutely nothing for clock speeds. In fact, I would definitely say that it gets worse. Generally speaking, I believe smaller cores yield better results and have a much better opportunity to scale. Large dies don't and have the added issue of yields being lower since it's not like you can disable one core if it's bad or something for lower models. All in all, it's not a cost effective or a good way to improve performance... and it costs more money. All in all, it's just a bad idea.

 

[Vayra86]

For the same reason two hands can do more than one, or two sets of eyes see more than one.

We don't need car analogies

 

[Aquinus]

For the same reason two hands can do more than one, or two sets of eyes see more than one.

We don't need car analogies

Well, I think that's even more inaccurate because 1 core can do everything two cores can do. There are tasks that (typically,) require 2 hands to be done effectively. Same thing with sight, two eyes helps with things like depth. The car analogy works because, like engines, there are consequences to making things too big.

 

[Vayra86]

Well, I think that's even more inaccurate because 1 core can do everything two cores can do. There are tasks that (typically,) require 2 hands to be done effectively. Same thing with sight, two eyes helps with things like depth. The car analogy works because, like engines, there are consequences to making things too big.

Nope, 1 core can't do everything simultaneously while two cores can, it has limited resources/access capabilities that a single type of task can saturate. Also, there is the sequential nature of things, and with two cores, you can do two sequential things side by side.

 

[Aquinus]

Nope, 1 core can't do everything simultaneously while two cores can, it has limited resources/access capabilities that a single type of task can saturate. Also, there is the sequential nature of things, and with two cores, you can do two sequential things side by side.

...it can do the same tasks. Might take longer but, it's not incapable of doing it. When you lose an eye, you lose something, a capability of seeing some level of depth that you had before. More cores just means you can do more at once, you're not gaining the ability to do something you couldn't do before.

In a technical sense, a single core is still a Turing-complete computation engine. Adding cores doesn't change that.

 

[FordGT90Concept]

Some great responses above. To them, I'll add one point: Processors are a race against time. The more work it can do in a given timeframe, the more powerful it is. It is impossible to make one logic circuit do the work of 16+ in parallel in the same power envelope because of physics.

To extend that: a clock is literally a window in which a processor can perform an action. If you have a dual core running at 1 GHz, it would take a single core of approximately 1.8 GHz to equal it (90% efficiency accounting for thread management overhead). Quad core, 3.6 GHz. Octo core, 7.2 GHz. I'm already up to clock speeds that aren't possible and we already have octo-cores running close to 4 GHz which would require a single core operating at a staggering 28.8 GHz to match. There simply isn't enough time in a second to cycle a processor 28.8 billion times without parallelism.

 

[FreedomEclipse]

I've seen many die shots detailing sections of a processor, and I'm sure many of us here have seen the same shots. All these chips with 4 or more cores... what if they just designed all that to be one big single core?

Im more in the field of AMD.... They know how to make Dual Core, Quad Core and Hexa Core CPUs, why not apply that knowledge and make DC, QC or HC GPU Processors??

 

[FordGT90Concept]

Overhead. SLI/Crossfire worked by sending all the things to all the cards and the driver instructing each card what to do with the data individually. This approach is somewhat efficient at doing work but wasteful on cache resources.

If they're going to make SLI/Crossfire work in silicon, they need something like Vega's HBCC to take care of the cache management. HBCC was a necessary step towards Navi which is a dual (or more) core GPU.

 

[Vya Domus]

Apple is a prime example of designing wider and wider cores compared to other ARM or semi-custom cores. Each year they extract more and more performance without much increase in power consumption and with untouched battery life.

But it's not for free , they most likely pay more than any other competitor for their dies as they are bigger and more prone to defects. And it's not just that , at some point this tactic will bite them back the way Netburst did to Intel. They will reach a point when they wont be able to increase their cores size by much and they will end up with an incredibly complicated monster of a core whose power consumption will skyrocket even with a minimal increase in clock speed.

 

[qubit]

That's a common misconception. Clock rate did scale pretty much how everyone expected it to , the problem is core width and count did as well , that's why you don't see 20 Ghz processors. I am pretty sure you can have a 20 Ghz 80186 if you wanted to.

Not really. The problem is the amount of power consumed by the CPU as the clock speed goes up, which goes up massively. Yes, you can get a very simple chip to run at that speed, but I doubt that even a 486 will run that quickly without consuming masses of power and putting out gobs of heat.

The original Pentium 4 was intended to run at 10GHz and above on later models, but that didn't happen due to this problem. If you Google Pentium 4, I'm sure you'll find articles that explain this issue. The only way to go after that was sideways, ie multicore.

From what I can see, a slower multicore is better than a faster single core, because it's harder to bottleneck it, so this situation is probably not altogether a bad thing.

 

[Bill_Bright]

All these chips with 4 or more cores... what if they just designed all that to be one big single core?

Because BY FAR most computing tasks (number crunching tasks) are tiny. When you hand off 4 small tasks to 4 small cores, those tasks are completed in the same amount of time as a big single core takes to complete just 1 of those small tasks. This is because if you have only one core, you have to hand those tasks off one at a time sequentially - waiting for the first to be completed before even starting on the second.

And as qubit and others noted, Moore's Law applies. Processors can only work so fast (the Laws of Physics gets in the way). Those 4 small cores are working at the same clock speed as one big core.

So because there are 4 small cores vs 1 big core, and because most tasks are tiny, even if the 4-core processor has a slower clock speed, for most computing jobs, it will be faster than a faster clocked 1 core processor.

Remember too while you, the user, may only be doing one thing at a time with your computer, the operating system is doing many things at once. It is multi-tasking big time. It is managing memory, updating graphics information, scanning every I/O port, waiting for mouse and keyboard input, checking the Internet for updates, scanning for malware and so much more.

 

[cakehunter]

Contrary to popular beliefs, you can multitask on 1 core... be it old athlon/duron or p3/p4.
Running a game, playing a mp3, a text editor in background, browser, voice-comms (okay skype a bad example), managing plug and play USB- all of this was possible on 1 core without major slowdowns, if your processor was adequate for the task (adequate being relative, cause you dont run a datacenter off todays i7 or Ryzen).

Power saving was introduced with Athlon I believe and P4? Because of Spectre/Meltdown now we all know what branch prediction is, and it was present on Athlons and P4`s already.
But I think more cores arent neccessarily the solution. Imagine 1 million cores, each corse running a simple trivial task, say 1+1 or 1+2. Now this needs some managing seeing what core is free, what is running what, and what is the result of that calculation, it will need a almost a separate managing processor. IIRC there is some stuff in x86 that runs at 1mhz (or was it 100 mhz?) always and governs some internal stuff. Some popular reposting target:

 

[FordGT90Concept]

Imagine 1 million cores, each corse running a simple trivial task, say 1+1 or 1+2. Now this needs some managing seeing what core is free, what is running what, and what is the result of that calculation, it will need a almost a separate managing processor.

That's basically a GPU.

 

[Bill_Bright]

Contrary to popular beliefs, you can multitask on 1 core... be it old athlon/duron or p3/p4.

But that is not true multitasking. It just seems like multitasking because it is happening so fast. It is still juggling with one hand. You can only have one ball in the hand at once.

 

[Nosada]

Not really. The problem is the amount of power consumed by the CPU as the clock speed goes up, which goes up massively. Yes, you can get a very simple chip to run at that speed, but I doubt that even a 486 will run that quickly without consuming masses of power and putting out gobs of heat.

The original Pentium 4 was intended to run at 10GHz and above on later models, but that didn't happen due to this problem. If you Google Pentium 4, I'm sure you'll find articles that explain this issue. The only way to go after that was sideways, ie multicore.

From what I can see, a slower multicore is better than a faster single core, because it's harder to bottleneck it, so this situation is probably not altogether a bad thing.

It's not even solely a heat or power issue. The fact of the matter is the substrate simply can't handle those speeds and "leakage" pops up long before you ever get there. It would be like saying you can take a small one-cylinder engine and compensate for its lack of cylinders and displacement by making it run 100.000 rpm. Not only will that not yield the performance you're looking for, the engine will simply blow up because you are running into a whole scope of issues that simply don't exist below 10.000 rpm.