why intel and amd just combine multiple cpus into single big cpu?

[aqarwaen29]

lets take for example i13600k. 

why not just take 4 same cpus and make them into single big cpu for better performance?

i know that you need new socket but that would smallest issue i think.i think real issue would be cooling,but this could easily sloved with cooling solution.so why not do it?

[CommanderAlex]

Cause Intel got jealous after Apple dropped them in their Mac lineup...

 

/s 

 

They want to try and emulate power consumption of ARM on x86 with the big.LITTLE design by putting the big performance cores to sleep that take more juice than the efficiency cores that "sip" power for tasks that don't need a lot of CPU performance. 

 

I know there's probably more to it but just a jist of what's going on. 

[toasty99]

They do?

 

That's literally what AMD Ryzen was. AMD was took two CPU chiplets and used the infinity fabric technology (or as Intel called it at the time "glue") to put two separate CPU chiplets together.

 

AMD was able to scale this into AMD Threadripper and AMD Epyc.  On the consumer side, you're looking at two chiplets maximum. On Epyc, they have used as much as 12 chiplets.

 

Intel is trying their own version of this. Look up Level1 techs if you want a deep dive into it but the tldr is that Intel is going to try using a tile structure to seperate different parts of the CPU.

[Blue4130]

6 minutes ago, aqarwaen29 said:

lets take for example i13600k. 

why not just take 4 same cpus and make them into single big cpu for better performance?

i know that you need new socket but that would smallest issue i think.i think real issue would be cooling,but this could easily sloved with cooling solution.so why not do it?

 

AMD does that. That's what chiplets are. There are downsides though. It's cheaper, but the communication between chiplets hurts performance.

 

[Godlygamer23]

It's already sort of done, but you have to keep things within certain power envelopes. Intel's server CPUs will take an architecture that already exists on the consumer side, and scale it up to have additional cores(I'm seeing up to 36 Performance Cores), but cut back on frequency, unless the situation calls for a single core being stressed, in which case Intel might throw some of that out the window and allow single cores to push the frequency much higher.

 

For example, the Intel Xeon w9-3475X CPU offers 36 Performance Cores, and when the CPU is totally stressed, the frequency will drop to 2.2GHz, but when a single core is stressed, that core can turbo up to 4.8GHz. It can get more complicated than that based on the workload, but both Intel and AMD already basically scale up architectures, which is basically taking lower end CPUs and putting them together, although it's not quite like that, since server and enthusiast CPUs offer additional features that the consumer counterparts don't offer. 

[RONOTHAN##]

40 minutes ago, Blue4130 said:

 

AMD does that. That's what chiplets are. There are downsides though. It's cheaper, but the communication between chiplets hurts performance.

 

Not to mention that it's only cheaper on the high end. There's a fixed cost to doing chiplets, so while it might be cheaper to produce a chipset 16 core than a monolithic 16 core, it's about even for 8 core chips and it starts getting cheaper for monolithic chips for the 6 core parts. There's a reason why they haven't released a chiplet based Ryzen 3 chip since the 3100/3300X, those chips just aren't cost effective unless your yields are really bad. It's also the main reason we haven't seen a Ryzen 3 7000 series chip yet, those chips are going to be monolithic and come out when they release the APUs. 

[8tg]

Been there, done that, 2005.

”hey dawg, I heard you liked the pentium 4, so we put a pentium 4, in yo pentium 4”

the reason why you don’t see this often is because unless the base dies used are fairly efficient, they quickly turn into power hungry, latency ridden house fires. 
AMD did amazingly with this concept with threadripper in recent years, but that’s not always suitable to fabricate. 
You don’t really see it done in the efforts of simply raw performance because it makes a poor performing product on paper most of the time. Do you really want a cpu with 4 times the thermal output of a 13700k? Even if it gives you 4 times the performance? As cool as it sounds, god no. It’s a cooling and power delivery nightmare.

So it’s usually done with a bunch of lower power chips that allows for the manufacturing of high core count, still reasonably powered large chips. 
 

There are also limits with sockets and physical cpu package a die size, but that’s a whole other ordeal, as it’s limited by 

1) the speed of electrons transferring through the material

2) communication latency between parts of the cpu as a direct result of limit 1

 

if you make a dinner plate sized cpu with a bajillion cores it’s going to be really good at basically nothing, because it won’t be able to communicate with anything in a timely manner, so even if it has the potential to do amazing amounts of math, it can’t actually take in or put out information quickly enough 

 

[Blue4130]

1 hour ago, 8tg said:

 

if you make a dinner plate sized cpu with a bajillion cores it’s going to be really good at basically nothing, because it won’t be able to communicate with anything in a timely manner, so even if it has the potential to do amazing amounts of math, it can’t actually take in or put out information quickly enough 

 

Epyc does it pretty well They work great within limitations. Things like VM's where you use chunks of the cpu don't care about other chunks. Great for that. But not so good for things like gaming. Big chips with big core counts on multiple chiplets are great, but only for very specific tasks.

[starsmine]

wait, is OP talking about a super fast single core?

[Godlygamer23]

2 hours ago, starsmine said:

wait, is OP talking about a super fast single core?

Honestly, this can go in a lot of directions as far as what they really mean. However, the reality is, is that CPUs all use a base architecture that is then scalable up to higher levels, or down to lower levels. Naturally then, it also allows you to sell tiered items at different price points based on customer needs. 

 

CPU(and GPU, for that matter)design is very complicated and a lot of factors are considered when designing the architecture that the different products are based off of. 

Edited November 20, 2023 by Godlygamer23