• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Have CPUs reached the limit?

Associate
Joined
25 Jul 2007
Posts
1,768
Location
Essex
With the somewhat lacklustre increase in performance from the AMD 9000 Series, is it possible that we're approaching the limit as to how much performance a CPU can achieve?
 
Even if there are limits for die sizes such as going smaller than the current 3nm processes, no doubt the boffins at the various manufacturers will work out how to increase performance. The techniques used on Apple Silicon might be the way forward?
 
The node limits are mostly economic.
Sure TSMC will reach their 2nm node but each wafer will cost around $30k so the cost per transistor will be stagnant. Meanwhile designing, validating, and making the masks for each 2nm design will be astronomical.

Chiplets and tiles will be required even if monolith would perform better (and monolith should perform better all else being equal - which sort of highlights how far behind Intel's huge P cores were since Zen chiplets outperformed then despite being chiplet).

Other tricks taken from phone SOCs and things like Apple Silicon will also be required.

Things like on-package memory, ASIC accelerators, NVMe drives directly connected to the CPUs memory system/cache, etc.

All things which will probably make upgrading and DIY impossible. Apple started soldered memory because they are greedy and anti consumer, but there now are technical reasons too.
 
Last edited:
There will never be a limit, just more cores and more distributed, I guess only factor will be cost and distribution of work over ever increasing bandwidth and faster networks
 
There will never be a limit, just more cores and more distributed, I guess only factor will be cost and distribution of work over ever increasing bandwidth and faster networks

This.

We might be reaching a limit on process node size, 10 years ago intel CPUs were at 22nm, now we're approaching 2-3nm processes. Are we likely to be at 0.2nm in 10 years time? Obviously not. But we can just make the die bigger to fit more transistors and more logical cores and boom that's a faster processor. This is already happening, the flagship desktop intel CPU has less raw power than a threadripper from a couple of years ago. How? The threadrippers are huge and have dozens or hundreds of cores. The only reason that the entire domestic computing market hasn't gone that way is because videogames and a few other tasks still scale with single-core performance, once developers and development frameworks catch up that should no longer be a factor.

Also there are huge gains to be made in efficiency. Performance-per-core or performance-per-die matters less than performance-per-watt in many applications and that trend will only accelerate.
 
Last edited:
This.

We might be reaching a limit on process node size, 10 years ago intel CPUs were at 22nm, now we're approaching 2-3nm processes. Are we likely to be at 0.2nm in 10 years time? Obviously not. But we can just make the die bigger to fit more transistors and more logical cores and boom that's a faster processor. This is already happening, the flagship desktop intel CPU has less raw power than a threadripper from a couple of years ago. How? The threadrippers are huge and have dozens or hundreds of cores. The only reason that the entire domestic computing market hasn't gone that way is because videogames and a few other tasks still scale with single-core performance, once developers and development frameworks catch up that should no longer be a factor.

Also there are huge gains to be made in efficiency. Performance-per-core or performance-per-die matters less than performance-per-watt in many applications and that trend will only accelerate.
Yep exactly, Also for anyone thinking dies cant keep increasing, well yes they can just vertically.
 
Some would argue that they reached a limit years ago!
It depends what you call a limit. Gone are the days when a new CPU would be twice the power of the previous generation! For a long time, generational increases have been relatively small.
We have reached a limit of size. You just can't make them much smaller because quantum physics says so. There is a thing called quantum tunnelling, where electrons can jump past junctions. The smaller the transistor the larger the problem. That's why chips generate so much heat these days. The electrons are literally jumping out of the transistors because they are so small, doing nothing other than generating heat. Even to the point where they can't make the chip itself occupy a larger area because the substrate deforms. This very problem has held back the 5000 series GPU's from NVIDIA. Glass is the next breakthrough, apparently.
The problems will be solved over time, but it's getting more and more difficult to squeeze more hardware power for the gamer. NVIDIA has been building new features in to the hardware, because it knows that's the only way it can really improve on the previous generation. Well, that's my opinion anyway.
 
Last edited:
When will the CPU equivalent of NAND QLC come, and what major caveats will it bring!
I imagine cooling is going to be the biggest issue, I read somewhere that we may end up with some clever liquid cooling much like the veins in an internal combustion engine. bit above my knowledge though. Interesting times though given AMDs progress with 3d cache etc
 
Last edited:
Back
Top Bottom