Why are CPUS not getting faster?

[beany_bot]

I was watching and reading about CPU history last night. One thing really stuck out to me. The jumps in clock speed from one release to the next were HUGE. But for the last ~10 years it has really slowed down. Now the focus is more cores.

Is this some sort of limit of silicone that we have reached? around 5Ghz? Or have Intel just been milking us?

 

[spoffle]

It's a bit of both Intel milking and limitations of clock speed. IPC (instructions per clock) has been increasing too, not by much but that's part of the Intel milking.

However, more cores really is the sensible way to go. As it means a lot more can get done at once. Increases in IPC haven't been all that necessary, as software has been quite inefficient when it comes to games.

Which is why you've seen things like Mantel, Vulkan, and DX12. They've been attempts at bringing the CPU overheads down as things just became a bloated mess.

Essentially more Ghz = brute force increases in speed. More cores and better software optimisation is where it's had to go due to the limitations of pushing CPUs faster and faster.

 

[ubersonic]

Bit of column A bit of column B, the Pentium 4 was supposed to hit 4GHz by the end of 2004 and then 8GHz before the end of 2005. However the rising heat from these increases meant that the 4GHz model had to be cancelled before launch as it required an AIO to prevent throttling, and higher clocked chips were abandoned. Intel then tried everything else to increase performance (more cache, higher FSB, Hyper Threading, then eventually more cores).

Of course that doesn't mean they haven't been milking us, for the past 7 years they basically had no competition, and as a result they were effectively walking the race. Improvements came slower and the "tick tock" cycle became the bane of any hardware enthusiast.

 

[EsaT]

Both.
Basically clock speed limit has been reached for current materials and manufacturing technology.
Any attempts at increasing clock speeds tend to need more volts causing fast rampup of power consumption/heat production.

Just for a comparison I overclocked my Q9550 Yorkfield from 2,83 GHz to fair 3,7 GHz while unvolting it from stock 1,25V to 1,2V
How many of todays high end CPUs can be overclocked any while undervolted?

Also increased possible transistor count doesn't help much anymore.
When every operation is already done in one clock tick adding new transistors doesn't help.
15-20 years ago lots of operations were divided to multiple stages which slowed down their execution lot.
Increased transistor count allowed newer CPU to have processing units capable to executing operation in single clock tick increasing speed without rising clock speed.
That road ended roughly decade ago.
Adding entirely new commands/operations doesn't help with existing codes and the more stuff you put into CPU core the harder its "bureaucracy" becomes to manage.


Though Intel has been and is definitely milking people in number of cores and threads.
Only i7 7700 can handle more than four simultaneous threads without huge performance penalty.
And then there are all those lower models with only two cores.
That's also one of the reasons why games are so badly multithreaded today.
Game developers can't rely on having guaranteed CPU resources for really more than one thread.

In comparison console CPUs have total raw processing power in class of one or two average desktop PC CPU cores.
But because there's certain number of dedicated cores available game developers can more easily divide load between them as much as it can be multithreaded.

 

[alec]

Cache size remains an ace in the sleeve. I bet when pushed Intel will release CPUs with quadruple cache size for awesome gains in IPC.
But other than that, 4-4.5 GHz looks to be the limit.

 

[beany_bot]

If, overnight say, we suddenly cracked room temperature super conductors. So basically the CPU would have no resistance in it ad make basically no to very little heat. How fast do you think we could go on silicone?

 

[riskman99$]

Quantum computing is where its all going for the next big speed step-up. Might still be a few years wait though...

 

[EsaT]

Cache size remains an ace in the sleeve. I bet when pushed Intel will release CPUs with quadruple cache size for awesome gains in IPC.

Actually cache size is two edged sword.
Increasing cache increases it physical size along with "bureucracy" increasing in turn latency in clock ticks.
And of course if execution units have to wait longer for data advantage is questionable.
So the most sensible approach would be adding it as L4 cache.
But adding more cache levels again needs more circuitry controlling bureaucracy and there's that thing of diminishing returns.

Of course if continuously used data happened to fit to new cache size then it would give clear advantage, but overal situation isn't that easy.

 

[melmac]

Another thing is that for most uses, CPU's are now more powerful than needed. Back when there was large jumps in performance for each generation of CPUs, even with just running office programs benefitted from a CPU upgrade. But since Sandybridge, and probably before it, there has been no real need for most people to change their CPU.

So that's part of the reason that things are shifting. Going from doing things faster to doing more things at once with as little power use as possible. I remember reading somewhere that die shrinks in the future are going to bring power savings but not much performance.

 

[Burr]

It seems that the process of making smaller transistors that can use less voltage is becoming harder and harder and with AMD (up until recently) being weak competition, Intel hasn't felt the pressure to step up their game. The next few years should be interesting and the next 10 - 15 years even more interesting since we're nearing the end of silicon.

 

[Cold_And_Miserable]

If, overnight say, we suddenly cracked room temperature super conductors. So basically the CPU would have no resistance in it ad make basically no to very little heat. How fast do you think we could go on silicone?

Silicon.
Clockspeed = electrical signal speed ~0.6c / length of die so maybe ~50Ghz for a Skylake sized chip with no resistance.

The simplest explanation for the stagnation is the cost per transistor has stopped decreasing beyond the 22nm? process.

 

[Panos]

I will agree. The shrinking of the transistors doesn't help to crack up higher speeds, while there is cooling issues and power consumption.

More packing of even smaller transistors creates issues on cooling while power raises the issues of leakage and interfering electrically corrupting the operations.

 

[Caracus2k]

the clock spend on silicon CPU's just can't reliably go much above 4.5-5Ghz without huge issues with voltage and cooling.

Mhz/Ghz race is largely dead for silicon CPU's. Improvement going forward will rely more on ipc improvements/ core count

 

[Orcvader]

A couple of years back Intel stated they plan to move away from silicon on the 7nm node. Whether that will allow higher clock speeds or not, we'll need to wait and see. Would also be interesting to see what material they pick. Graphene should be the cheapest right?

 

[smilertoo]

As soon as AMD stopped being competative intel started releasing the smallest updates they could get away with.

 

[Matkinson]

As soon as AMD stopped being competative intel started releasing the smallest updates they could get away with.

Yep, now that Ryzen has been released, Intel have miraculously announced an 18 core cpu. They have also moved their timeline forward by 6-12 months as well?

Pretty sure I also saw a leaked slide showing their new coffeelake, 8700k?, coming 2H 2017 will be 6 cores. This however may have been planned beforehand.


I also believe we are simply topping out the clock speeds and IPC for current technology, as the IPC of processors seems to have stagnated, regardless of how hard both AMD and Intel try.

 

[AWPC]

Intel milking it all due to no proper competition, few home user software requiring many cores or clock speed, software + games not optimised for anything more than the mainstream CPU due to how expensive it is to develop for multi core. If they wanted to include an AIO as std then CPU speeds could easily be 6-7Ghz years ago by backing of cache sizes & FSB.

 

[Orcvader]

I also believe we are simply topping out the clock speeds and IPC for current technology, as the IPC of processors seems to have stagnated, regardless of how hard both AMD and Intel try.

Clock speed yes, but IPC maybe not, Intel seemed more focus on the IGP as well as the power efficiency instead of actual CPU per core performance.

 

[cheesyboy]

TDP

CPU wattage went up and up for about 10 years, now it's going down and down.

 

[AWPC]

5.7Ghz here for one of the new mid range Core i9! http://www.pcgamer.com/intel-core-i9-7900x-hits-57ghz-and-crushes-cinebench-record/