Why no higher clockspeeds?

[HangTime]

Seven years ago, if somebody had told you that in seven years time, the fastest processor on the market would be clocked at 3.4ghz, they would have been laughed out of town. Until that point, clock speeds had steadily increased, I'd say they probably doubled every 2 years, something like that. Even in spite of improvements in architecture.

Now don't get me wrong, I've been around the block enough times to know that clockspeed isn't everything. Nowadays cpus have multiple cores, and even a single Sandybridge core at 3.4ghz is going to be way faster than a P4 even at 5 or 6ghz. It's good that cpus are getting more efficient and becoming faster clock-for-clock.

However, what I've also wondered about is why to my knowledge we still haven't had a retail cpu at 4ghz or more? To my knowledge the fastest (in clockspeed terms) we've ever seen was 3.8ghz from Intel back in 2004. AMD have a 3.7ghz part on the market today. But still no 4ghz.

Why do I care so long as performance keeps increasing? Well, I just want to have my cake and eat it. I want more better performance per clock AND higher clockspeeds. Pentium was faster than 486 yet had better clocks too. P2 was faster than Pentium yet had better clocks too. Surely Intel could get a slightly better retail cooler, bump up the voltage a tad and come out with a 2700k or whatever at 4ghz bearing in mind the overclocking success people have had.

It just seems a bit alien to me running cpus with the same headline clockspeed for so long. I even went through three upgrades (P4-1.8A, A64 3000+ Venice, C2D E4300) all at the same clockspeed of 1.8ghz, that's when I started to notice things were starting to plateau on that front. A far cry from previous upgrades which had always seen my clockspeed more than double (650mhz, 300mhz, 133mhz etc).

So is it some sort of technical limitation with heat dissipation or what? Just a question of manufacturers prefering to pile on more cores and getting better efficiency from that? I mean looking back to what I would have projected back in the day, by 2011 I'd have been expecting to be running a 5-digit clockspeed.

 

[PhillyDee]

I believe it is a limitation with the transistors themselves. There is only do fast you can have them switching. Thats why the new architectures do more wit the exisitng speed than up the speed themselves.

 

[james.miller]

it's not sustainable. heat and power becomes more of a problem the higher you clock a transistor. It also becomes more of a problem the smaller you make that transistor. combine the two and there's your answer. the switch was made from pushing clockspeeds to improving efficiency and parallel processing, in a nut shell, as this allowed the performance increases to keep coming whilst keeping withing reasonable power envelopes.

clockspeed eventually became unimportant in the race for processing power.

 

[iviv]

From how I see it, they were getting faster by shrinking the transistor size so they could fit more into the same space. More transistors mean more computing power. But unfortunately there's a limit to how small we can make things, and we're pretty much butting heads with that wall.

 

[ian54]

Distributed processing and n-tier technology means that modern software development is more dependent on multiple parallel processors than raw clock speed.

As mass market virtualization comes of age, the ability to dynamically load balance determines the overall effectiveness. An analogy would be that if you want more traffic on the motorway it is better to add lanes than increase the speed limit.

"Industrial" CPUs often run at below 2.5Ghz, but they have 8+ cores each running multiple threads, and there is usually more than one CPU in a box.

The future technology and business model of online gaming (and possibly everything else) will be that all processing is done server side, your PC will be nothing more than a dumbed down graphics card and the CPU wont matter any more.

 

[HangTime]

An analogy would be that if you want more traffic on the motorway it is better to add lanes than increase the speed limit.

The problem with that analogy is that it is based on the assumption that all you care about is the amount of traffic on the motorway. If instead you care about, say, a subset of the traffic getting from A to B in the quickest time, adding more lanes won't necessarily help as much as increasing the speed limit.

 

[jonny512379]

The problem with that analogy is that it is based on the assumption that all you care about is the amount of traffic on the motorway. If instead you care about, say, a subset of the traffic getting from A to B in the quickest time, adding more lanes won't necessarily help as much as increasing the speed limit.

Isn't this the same thing though, (in a way as the analogy may not relate exactly) as the traffic has always got from A to B in the same time, no matter what CPU or even a single transistor?

Isn't the speed of the CPU how fast it can switch its transistors (or how many time it can do this a second) so with out increasing the speed they switch at and with more transistors = more calculations per second and the time from A to B is still the same..... but you can still fit more in giving the false impression of being faster...

 

[bru]

clockspeed eventually became unimportant in the race for processing power.

I know the point your making and I do agree with it, but I must just say that line made me chuckle, especially on these forums where things are never fast enough

 

[james.miller]

I know the point your making and I do agree with it, but I must just say that line made me chuckle, especially on these forums where things are never fast enough

not really, im talking purely about the design of the architecture. if your looking at different models within the same family of cpus then yeah of course it's important

 

[ATIorNvidia]

With current technology, CPU's would be incredibly inefficient and draw absurd amounts of current if we had much higher stock speeds. You get far more performance per watt with lower clock speeds.

 

[HangTime]

Isn't this the same thing though, (in a way as the analogy may not relate exactly) as the traffic has always got from A to B in the same time, no matter what CPU or even a single transistor?

Isn't the speed of the CPU how fast it can switch its transistors (or how many time it can do this a second) so with out increasing the speed they switch at and with more transistors = more calculations per second and the time from A to B is still the same..... but you can still fit more in giving the false impression of being faster...

What I'm getting at in terms of relating the analogy to computers is this.

Lets say you have a 2 lane road with a speed limit of 60mph [Dual core cpu at 6ghz]
Compare this with a 4 lane road with a speed limit of 30mph [Quad core at 3ghz]

In cases where you are running a lot of fairly equal or well threaded applications, one would expect roughly comprable performance (or in terms of the analogy, you have a large huge number of cars that all need to go at 30mph). This is fine because in a given time period you are getting the same amount of work done, and shifting the same amount of cars down the road.

Now lets look at a situation where you are running very few applications of which one has a single, very demanding thread. This is equivalent to say a single car that wants to drive at 60mph. In this situation, the dualcore system will perform better, or in the analogy, the two-lane road will perform better. 4 lanes is only any use if you have enough traffic to make proper use of it.

 

[Combat squirrel]

Weird Answers ?!? It's because of physics, no need for weird analogies. It's because a transistor cannot turn off and on much quicker than 4-5ghz depending on it's size. It super simple laymens terms turning something on and off quickly produces heat, sort of. It's also to do with how fast the electrons are whizzing around the processor paths, you can make less heat by using less power, to do that u need to use smaller lithography, were currently on 35-45nm, older processors were 90nm, next years will be 22nm, the smallest were likely to get is 9nm, any smaller quantum effects come into play as well as severe electron migration, so thats it, the teeny tiny wires can only carry so much electrical load, to much it gets to hot , so 'slowing' the processor or keeping it within certain physically dictated speed limits is the only way, I think with processors 5 years from now they will all top out at 5-6ghz absolute max.

 

[HangTime]

Even 130nm processors reached 3.4ghz so I'd have expected more from 35-45nm, although I suppose they have more cores producing heat. Normally we got faster cpus with die shrinks but since 2004 that seems to have dried up.

Maybe turbo mode needs to be a bit more extreme to achieve my vision, having one or two cores ramping up a lot more (I understand some mobos let you tweak this?)

 

[james.miller]

Even 130nm processors reached 3.4ghz so I'd have expected more from 35-45nm, although I suppose they have more cores producing heat.

multicore was the answer to the problem, not the cause of it.

 

[Redmint]

Even 130nm processors reached 3.4ghz so I'd have expected more from 35-45nm, although I suppose they have more cores producing heat. Normally we got faster cpus with die shrinks but since 2004 that seems to have dried up.

Maybe turbo mode needs to be a bit more extreme to achieve my vision, having one or two cores ramping up a lot more (I understand some mobos let you tweak this?)

Shrinking transistors (past a certain point) unfortunately doesn't actually make them easier to clock. I think the record is still 8.2 GHz on a Celeron D (65nm).

The clock speeds are more due to the limits of silicon itself rather than the transistors, I doubt we'll see much higher than 5GHz stable overclocks in the next few years.

 

[keyser]

There's a few reasons.

Moore's Law transistor counts doubles every 18-24 months, it doesn't mention about related performance increases. Actual performance is only a nice bi-product of increased transistors. So doubling the transistors doesn't equal doubling performance.

Back to clock speed..

Jim Gray explained it best in a channel9 interview a while back [requires sliverlight]. The speed of light through a vacuum is 300 million meters per second, or ~7cm in 4 billionths of a second, and less for electrons through silicon. So information can only travel from one side of a chip to another in that distance. Granted current chips are about 1cm long, but the circuit length maybe longer..

Mentioned in this blog too.

 

[Phenomenologica]

Actually the electrons don't flow through wires (or silicon) in the way you imply. Yes they move incredibly fast, but as they do so in three dimensions and collide with atoms in the wire electrons move along a wire incredibly slowly. It varies depending on the current and the size of wire, but it's on the order of centimetres per hour. Their speed is great, but their net speed in any particular direction is tiny. It's a set of electromagnetic waves propagating down the wire that actually carries the energy from one point to another, not the electrons themselves. But this Gray person is correct in that there is a limitation on how far a bit can be transmitted down a wire in a single clock cycle.

 

[topdog@OC]

It's funny you should choose 7 years since that's about as long as we've been warned this is going to be happening anyway (re: Pat Gelsinger's keynote at intel's developer conference in feb 2004). It's just taking and taken the software industry a time to wake up and face the facts, outside of server and specialist apps, which would support multi-core regardless of the clock speed change rate anyway.

Here's one of the more well known articles still intact from 6 years ago, though not the first:
http://www.gotw.ca/publications/concurrency-ddj.htm

"Arguably, the free lunch has already been over for a year or two, only we’re just now noticing." (March 2005)

Roll on 2010+ and still it seems "only we're just now noticing". Development cycles take a long time to change, but there's been time enough for a couple of full cycles in this time at least.

As for the limit to the processor itself, heat is a big factor, the common quote is that the heat within the processor would soon reach/exceed the surface temperature of the sun, if they didn't switch to alternative means like specialised instruction sets or hyperthreading and multicore architectures.

 

[ian54]

Now lets look at a situation where you are running very few applications of which one has a single, very demanding thread.

But you wouldn't design software that way any more, that's my point, so clockspeed isn't as important as parallelism.

The analogy with the traffic should be amended, if the traffic was carrying freight, it is better to have lots of little vans to take advantage of the extra lanes, rather than one big lorry.

 

[Corasik]

There the multiple cores, but dont forget that the IPC has gone up too. So single threaded apps are still way faster on a modern processor than they were on a a P4.

Sure you you take a very very simple instuction, such as an add, the highest clocked P4's actually did quite well. But for an average singled threaded application with a wide range of instructions and branches, with a low IPC processor its like driving down a congested road. Sure your top speed is 70mph, but your stuck in traffic doing 10mph.

With a processor with a higher IPC, its more like driving down an empty road, with no corners, you can put your foot down, hit 70, and fly along.

Modern intel and amd processors arnt just about multi tasking. They are about making sure that single threaded apps dont end up stalled, in long congested pipelines, combined with incredibly good branch prediction to keep everything flowing nicely.