Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Processor race over?

EvilDoc

EvilDoc

macrumors 6502a
Original poster
Jun 30, 2005
513
0
Circumventing the multi-verse...
Hey just a question, im just wondering how much longer we will be hovering around 3ghz? We broke the 3 ghz barrier back in 2003 and now 4yrs later we are still hovering around 3ghz. I do realize that processors have seen other updates ( dual cores, smaller chips, less heat to name a few ). And i do realize that there are and people have over-clocked chips to 4ghz, but when will 4ghz become main stream? We went from 200mhz to 3 ghz in 10 years now we have been at 3ghz for almost 5. What are you thoughts?
 
I understand the idea that processor speed is not the deciding factor but at some point we are going to break the 4 ghz barrier. Are processor speeds still going to be in the 3ghz for the next 5 years? I by no means need a 4ghz computer but i can help to wonder when the next big bump will be. Its not like im holding out on buying or anything like that, its just a question ive had rolling around in my head.
 
Apparently, all the chip-makers have hit the ceiling below 4 ghz, possibly because of the heat consumption, power consumption, something like that.

So, instead they use Dual/Quad Core systems at lower clock speeds. They are still pretty efficient.

We are going to see Quad Cores hit main stream in the next year or maybe even sooner.

Octo-Core is falling into the Mac Pro lineup slowly, and soon the whole Mac Pro lineup will be Octos.
 
you know what's ironic... i now start to associate lower clock speed with faster operation. What I mean by that is, I know a 3.8GHz Pentium 4 is way slower operation-wise than a 3.0GHz Athlon X2. And a 2.6GHz Core 2 duo is faster than that.

As far as 4GHz is concerned, my guess is that we'll probably get there when we get to 32nm or late 2008. The reason being, it takes a lot of energy to get to 4GHz. And really, we're not seeing too much performance increase because of those clockspeeds. Where we'll see more improvement is in the area of caching, front side bus increases, integrated memory controller after that, the ability for cores to talk to each other better, and more importantly... better software to handle all of these improvements.
 
Clock speed is important, but it is no longer the most important aspect.

As long as newer processors do more "work" than the old procesors, I dont care if the chip is runnning at 66MHz....

... i mean, you get my point.
 
The race is still on. It's just changed.
The transistors are like capacitors. E = energy, C = capacitance, V= voltage.

E= 1/2 C*V^2

As you toggle these transistors faster, the power goes up proportionally.
But to go faster, you need more voltage and is related with a square. So doubling the voltage causes power to go up by a factor of 4!
The die shrinks, from 130 to 90 to 65 to 45 ... causes the C factor to drop. So dropping from 130nm to 65nm the "C" (capacitance) should in theory drop in half. Not totally true, but close enough. But shrinking the die also means the total die area is smaller, so power density goes up and this is a problem. For example, say a 130nm cpu has 2 in^2 area and dissipates 100 watts, thats 50 watts per square in. But now take a 65nm cpu with 1 in^2 area and still at 100 watts. Now the power density is 100 watts per square in. Double!
Now back a few years, they could reasonably extrapolate speed and power for these CMOS cpu. Drop size by 2 (130nm to 65nm) and power dropped by a factor of 2. (equation above). But when they neared the 90nm mark the transistors started leaking more. This drop by factor of 2 didn't hold anymore.
Basically it gets hugely expensive to go faster now. So now they try to do more in parallel, hence dual cores & quad cores and more.

But its still about getting more work done is less time.
 
Processor performance is not stagnating. You are thinking in incorrect terms - "GHz = performance" is not true. Intel nearly reached 4GHz with the P4, yet that chip was not even close to the highest performance available at the time.

If you want to use the car analogy, GHz are like engine RPMs. Some motorcycles have a max RPM above 15,000. That does not make their engines more powerful than a Toyota Corolla at 5,500, or a freight train at 400. You would not want to "clock" a freight train's engine at 10,000 RPM because RPM is not the limiting factor on that engine.

Specially designed processors have been clocked to 100GHz in the lab. This does not mean they've performed 50x better than any given 2GHz chip. In fact, it doesn't mean much at all except a lot of publicity and drooling from people who don't know any better.
 
There is nothing inherently better about 4 GHz. Using GHz as a metric of processor performance has been highly misleading since about 2000, and was primarily perpetuated by Intel. Around that time AMD and Intel designs started to diverge, with AMD opting for doing more work in a clock cycle, while Intel continued pushing for more cycles that do less work.

To paraphrase from AMD's consumer level documentation, imagine an adult and a small child walking together. They both cover the same amount of distance, they both go the same speed. However, the adult (AMD) takes significantly fewer steps than the child (Intel). The child has to take many short, fast steps to do the same work that the adult does in one slow step.

This analogy is somewhat dated now, as it is comparing the AMD K8 Architecture vs the Intel Pentium 4 Netburst Architecture.

It is my belief that while 4GHz will eventually become commercially viable, there will be substantial diminishing returns for the cost benefit one gets from faster processors.

Heat and power have already been touched on in this article. Another issue is propagation delay. With each new generation of processors, the transistor count grows. The progression is approximately linear, with two cores having approximately twice as many transistors as a single core (technically there are more due to the interconnection overhead). So for a given core, as new instruction sets are added, better branch prediction is implemented, et cetera, the number of transistors between the data entry point and the data exit point grows. Each transistor takes a minute amount of time to respond to a change in voltage. Although this time is practically nonexistent, when you put a few million of these in a row, that time starts to add up. The longer the chain, the greater the minimum time a clock cycle can be.

This is why we are seeing the sudden expansion into multiple cores and multi-threaded programming. We are very close to the limit of how fast we can realistically make a processor. So now, to get more performance out of them, we are increasing the number of cores they have. Eventually you can expect to see special purpose and general purpose multicore hybrids (like the Cell Processor on the PS3). GPUs will also follow suit.

This is strictly my opinion, mind you, and applies only to the current architecture of AMD/Intel processors. Additionally, this is all null and void if we move to quantum computing or some other new foundation.
 
Intel also changed the way they calculated the clock speed after the P4. P4s went up to like 3.4 GHz or something, but when the core microarchitecture came out, they changed the way they calculated it, so a 2GHz chip was still actually faster than the 3.4GHz P4.
 
Flynnstone said:
The race is still on. It's just changed.
The transistors are like capacitors. E = energy, C = capacitance, V= voltage.

E= 1/2 C*V^2

As you toggle these transistors faster, the power goes up proportionally.
But to go faster, you need more voltage and is related with a square. So doubling the voltage causes power to go up by a factor of 4!
The die shrinks, from 130 to 90 to 65 to 45 ... causes the C factor to drop. So dropping from 130nm to 65nm the "C" (capacitance) should in theory drop in half. Not totally true, but close enough. But shrinking the die also means the total die area is smaller, so power density goes up and this is a problem. For example, say a 130nm cpu has 2 in^2 area and dissipates 100 watts, thats 50 watts per square in. But now take a 65nm cpu with 1 in^2 area and still at 100 watts. Now the power density is 100 watts per square in. Double!
Now back a few years, they could reasonably extrapolate speed and power for these CMOS cpu. Drop size by 2 (130nm to 65nm) and power dropped by a factor of 2. (equation above). But when they neared the 90nm mark the transistors started leaking more. This drop by factor of 2 didn't hold anymore.
Basically it gets hugely expensive to go faster now. So now they try to do more in parallel, hence dual cores & quad cores and more.

But its still about getting more work done is less time.
Click to expand...

Wow AMAZING...Never knew it worked that way. I feel smarter LOL
 
Seymour Cray (1925-1996) put it well:

Cray said:
If you were plowing a field, what would you rather use? a strong ox, or 1024 chickens ?
Click to expand...

I remember at uni, 15 years ago, they just got in a Cray T3D or something like that. That machine ran at about the same speed as a Pentium 100, but had something like 2TB of RAM, and 4PB of disk storage. And insane bandwidth to the RAM and disk array.

Try going through a 1TB randomised database on a modern high end workstation ($10k+), and you'll still spend days waiting for the disks to spin round. Even if a 1TB HD only costs $500, the CPUs will still be idle most of the time. That Cray machine could hold it all in RAM and shovel it into the CPUs and sort it all out in a couple of hours. And that was 15 years ago.

Course, nowadays, people are saying lets add a second ox, instead of breeding an even bigger ox.
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back