It's 2010 - why aren't CPU speeds faster?

Aaahhh, that makes perfectly sense. Considering how long we have multicore for already, when are the software developers are going to pick up I wonder ?
Do they already classify what PC games have been optimized for 2 or even 4 cores ?
And by the way .... how on earth should I be able to pronounce your screen name, let alone remember it ?
Like Drescherjm said, it depends on the developer. Multisocket systems have been around for many years, however because of (mostly) the price it has remained mostly a high end arena for servers and power users (think content creation). So while multithreading applications is nothing new the vast majority of applications were still single threaded. Basically only things like Adobe, film editing, and web/data applications would have the ability to utilize multiple processors. And if you pay attention to benchmarks, content creation apps display the best scaling with # of cores. Games don't display as much, they've only just started to have multithreading (until about 2006-7 99% of games were single) and are still fairly immature implementations of multi-processor support, but it's always getting better.

There is a bit of performance overhead just from needing to coordinate multiple processors that won't allow double the cores to mean double the performance. How well the application is multithreaded makes a big difference. The threads won't necessarily have the same workload or be needed simultaneously. There are many, many dynamics to it. But generally more cores is better.

On a single core system things would run like this example:
You'd be encoding a video and trying to browse the web at the same time

The processor has to switch between tasks constantly to try and keep your browsing going smoothly and keep making progress on the encode. Very difficult to keep things running smoothly when encoding can use 100% of your cpu's capacity, and would use more if it could. This switching will make things seem laggy, stuttery, and so on.

While on a dual core the entire encode can be offloaded to processor #2 and can be running 100% cpu usage all day long. Meanwhile cpu #1 is sitting at 15% with whatever you're doing. In terms of user experience you will be running applications just as smoothly as when the single core is not encoding video, but you are doing encoding that video at the same time.

The power of multicore lies in that. You can run demanding applications on separate cores with virtually no effect on the performance of CPU1 (the core that windows is running on) and can do your normal web browsing/word processing and so on just as fast as a single core processor that isn't running those demanding things. So it all relative, the performance increase entirely depends on what you are doing. It's almost impossible to quantify the improvement in user experience of running multiple things at once with no noticeable performance loss. But it sure is nice, that's for sure.
 
This is the mantality I just do not understand. What is the point, or even logic in having several appz running at the same time? Like for example, how can it make any sense whatsoever say doing a doc for school in MS publisher while decoding a movie with a game running in the back ground all the while talking on MSN messenger and using Ventilio-AND having 14 porn web sites open at the same time?

LOL

I laugh, but seriously how does that make any sense at all? If I was doing a report for school I wouldn't be doing anything else period! I wouldn't be trying to decode a stupid movie. I wouldn't be talking on MSN to some idiot. I wouldn't be veiwing 45 web sites at the same stupid time. Do you see what I'm getting at here? The whole mantality of doing all these things at the same time is retarted to say the least and to me is nothing but a silly excuse to justify a quad core or 6 core CPU. Geez.....

If you are writing a report and pulling information in from various websites and an outlook email account and you are collaborating on the report with another person you may need all those programs open at the same time. Plus if I am making a multimedia presentation that has a video in it, I may need to resize/re-edit it and reencode. All of this is perfectly plausible in a modern work environment. On top of that, most of us don't get to work on one thing from beginning to end. I have to jump from project to project.

How can you seriously NOT see the purpose of multitasking capability.
 
How can you seriously NOT see the purpose of multitasking capability.

Because some people are rather boring people who'd rather denounce technology they have no use for because all they do is chat on MSN while browsing some sites and playing Second Life while listening to music and checking their Facebook/Myspace inbox every ten seconds and...

Oh wait, even boring, low-tech people seem to be able to use at least a dual-core CPU :p
 
Be happy it has turned out this way. They have been able to focus more on improving the core architecture of the cpu, which means we'll get better cpus. Higher ghz eventually with all the evolution of current cpus and future cpus.
 
While the advances in architecture, die shrink, etc have been good,

They still don't compete with raw power. Hence why most of us are so eager to take our processors and overclock them to the max they will handle to get every last mhz out of them.

Dual/quad cores are fantastic, but there reaches a point (we are well past it in most cases) where the 2nd-4th cores are doing nothing more than running our background apps and windows, meanwhile our active application or game is using 1-2 cores only.

Compare an original Core 2 Duo to the newest i7 6 core extreme thats coming out and I think you'd be pretty disappointed in clock for clock performance in your average application/game/windows experience. Sure there are cases where the new architecture or extra 2 cores MAY make a difference (and in huge in some cases that will use all 6 cores actively), but for the most part you gain almost nothing. The increase in cache makes most of the difference you actually see, along with things like upgrading to the triple channel DDR3.

Also this comparing a 6x2ghz to 12ghz is just silly. Almost no applications or games used on consumers desktop utilitize 4 cores effeciently, let alone 6. Not to mention the scaling loss. a 6 core 2 ghz machine would have a hard time beating a dual core 3.0 in most real life consumer uses, forget trying to compete with a hypothetical 12ghz cpu.
 
I feel baited into this discussion for some reason. Probably due to my semiconductor processing background.

While the ignorant (I don't mean that in a bad way, I don't have a clue about a ton of stuff) are unimpressed with the lack of clock advancement, I'm pretty amazed with the work they've done.

Things off the top of my head why we aren't seeing bazillion ghz processors:
Physics is a bitch. Velocity saturation is about 10^7 cm/s, across a 22nm (~16nm actual) channel is not the issue, as that's ~10^-14 s (ideally, although I'm not sure how close they get to VSat), the problem is clock propagation through the wires (huge!), accumulation and relaxation time (static and dynamic capacitances), power consumption, traps, and a dozen other things I don't feel like saying.

Will they overcome some of these problems? Absolutely. Sort of like moving to HfO2 enabled physically thicker gate dielectrics which, due to the much higher relative dielectric constant yielded thinner oxide equivalent gates. Of course the dirty secret was that getting the gate metal right was/is the much tougher part of the deal. Other solutions will come and go, which is exciting. A pentium 4 on a modern process would go faster than it did when it ran into a wall, but the gamut of other issues are still there (like the whole ultra-long pipeline architecture). Smaller, parallel cores makes a lot of sense, since it dramatically shortens the data path of any operation (how many miles of copper the signal is propagated over), which solves a LOT of problems.

That said, if it were easy, everyone would be doing it. There's no doubt they could make a higher clocked dual core uberprocessor, but the market would be small and the expense huge. That ship has sailed.

In short, process engineers, materials scientists, and the whole crew had a paradigm shift away from pushing the clock speeds to oblivion to a more balanced, energy efficient (watts per calculation), processors operating under safe(r) operating conditions. The concept of turbo is awesome. There's a reason you can overclock the crap out of present systems--they're built pretty safely.

The hardware certainly precedes the software, but I have no doubt that over time as multiprocessors become universal that developers will find a way to build systems around it. Branch prediction and scheduling will continue to improve. It's all good.
 
The Netburst design was completely aimed at the MHz race, with its huge pipeline and small cache. It was supposed to scale to 10 GHz. The fact that Intel basically resumed developing the PIII design says a lot about the viability of the MHz race, i.e. that it was a bad idea and good riddance :)
 
I feel baited into this discussion for some reason. Probably due to my semiconductor processing background.
[...]
the problem is clock propagation through the wires (huge!), accumulation and relaxation time (static and dynamic capacitances), power consumption, traps, and a dozen other things I don't feel like saying.

So the main issue is not clock propagation. It's not particularly difficult to run a super high frequency clock on a chip.

The actual issue is that you can't put a useful amount of logic between flip flops and still meet setup and hold times when running on these really small clock periods. Smaller processes haven't been speeding up at the same rate they used to, if it all. The P4 was an attempt at circumventing this by having small amounts of logic between flops, culminating in an unusually large pipeline which ultimately hurt performance.

And I say this as someone who designs integrated circuits for a semiconductor company that does x86 CPUs, among other things.
 
So the main issue is not clock propagation. It's not particularly difficult to run a super high frequency clock on a chip.

The actual issue is that you can't put a useful amount of logic between flip flops and still meet setup and hold times when running on these really small clock periods. Smaller processes haven't been speeding up at the same rate they used to, if it all. The P4 was an attempt at circumventing this by having small amounts of logic between flops, culminating in an unusually large pipeline which ultimately hurt performance.

And I say this as someone who designs integrated circuits for a semiconductor company that does x86 CPUs, among other things.

Thanks for the specific clarification--I had forgotten the motivation behind the ultra-long pipeline in the p4 architecture, but knew it had something to do with clocking. Just had the wrong part ;-) I'm glad you pointed that out.

Still comes down to transistor delay and the inherent capacitances of the system lengthening the response time of each logic step.
 
I feel baited into this discussion for some reason. Probably due to my semiconductor processing background.

While the ignorant (I don't mean that in a bad way, I don't have a clue about a ton of stuff) are unimpressed with the lack of clock advancement, I'm pretty amazed with the work they've done.

Things off the top of my head why we aren't seeing bazillion ghz processors:
Physics is a bitch. Velocity saturation is about 10^7 cm/s, across a 22nm (~16nm actual) channel is not the issue, as that's ~10^-14 s (ideally, although I'm not sure how close they get to VSat), the problem is clock propagation through the wires (huge!), accumulation and relaxation time (static and dynamic capacitances), power consumption, traps, and a dozen other things I don't feel like saying.

Will they overcome some of these problems? Absolutely. Sort of like moving to HfO2 enabled physically thicker gate dielectrics which, due to the much higher relative dielectric constant yielded thinner oxide equivalent gates. Of course the dirty secret was that getting the gate metal right was/is the much tougher part of the deal. Other solutions will come and go, which is exciting. A pentium 4 on a modern process would go faster than it did when it ran into a wall, but the gamut of other issues are still there (like the whole ultra-long pipeline architecture). Smaller, parallel cores makes a lot of sense, since it dramatically shortens the data path of any operation (how many miles of copper the signal is propagated over), which solves a LOT of problems.

That said, if it were easy, everyone would be doing it. There's no doubt they could make a higher clocked dual core uberprocessor, but the market would be small and the expense huge. That ship has sailed.

In short, process engineers, materials scientists, and the whole crew had a paradigm shift away from pushing the clock speeds to oblivion to a more balanced, energy efficient (watts per calculation), processors operating under safe(r) operating conditions. The concept of turbo is awesome. There's a reason you can overclock the crap out of present systems--they're built pretty safely.

The hardware certainly precedes the software, but I have no doubt that over time as multiprocessors become universal that developers will find a way to build systems around it. Branch prediction and scheduling will continue to improve. It's all good.


wut :confused:

need the layman's translation. :D
 
wut :confused:

need the layman's translation. :D

The electron moving gnomes got lazy. So we had to use a lot more of them instead to get the work done. :D

Here's the condensed version of what I said:

The computer architecture has changed to take advantage of the present level of semiconductor manufacturing instead of building an architecture designed around a certain clock speed and ramming it into submission. That's why clock speeds haven't increased, but present processors are able to do a lot more per clock cycle, thus are a lot faster.

They (being the process engineers and the like) will continue to make improvements to the fabrication process (although the advances might slow to sub-Moores law levels, maybe not) which new architectures will take advantage of.

Then I said some wrong things (or just had them mixed up in my head) and some other things to frighten people. No biggie :D

Hope that helps
 
Thanks for the specific clarification--I had forgotten the motivation behind the ultra-long pipeline in the p4 architecture, but knew it had something to do with clocking. Just had the wrong part ;-) I'm glad you pointed that out.

Still comes down to transistor delay and the inherent capacitances of the system lengthening the response time of each logic step.

Not to mention the need for more clock domains when you crank up the Herz as otherwise even on a tiny chip you'd get terrible clock skew. Not that synchronizing clock domains is that easy :p

The Tb/TB telecom industry chips my housemate designs at his company rarely run much above 500 MHz for a reason :)
 
A big "Thank You" to everyone who posted useful information - I learned a lot more than I expected. :cool:

Just to be clear, I posted the question out of curiosity and not out of a sense of disappointment.

I'm quite happy with the performance of the the cpu's I've purchased over the last year or so, even if they don't run at 10GHz... :)
 
cpus have hit a speed wall notice the rush away from single core to multicore but speed has remained the same or nearly the same as previous generations i expect this trend to continue for a while till x86 is replaced
 
cpus have hit a speed wall notice the rush away from single core to multicore but speed has remained the same or nearly the same as previous generations i expect this trend to continue for a while till x86 is replaced

What about the x86 ISA do you believe is preventing clock speed increases? Process improvements are independent of the chip architecture.
 
What about the x86 ISA do you believe is preventing clock speed increases? Process improvements are independent of the chip architecture.

The x86 ISA these days is this huge blob of ucode-filled, uOP decoding engine stuff pasted as frontend onto a lean RISC-like architecture. On Atom it's one of the largest sections and thus not insignificant in its overhead and power usage. ARM in comparison uses a fully hardwared ISA and thus skips all this overhead while simultaneously making Atom look fat :p
 
not to mention 1 core of an i7 @ 2.6 is more efficient / faster than 1 core of a p4 @ 2.6 or even 3.6 most likely

i7 @ 2.6 (no HT) is probably about equal to a P4 @ 5.5+ ghz. 3.6 ghz is a far too conservative guess.

Since a core 2 in single core was about equal to a pentium 4 at twice the mhz in most cpu limited game benchmarks/fps counters, usually tested best in older games)..
(tested on my Pentium 3 (which was clock for clock faster than a P4), P4 at various speeds, and my core 2 quad, in the game VietCong, on the introduction level (which seems to have the highest FPS hit).
 
The x86 ISA these days is this huge blob of ucode-filled, uOP decoding engine stuff pasted as frontend onto a lean RISC-like architecture. On Atom it's one of the largest sections and thus not insignificant in its overhead and power usage. ARM in comparison uses a fully hardwared ISA and thus skips all this overhead while simultaneously making Atom look fat :p
Right, I'm actually peripherally familiar with x86 micro-architectures as I do IC design for a company which makes them.

Unless the logic which decodes to uOPs is the critical timing path, it's impact to clock speed is mostly negligible. It may well add some additional cycles of latency onto execution, but again, that won't impact clock speed.

Additionally, even if the micro-architecture is comparatively slow, switching away from it isn't going to usher in a new era where we see consistent clock speed increases. You might get an initial boost from switching to faster architecture (clock-wise), but that's only once.
 
Back
Top