Is CPU technology topping out?

[Fjodor2001]

It's true that CPU gains from new generations are lower. But gains weren't free. 486 CPUs used low enough power to be used in cellphones and Pentium chips had lesser cooling than some PCH chips today! We can't afford to have that magnitude of increase in power usage anymore.

Going from PII 233 MHz to P4 2.0 GHz increased the clock frequency alone by 858%, but only doubled TDP (from 34.8W to 71.8W as you wrote).

Even if Intel could "afford" to double TDP in the next CPU generation(s), they can no longer do such impressive gains in clock frequency. Otherwise going from 3.5 GHz 77W IB CPUs we'd get:

77W x 2 = 144W TDP
3.5 GHz x 858% = 30 GHz

Now that would be impressive!

But for some reason the CPU frequency increases we've seen all the way going from 5 Mhz 8086 to 3.5 GHz 3770K has topped out. Or actually, it already started topping out with the 2.0 GHz P4 and later 2.4 Ghz C2D (C2D has completely new CPU architchture so its frequencies cannot be directly compared with P4). That P4 was 11 years ago, and the C2D was 6 years ago. Since then the clock frequency has not increased much...

 

[IntelUser2000]

You are not wrong. I guess this situation is a good example of low hanging fruit being taken advantage of in the last decade.

Every new process generations gave significant reductions in voltage, when nowadays it barely gets reduced and lots of tricks have to be employed. The days where transistor drive current was the sole factor in clock speeds may be over as well. 15 years ago even ma and pa of the world needed faster computers, nowadays even enthusiasts don't need them.

 

[Ferzerp]

15 years ago even ma and pa of the world needed faster computers, nowadays even enthusiasts don't need them.

To me, this says more about the lack of advancement in other areas than anything else. We can always find a use for more speed, but if something else is holding us back, we do have to address that area first.

My view is a lot of it is due to the scale out vs. scale up design philosophy. Scale up will improve everything, but it is very difficult, and so a lot of effort has been invested in scale out designs. Scale out designs though, require software and problem sets well suited for scale out architectures.

Put more simply, I'd rather have a dual core + HT processor of architecture X running at 7Ghz, than a quad (or likely even hex or octo) core+HT processor of the same architecture running at 3.5Ghz. That's not to say clock speed is everything, if a processor can accomplish more at a given clock, that is fine too. What isn't fine is just throwing more cores in to processors and saying "Ok, all you software developers, *you* make things faster now". That philosophy pushes performance improvements from the people who are best at it (the cpu companies) to *every single developer in the world*.

Which of those two scenarios creates for a more consistent improvement, would you think? The Intels, AMDs, nvidias, etc of the world working on the problem and improving the performance of every application (that is processor bound), or expecting every single developer to suddenly start using perfectly parallel code and not just pushing out something that functions?

 

[PaGe42]

Which of those two scenarios creates for a more consistent improvement, would you think? The Intels, AMDs, nvidias, etc of the world working on the problem and improving the performance of every application (that is processor bound), or expecting every single developer to suddenly start using perfectly parallel code and not just pushing out something that functions?

There is a third scenario: let the Microsofts, Oracles, or perhaps you working from your garage (or living room, if you prefer) create a software infrastructure that allows every single developer to create parallel code just as easy as serial code. How to do it? I don't know. If I knew, I would create it.

 

[poohbear]

The shift has obviously gone to the mobile sector, the focus is entirely on what they can do for that sector. Desktop advancements isn't really a huge concern as modern CPUs can already handle everything thrown at them & then some.

 

[PlasmaBomb]

160W? That's doubtlessly an 8C/16T unit, if not more. Eight-core SB-E at 2.7 GHz only pushes 130W.

Unless the article is wrong (which is possible) it's a 16C/32T processor...

 

[lehtv]

Unless the article is wrong (which is possible) it's a 16C/32T processor...

I think 160W for a 16C Haswell is plausible. While it's not a die shrink, you can still expect a bit more optimized power consumption. And the 16 cores don't have to be fast ones, that CPU is obviously built for heavy multithreading instead of per-core performance. An Ivy Bridge i7-3770S has a TDP of 65W, the Haswell should achieve 4C/8T at 40W with a combination of new architecture (integrated voltage regulator, anyone?), sub-3.0GHz clock speed and lack of IGP. If an eight core haswell needed 160W to operate, that would indicate a 4GHz+ clock.

 

[Ventanni]

And the TDP went from 34.8W to 71.8W. The fact is last few decades or so of massive CPU gains came from power usage increases that equal the performance gain. That happened with GPUs too. Which is why Nvidia's claim of GPUs being on a 3x Moore's Law cadence deteriorated pretty quickly.

I'm in full agreement with the above sentiment. I feel CPU performance is rising just as quickly as it ever had, but the only difference today is that we have practical TDP limits to contend with. The same is true for the GPU sector. Sure, it was nice to double the performance back in the day from generation to generation, but we also saw a general, sometimes significant rise in power consumption as well.

The "excitement" however is in the mobile sector these days. I'm wondering just how long it'll be before we just put our phones in a cradle and that becomes our PC tower.

 

[IntelUser2000]

Put more simply, I'd rather have a dual core + HT processor of architecture X running at 7Ghz, than a quad (or likely even hex or octo) core+HT processor of the same architecture running at 3.5Ghz.

Assuming everything is being compared at peak, 2x frequency equals 8x the power use. If we say cores take up 80% of CPU power, then that equals in 6.6x the TDP.

At 600W or so you'll probably need at least the best watercooling setup so it can run without frying itself. Don't forget that overclocking will need even more than that.

While quad core users with 3.5GHz base clocks can up theirs to 5GHz with likely less cooling than the stock 7GHz one.

Not to mention Intel will charge like a top-end Xeon E5 chip since having it yield reliably at 7GHz is extremely low.

The Pentium 4 Northwood had more than 2x the pipeline stages of Pentium III to get less than 2x the frequency boost at the same process node. The chip also got more complex, used more power, and lost IPC. At some point of diminishing returns I think its just smart to try to pursue something else to even the "costs"(which can be $$, die sizes, time, etc).

To me, this says more about the lack of advancement in other areas than anything else. We can always find a use for more speed, but if something else is holding us back, we do have to address that area first.

This is basically showing lack of knowledge than anything else IMO. The same can be applied to battery and other renewable technologies. Some believe in the conspiracy that rich smart people are deliberately hindering advancement in battery technology.

Although I can't give a good example for battery technology, I can give for why solar panels have hard time in reducing prices.

I'd like to acquire a cheap solar panel system, so I've computed some costs from eBay. They have major manufacturers selling there so I figured why not.

Sure you can get a solar cell that costs about $0.50/watt or so. That seems wonderful, doesn't it? But add the costs of encapsulation, the glass+frame, the wire connections, electronics, and can easily go over $2/watt. In fact the professionally built systems can go under $1.50/watt(making DIY setups hard to justify) meaning they basically cut costs with mass production. Bigger setups for covering your house may end up at $1.25/watt or so.

Now imagine some manufacturer decides that they'll give away their solar cells, making the solar cell portion $0/watt. You'll still end up with almost $1/watt because of what I call the "support infrastructure".

I assume its the same with EV technologies and battery. If the pace of innovation could have been maintained, it likely would have been.

 

[Ferzerp]

Assuming everything is being compared at peak, 2x frequency equals 8x the power use. If we say cores take up 80% of CPU power, then that equals in 6.6x the TDP.

At 600W or so you'll probably need at least the best watercooling setup so it can run without frying itself. Don't forget that overclocking will need even more than that.

While quad core users with 3.5GHz base clocks can up theirs to 5GHz with likely less cooling than the stock 7GHz one.

Not to mention Intel will charge like a top-end Xeon E5 chip since having it yield reliably at 7GHz is extremely low.

It was a conceptual device, not to be taken literally.

If you'd prefer, I'd rather have two cores (plus HT) of double the single threaded performance than 4, 6, or 8 cores of 1x the single threaded performance.

 

[IntelUser2000]

If you'd prefer, I'd rather have two cores (plus HT) of double the single threaded performance than 4, 6, or 8 cores of 1x the single threaded performance.

I don't think that changes the overall picture too much. There's something common with three following scenarios:

-Light strength per area of unit relative to distance
-Number of transistors and single threaded performance
-Why an ant can haul several times its own weight while human can barely haul its own weight

So what's common? The relation is all exponential, or specifically square of the number. 2x the distance away from a light source results in 1/4x the light strength. 2x the performance per clock results in approximately 4x the amount of transistors. Longer muscles aren't stronger but fatter ones are(basically mass grows LxHxW while strength goes up by HxW).

Comparing CPU cores that are in similar process node and with the designs at the peak, Intel cores are 2x as large as AMD's(at least until Bulldozer muddied the comparisons). And the IPC advantage of Intel cores are roughly in average square root of size difference.

Pentium M was 30% faster per/clk than Pentium III, Core 2 Duo was 20% faster than Core Duo and 30% faster than 90nm Pentium M. Then Intel announced that they'll stick to a rule in Nehalem where they'll only implement circuits if it gives them 2% improvement in performance for 1% increase in power. AMD also said with Bobcat and Bulldozer they'll go for "Knee of the Curve" improvements(or in other words, the "low hanging fruit").

Eliminating all the enormously complicated things that happened, since then on the Intel side the new architecture gain went down from 20-25% down to 10-15%, while AMD got reduced IPC with Bobcat and Bulldozer.

 

[Ferzerp]

I think you're missing the point by concentrating on the exact details. It's just to illustrate that improving per core performance is much better for most scenarios than just throwing more and more cores at it (which once you have a few, improves absolutely nothing unless you are using one of the few applications that it helps on).

Gluing on more cores is easier, but it's not a direction that I want to see continue.

 

[IntelUser2000]

Actually there's a much better example. You should look at the ULV CPUs before the first gen Core.

Pentium M 0.13u ULV
-1GHz, 7W

Pentium M 90nm ULV
-1.2GHz, 5.5W

Core Solo 65nm ULV
-1.33GHz, 5.5W

Core 2 Solo 65nm ULV
-1.2GHz, 5.5W

Core 2 Solo 45nm ULV
-1.4GHz, 5.5W

Gluing on more cores is easier, but it's not a direction that I want to see continue.

I think you'll be even more unhappy if/when we move into hetereogenous era and performance becomes even more dependent on application, usage scenarios, and optimization. Or when desktops become an even smaller fraction of the market(which is inevitable whether or not Intel succeeds in their mobile strategy with Ultrabooks and Atoms). We all want to see the advancement at the rate we're used to, but I don't think we'll see it, at least in a traditional sense.

 

[Azuma Hazuki]

CPU tech may hit the limits of silicon around the 10nm process node. However, I think at that point or before we'll see advances in heterogeneous computing, solid-state storage, and maybe novel forms of memory like phase-change RAM.

If PRAM works out, or if some memristor-based format appears, it may obviate the need for a RAM/storage dichotomy entirely. Imagine something that's basically a bunch of CPU cores with four levels of cache, L4 being a combination of RAM and storage space in terms of usage. A thinking rock, basically, and it should be nice and efficient too.

 

[ShintaiDK]

CPU tech may hit the limits of silicon around the 10nm process node.

Seems both 7nm and 5nm is secure as well. Maybe sub 5nm will hit a limit.

 

[lehtv]

Seems both 7nm and 5nm is secure as well. Maybe sub 5nm will hit a limit.

Source?

 

[ShintaiDK]

Source?

http://www.engadget.com/2012/05/14/intel-sets-sights-on-5nm-chip/
http://www.slashgear.com/intel-roadmap-reveals-plans-for-5nm-chips-14228167/

 

[lehtv]

Thanks. So exciting to see this technology is still developing fast. 22nm to 5nm is a huge, huge difference, similar to the difference between a 90nm Pentium 4 and a 22nm Ivy. That was eight years ago, so I guess we can expect 5nm around 2020.

 

[Fjodor2001]

Thanks. So exciting to see this technology is still developing fast. 22nm to 5nm is a huge, huge difference, similar to the difference between a 90nm Pentium 4 and a 22nm Ivy. That was eight years ago, so I guess we can expect 5nm around 2020.

Yes, the improvements in manufacturing process technology is still impressive. Unfortunately that does not directly translate into better CPU performance. The question is what Intel intends to do with future process technology advancements? Grow the IGP, add more CPU cores, or make smaller dies which consume less energy? Regardless, we will not see the kind of CPU performance increases we saw going from 8086 to P4 2.0 GHz or C2D, despite the process technology advancements...

 

[lehtv]

Indeed, performance won't see the same boost as going from 2004 to today, but the mobility of that performance will increase so much. I would not be surprised if in 2020 we could play modern games on a 4K monitor with the IGP of a laptop, and still get decent battery life

 

[cytg111]

What gives you that idea? AMD is all about moar "cores" regardless of "core" performance (and whether or not a "core" is actually a core)

As the entire industry is going mobile and tries to approx zero watts for their devices, wouldnt it be an vector for amd to go the opposite direction? Invent a package that can handle 400watt tdp and slam loads of loads of transistors at it..