Why have CPU's stopped at ~1.8 GHz?

[yada]

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

[BradC]

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

You might be a decade or so behind. CPUs have been able to hit a qualified 5GHz for a while now.

[Ampera]

My i7-4790k runs at 4.4Ghz, and it's several years old

[james_s]

Mine is 3.5GHz IIRC. Speeds have hit a plateau though and now the trend has been to add more cores and do other refinements. There are physical laws that we are really starting to run up against. Moore's law cannot go on forever. Fortunately computers are already more than powerful enough for most uses, it's not like 20-30 years ago where a 1 year old computer was hopelessly obsolete and wouldn't even run current software.

[Ampera]

Who knows. Intel has already said their 7nm chips are going on something other than Silicon. Maybe we will see a Moore's law revival?

[Vgkid]

The laptop I was using last year ran at 1.8Ghz, it was perfectly useable. Ran Vista as well.

[james_s]

Who knows. Intel has already said their 7nm chips are going on something other than Silicon. Maybe we will see a Moore's law revival?

For how long?

7nm is approaching the physical limit based on the size of the atoms, they simply can't go much smaller, at least not without a whole new paradigm.

[yada]

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

You might be a decade or so behind. CPUs have been able to hit a qualified 5GHz for a while now.

So pretend I said 5 GHz, why have things stopped at 5 GHz? Why add another core instead of making it faster like we have for the last 30 years? To me adding more cores seems to make more sense, but people tend to buy based on a bigger number. Look at cordless phones: 49MHz then 900 MHz then 2.4GHz then 5.8GHz even though 900 MHz has the best range through walls and buildings since you are just sending voice, which doesn't need much band width.

[Halcyon]

Why add another core instead of making it faster like we have for the last 30 years? To me adding more cores seems to make more sense, but people tend to buy based on a bigger number.

As has already been discussed, we're reaching the limit on what can be done with current technology and manufacturing processes.

But simply adding more cores doesn't necessarily make applications run faster -- It depends on the application. It's still not uncommon to see applications which only run on a single core (or support less than the number of available cores in the system).

[technix]

The overclocking world record was 7GHz on LN2, and it is reported that i3-7350K can often run at 5GHz or above. Hell I have even dialed in a small overclock for my Xeon E3-1231v3 (3.5/3.9GHz instead of 3.4/3.8GHz)

[David Hess]

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

It is a matter of power, efficiency, and memory latency.

Portable devices have a power envelope which limits the CPU to a less complex but more power efficient design which requires lower memory latency (1) and that limits clock speed.  The extra complexity of a desktop processor allows it to tolerate higher memory latency allowing a higher clock speed and greater performance but at a cost of lower efficiency.

A design made to take advantage of lower latency lower clock frequency memory cannot scale up to higher frequencies simply because it will spend all of its time waiting for results from memory.

(1) The core design is intimately linked with the load-to-use latency of the level 1 cache memory.  Low frequency cores might have a load-to-use latency of 2 or 3 clock cycles but high performance processors can tolerate 4 cycles of load-to-use latency.

[yada]

So some part of it is that the memory can't keep up with the CPU? Also as far as desk tops are concerned, if you can't get the transistors any smaller why not go to a bigger die size, stack them on top of one another, or even make the circuit 3 dimensions? Go from the size of a postage stamp to four postage stamps? I realize its not this simple but my question is why?  We started out on a quest to make things smaller way before we knew it would have to fit into a smart phone. More cores doesn't solve the problems that are linear; where the next set of instructions depends on the last.

[BrianHG]

DDR5 has passed 7GHz speeds already.

As for complex devices like CPUs, the next big speed hike after 7nm process will be graphene based transistors expected to be at least 20x faster.  Yes we are talking about 100GHz processors.

[technix]

This requires through-silicon via technology to be ready. Current we already have TSV-linked NAND and NOR stacks, but for processor cores nobody is doing it yet. Thermals may be the main problem here as all dies get very hot at the same time.

[Housedad]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

[David Hess]

So some part of it is that the memory can't keep up with the CPU?

The level 1 cache cannot even keep up with the CPU.  The big advantage of a complex out-of-order processor is that it can deliver higher performance for a given amount of level 1 cache latency and that applies to a lessor extent to the upper cache levels and memory as well.  A complex out-of-order core can take advantage of greater memory parallelism than a simpler core and memory parallelism limits performance.

Also as far as desk tops are concerned, if you can't get the transistors any smaller why not go to a bigger die size, stack them on top of one another, or even make the circuit 3 dimensions? Go from the size of a postage stamp to four postage stamps? I realize its not this simple but my question is why?

Desktops and servers *do* use larger dies.  The highest performance processors use the maximum die size available if only to have as much integrated cache memory as possible.

Power density prevents stacking high performance logic.  For the past few generations, high performance processors have been limited by how much power can be removed from the die so as the die has shrunk, the power *had* to be reduced.  On Intel's recent designs, the processors have to operate at a lower clock rate when power hungry features like 256 bit vector instructions are used to prevent areas of the die from becoming too hot.

If you check the highest performance grades for Intel desktop processors going back several generations, they all have a roughly constant power for a given die area.

We started out on a quest to make things smaller way before we knew it would have to fit into a smart phone.

That is just economics.  Moore's law is about lowering the cost of transistors which means placing more transistors in a given area.

More cores doesn't solve the problems that are linear; where the next set of instructions depends on the last.

That is called Amdahl's law and it is why more higher efficiency but lower performance processors are not a replacement for higher performance but lower efficiency cores.

[Nusa]

Measuring the systems actual performance is better than using the clock speed as a measuring stick.

But if you insist, here's the current records: http://valid.x86.fr/records.html

[Muxr]

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

As others have mentioned today's high performance CPUs run well into the 4Ghz+ (on laptops they are clocked more conservatively to save power).

But you seem to have a bit of a misconception about equating frequency to performance. Frequency is only a part of the equation. There is also something called IPC.. or Instructions Per Clock.

Modern CPUs are capable of executing more than one instruction at a time.. they accomplish this via tricks like out of order execution and superscalar instruction level parallelism. For instance a modern Ryzen CPU can technically work on executing 6 instructions at once. In practice however it is hard to keep all 6 execution units per core busy, because this requires ideal branch predictors and a particular mix of instructions.

Pentium 4 from 2003 could hit 4Ghz back then, but it was way slower than today's CPUs which run at 4Ghz. Because it had weak IPC among other things.

Also it is important to know how we got to 4Ghz in the first place. A monolithic execution pipeline where all the work is being done at once could not run at 4Ghz.. because you could never meet the timings. By the time the signal reaches the farther parts of the execution unit the execution window would be over and the new cycle would need to start before the work was completed, causing a crash. Such CPU would at best be confined to a few hundred Mhz. Yes process improvements usually called "die shrinks" do increase the f-max (frequency) ceiling, but that's only a portion of the story of how we got to 4-5Ghz CPUs.

Clever CPU designers came up with a way to reach even higher clocks. They split up the work in pipeline stages, where each stage does only a portion of the work.. and the work gets propagated to the next stage like a conveyor belt on each cycle. This allows each stage to be much smaller and therefore timings much easier to meet, allowing for high clocks 4Ghz +. For example Ryzen has a 19 stage pipeline (Intel's recent CPUs have 16).

The problem with this long pipeline approach is that every time you have a branch prediction miss.. the entire pipeline has to be evacuated.. because the instruction the CPU was executing was not to be executed yet, the CPU was executing the instruction under the assumption that it's the next in line to be executed but because it isn't it's the result is invalid, because the data the instruction operates on may have to be modified by another instruction, so for a 19 stage pipeline for instance this is 19 cycles wasted. This is where a good branch predictor is paramount. And in case of Ryzen AMD employs some neural net technology, the perceptron to improve branch prediction and keep branch prediction misses at a minimum.

So yes Moore's law is slowing down. A die shrink may give you 10-20% more frequency or it might not.. adding clever features can also improve performance, widening the CPU cores by adding more execution units for instruction level parallelism and other tricks can improve performance as well.. but at the end of the day it's all a matter of balance and power consumption budget, silicon budgets etc..

Leveraging parallelism is the best way to improve performance in applications when possible, because 6 and 8 core CPUs are becoming mainstream. And adding more cores is much easier than improving IPC or Frequency.

[cdev]

I've never really pursued overclocking because I'm afraid my system would become unstable, but I was able to get my computer to boot at well over 5 GHz without any modification whatsoever just by changing the multiplier.. I am using a fairly generic multicore AMD system thats normally clocked at 3.5-4 GHz. I leave it at its defaults because I like my PC to run whisper quiet.

[Ampera]

Overclocking is primarily a game of heat management. Stability isn't something that is going to affect you in the long term really. If it runs stable for a week, it should be fine. Tweaking voltage helps with some stability issues.

A few things about overclocking:

Non-binned CPUs are normally better. If you can identify one of those, they tend to live a longer, more stable life.

Intel CPUs really need the K suffix to be truly overclockable.

There is a limit. It's not a case of "If I pump enough cooling into it, I can throw it to a billion Ghz". There will be a point where the CPU just won't run stable, no matter how much power and cooling you pump into it.

It will void your warranty unconditionally. Even if they market it as "Overclocker friendly", it's still gonna void the shit out of your warranty.

It will decrease the lifespan of your CPU. The extent depends on your level of overclock. It's not gonna be like from a 20 year lifespan to 4. It's more like if you run it 24/7 at a good overclock, a number somewhere along the line will decrease.

Do it if you need too. For me overclocking is a case of I either have tons of money to piss away (I wish I had this problem) or a case of future-proofing. I suggest you get a good system ready for overclocking, and wait until you come to a point where you can't do what you want to do on your machine anymore with your current CPU. At that point, overclock. It's not a game of just gonna throw the specs up a bit, no biggie. Regardless of what people say there are consequences. You are literally operating your machine outside of it's designed specifications, even if it's designed to go out of those specifications.

If you have a stock cooler that came with your CPU, forget about it. If you have an air cooler, be very careful and monitor temps like you would monitor your family in a bad Detroit neighborhood. While having an unstable OC is usually reversible, a 110c OC is not usually reversible. The rule of thumb is never go over 80 @ full load for extended periods of time. Spikes to 90c are alright, but go past that and you are living on the edge. 100c is normally the danger zone for good Intel CPUs, and if you see 100c on your silicon, cut the power, and cut your OC. Modern CPUs aren't as dangerous to overheat as older ones. It used to be that if you were to pump the multiplier too high, you would pop your CPU, but today you can normally be safe with throttling and emergency cutoffs. It's still never a thing to rely on.

Ask people. If you're unsure about OC, and experimental and strange computer configurations in general, ask people. Waiting a day to get a second opinion can mean the difference between busted box and busting box. Nobody is an expert here, not even the people who made the chips. By overclocking you are cementing your feet firmly into hobbyist territory, and we hobbyists almost always ask other hobbyists. It's just the smart thing to do. Follow common sense rules, don't make impatient actions, if it's too good to be true then it is, don't act like you know everything.

I have never overclocked anything in my entire life. All I have done is gather information for the day that I say I am too far behind, I need to put the hammer down, and give my chip a bit of a boost.

To answer the OP's question in my own words, the higher you go on the clock speed, the higher you go on the power and heat. You can see this through the evolution of PC power consumption and cooling. Up until the DX4 line of 486 chips, CPUs were rated for no-heatsink operation. You went from smaller, 8-10 fin heatsinks with a small fan on Socket 3 and 4 machines, to slightly beefier coolers on Socket 5, 7, and 8, and when the line got to Slot 1 and Socket 370, you started to see block coolers, with later sockets like the first Intel LGA sockets and AM2 having larger and larger coolers.

Today AMD has sold chips that require water cooling because with their limited architecture (Before Zen) they had to kick the speed up to stay even remotely relevant. This is where you get the old joke "AMD CPUs are space heaters"

Of course transistors have gotten smaller and more efficient, but the higher you go on clock speed the higher you go on heat and power. With more cores it's somewhat different. Of course it's more heat and power, but not to the extent of clocking higher. This is why most of the Xeon chips have loads and loads of cores at slower clock speeds with fairly moderate TDPs, all on air cooling.

Nobody knows what is going to happen. You can theorize all day, but at the end of said day, there will always be another person, back in history, saying a very similar thing, to be proven wrong later. People for the past thousand years have been making these predictions, and each and every one of them are wrong today. I am willing to believe history in that respect will repeat itself, but I too can be proven wrong (It's quite annoying actually). There may be new transistor tech, there may be new architecture tech. We may fundamentally re-envision the entire concept of computing at it's very core, but nobody knows as for yet.

My personal opinion is that there will be a way around. There always has. Stopping at a wall just gives time for people to work around it.

Have a good day.

[abraxa]

To add to David's and muxr's great responses:

If you want a high operating frequency, you need a small die so you can meet the timings. (Hint: signals travel at the speed of light at most, which becomes very noticeable at such clock frequencies; example: 5GHz=6cm wavelength, meaning a clock cycle's data can only travel 6cm before the next clock cycle begins)
If you increase the frequency, you increase the leakage and switching losses in the transistors, so smaller transistors generate more heat at the same frequency as bigger ones, causing issues with thermal management. (gate isolation is one of the issues)

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

I'm assuming you refer to CPUs used in mobile devices. As David pointed out, efficiency is an issue. I'd like to add to that and say that efficiency is an issue because of both the limited energy supply and the thermal management. Even if you had an unlimited power supply, the CPU ultimately warms the case (not a good heatsink) which you hold in your hand. There is no fan like in a desktop, so power dissipation has to be limited even when supplied with unlimited power, making efficiency paramount.

[Syntax_Error]

Watch this presentation.



In a nutshell: They're working on it. A lot of people are working hard on it, but all the low hanging fruit is gone. Economics is more of a constraint than physics. Physics ultimately will be the final constraint, but we aren't there quite yet. But economics is making even the incremental gains slow down.

[Ampera]

It's probably gonna cost them an ARM and a leg to get past it!

Badum-tss!

[coppice]

Economics is more of a constraint than physics. Physics ultimately will be the final constraint, but we aren't there quite yet. But economics is making even the incremental gains slow down.

Economics will always be the final constraint. The current state is high value things can use quite expensive processes (e.g. III/V), but most things have to make use of something lower cost (e.g volume CMOS). That won't change. Even the high value applications will never be of high enough value to push things as far as physics will allow.

[coppice]

It's probably gonna cost them an ARM and a leg to get past it!

Chips with legs are really old fashioned. If you want to push device technology as far as it can go, its usually just an ARMs race.

[T3sl4co1l]

If they go to a different semiconductor, I'm curious how much dev time and cost they're going to incur on that.

It's one thing to change everything about your fab and dial in good yields again; it's a whole 'nother thing to do that AND do it at the finest pitch yet.

Tim

[coppice]

If they go to a different semiconductor, I'm curious how much dev time and cost they're going to incur on that.

It's one thing to change everything about your fab and dial in good yields again; it's a whole 'nother thing to do that AND do it at the finest pitch yet.

Tim

That's why a lot of research into new semiconductor materials is based on growing them as a layer on silicon, where so many problems have already been solved.

[David Hess]

If they go to a different semiconductor, I'm curious how much dev time and cost they're going to incur on that.

It's one thing to change everything about your fab and dial in good yields again; it's a whole 'nother thing to do that AND do it at the finest pitch yet.

Doing this would conflict with Moore's Law if the cost per transistor is higher and the way to avoid that is to place more transistors in a given area.  Some of the transistor performance in later process generation steps went *backwards* but what mattered was cost per transistor which required a density increase.  Density increases also require lower power per transistor which is what made that an overriding concern over performance in every later process generation; the performance may be worse but the power and density *must* be better.

What is being suggested with the use of a dense but exotic high performance process has been tried in the past and it always lost compared to denser silicon CMOS even when the later was lower performance.

Maybe things will change but is this same exotic process going to be used for the cache memory?  Because unless it is also denser, it is going to have to compete with processors built on silicon CMOS processes which cost the same or less but have many times more transistors available to build more complex out-of-order processors that can operate at higher frequencies despite the same cache memory latency.

To me that says Intel and the server processor manufacturers will be the first to take advantage of an exotic process because it is the only way they can get higher performance per core and they have customers who will pay for it.  ARM still has the option of relying on power and density scaling to make a more complex higher frequency processor and Intel does not because they already did that several process generations ago.

[Muxr]

Yeah. The latest crop of FinFet die shrinks didn't result in any money saving. The cost has gone up but it's at a break even point because of more die per wafer. The advantage in power consumption and f-max is the main advantage to going to FinFet, used to be you also got a cost saving by doing it.

The other thing that's getting really expensive is masks. It's is starting to cost exponentially more to tape out chips on these new processes. Each time the mask set for a new design pretty much doubles the previous one. For big complex chips we're already approaching a $100M territory. Just to tape out a chip. So unless you're a big manufacturer who sells these chips in millions it's not worth the cost.

This is actually a good thing for FPGAs. As more and more applications will find FPGAs to be the right balance between leveraging a new process and keeping the costs down. This in turn should lower the cost of FPGA. I am purely speculating though.

[David Hess]

Yeah. The latest crop of FinFet die shrinks didn't result in any money saving. The cost has gone up but it's at a break even point because of more die per wafer. The advantage in power consumption and f-max is the main advantage to going to FinFet, used to be you also got a cost saving by doing it.

Which just says to me that Moore's Law already ended and the new law involves the number of transistors for a given power instead of price.

A couple months ago there was a great discussion in a video on this subject with graphs and numbers and everything from an older Intel fellow but I lost the link and have not been able to find it.   I do not remember it showing that Moore's Law had ended yet.

Update: just because I love you guys so much (ha!), I went back months in my Slashdot log of posts and managed to find a link to the video by Intel's William Holt, "Moore’s Law: A Path Forward".  The original video that I watched is at the Vimeo link but I hate Vimeo so I also found a YouTube link which may be the same thing but it is significantly longer and I have not watched both again for comparison:

https://player.vimeo.com/video/164169553

[coppice]

The other thing that's getting really expensive is masks. It's is starting to cost exponentially more to tape out chips on these new processes. Each time the mask set for a new design pretty much doubles the previous one. For big complex chips we're already approaching a $100M territory. Just to tape out a chip. So unless you're a big manufacturer who sells these chips in millions it's not worth the cost.

The corollary of that point is that if you won't be able to get a lot of saleable products out of a single die, the die will only be economically viable in a few niches, like DRAM, when the volume for a single saleable variant is huge. Look at a massive family of parts, like an MCU family, and there are unlikely to be more than a few dies involved. Test time operations enable or disable areas of memory, peripherals, etc. Future MCU families will push this further. Its much cheaper to waste most of the silicon, rather than create more optimised mask sets.

[Muxr]

Yeah. The latest crop of FinFet die shrinks didn't result in any money saving. The cost has gone up but it's at a break even point because of more die per wafer. The advantage in power consumption and f-max is the main advantage to going to FinFet, used to be you also got a cost saving by doing it.

Which just says to me that Moore's Law already ended and the new law involves the number of transistors for a given power instead of price.

A couple months ago there was a great discussion in a video on this subject with graphs and numbers and everything from an older Intel fellow but I lost the link and have not been able to find it.   I do not remember it showing that Moore's Law had ended yet.

Yup, it's dead as we know it. But it continues on a new journey. The important part is doubling of transistors each 18 months. Which is also getting increasingly harder to do, due to declining yields. Impurities and defects are a much bigger factor the smaller you go in terms of process pitch.

But there are ways around it. Companies are designing redundant features into the chips and designing them for harvesting/binning more and more, so they can be salvaged (with defective portions disabled).

Also packaging technologies to package multiple smaller dies in a same package. MCM packages have been around for awhile:


There is also really nifty die stacking going on. They use an interposer which is basically a silicon die built on a cheaper process like 32nm or similar, used to stack smaller die built on more expensive process on top.

[David Hess]

Yup, it's dead as we know it. But it continues on a new journey. The important part is doubling of transistors each 18 months. Which is also getting increasingly harder to do, due to declining yields. Impurities and defects are a much bigger factor the smaller you go in terms of process pitch.

This is a great point which I should have made.  Moore's law does not care about *how* the cost per transistor is decreased.  Larger wafers and better yields also lower the cost per transistor.

See my updated post above for links to video of Intel's William Holt discussing the economics of Moore's Law.

The corollary of that point is that if you won't be able to get a lot of saleable products out of a single die, the die will only be economically viable in a few niches, like DRAM, when the volume for a single saleable variant is huge. Look at a massive family of parts, like an MCU family, and there are unlikely to be more than a few dies involved. Test time operations enable or disable areas of memory, peripherals, etc. Future MCU families will push this further. Its much cheaper to waste most of the silicon, rather than create more optimised mask sets.

I have made this point before comparing Intel and ARM in connection with ARM replacing x86.  Like the conflict between Intel's CISC and other RISC processor manufacturers in the past, Intel is not facing a single ARM juggernaut.  Instead they face lots of different ARM manufacturers who all have their own individual product designs.

[timb]

It's probably gonna cost them an ARM and a leg to get past it!

Badum-tss!

Working on the cutting edge can be quite RISCy, but eventually there will be a breakthrough that SPARCs a new CPU ARMs race. Only time will tell which company ends up the ALPHA dog.

[Ampera]

It's probably gonna cost them an ARM and a leg to get past it!

Badum-tss!

Working on the cutting edge can be quite RISCy, but eventually there will be a breakthrough that SPARCs a new CPU ARMs race. Only time will tell which company ends up the ALPHA dog.

How did you get this INTEL? That's pretty EPIC tbh.

[Mattjd]

If I am remembering my electronics class correctly. A good deal of it has to do with being able to only decrease the size so small. If you look at the current equations of an nmos you'll some this constant K_n

K_n ultimately equals (U_n*C_ox / 2 )*(W / L) where

U_n is the electron mobility
C_ox is the capacitance of the oxide layer used to form the capacitor between gate and body (which can be decomposed further but is good enough for this explanation)

both C_ox and U_n are material dependent and are fixed for the most part.

W = width and

L = length

W/L is ultimately the ratio that is manipulated to change. The W and L are whats changed to increase the current. The problem is, we're getting to sizes so small that not many changes can be made anymore.

So this leaves us to having to manipulate C_ox and/or U_n. They could be changed, as there are materials other than Silicon that can used. The issue comes from manufacturing and companies refusal to change. There are limited facility that can actually print these processors, a lot of money would have to be invested to create them with new materials.

[BrianHG]

It's probably gonna cost them an ARM and a leg to get past it!

Badum-tss!

Working on the cutting edge can be quite RISCy, but eventually there will be a breakthrough that SPARCs a new CPU ARMs race. Only time will tell which company ends up the ALPHA dog.

How did you get this INTEL? That's pretty EPIC tbh.

Read these 2 articles:
https://www.extremetech.com/extreme/175727-ibm-builds-graphene-chip-thats-10000-times-faster-using-standard-cmos-processes
http://www.computerworld.com/article/2507086/computer-hardware/ibm-shows-smallest--fastest-graphene-processor.html

Now, I cant say for sure graphene is the future, but, it demonstrates that something new will come along improving silicon's speed performance in time as we are stuck in this current slowdown hump.  But yes, eventually physics will eventually put an end to it all as we one day get to the point where we can construct devices from the atom up...

[Ampera]

Who knows, maybe we can somehow build using electrons. Maybe we will start using subatomic quarks?

IMO once we hit the ultimate barrier of an Atom, it's going to become a game of who can bring the price down and the yield and efficiency up.

[Red Squirrel]

I think we will hit a point where we can only grow in parallel.  More cores/cpus and applications that need that power will just need to be designed to leverage multithreading so each thread can run on it's own core.
 

[Ampera]

That's my point. I think the game is going to become bringing the power consumption and heat down on more complicated dies.

[james_s]

I think we're already to that point, at least to some degree. It's not as if Moore's Law was ever an actual law of physics, it was just an observation of a trend, everyone knew it could not go on forever.

[Ampera]

The theoretical limit is around 5nm according to Wikipedia, with some prototypes working.

[SeanB]

I would also add that current processors already operate at an energy density roughly equal to the surface of the sun, and the biggest issue is to cool this die and get the heat out evenly across the surface, but still have a good insulator to keep the leakage curent to the thermal solution well below what is going to slow the rise and fall times of the signals. Another issue is the speed of light, while the 5GHz might go 6cm in a cycle, it really is only going to be sampled on the rising edge ( or falling edge depending on the logic in the particular part of the die) and will also be only available on the other edge on the origination side. Thus, with C in silicon being a crap load slower than in a vacuum, you are limited to under 1cm of total trace length with polysilicon, and slightly more with a copper interconnect ( thus the IBM patents for copper on silicon interconnects that garnered them so much revenue over the years) so that you can actually have a reliable data transfer of data, and you need to have all traces matched lengthwise and all will be transmission lines with a defined impedance as well.

Thus you see things like clock generation and distribution having most of the silicon die space, and that there are more data lines snaking through than logic, along with doing things in stateless logic and clockless logic whenever possible so that you can shave a tiny bit of time off them, along with shutting off whole swathes of the die between instruction cycles so as to save power lost in the really leaky transistors in there.

Pretty much all at the limits of current processes, and adding extra cores and getting the programmers to make software that will use them as efficiently as possible is the way forward. Thus you see graphics cards with essentially 1000 Z80 processors on board, simple and small, but with very fast memory and really fast sharing of data between them, so that you can do a lot of things in parallel when you need to, so they are more capable than the individual processor core itself. Kind of like an ant nest, where each individual is small and seems unable to do much, but in a collective they are capable of doing massive tasks in a short time, without much coordination other than simple semaphores and limited data transfer.

[T3sl4co1l]

I would love to see "lossless" logic take a spin at ~nm scale.  While the density will be considerably lower (4x or worse I think?), and the clock speed reduced similarly, the power level would make practical to stack CPU dice.  Multicore designs could stack up rather than "around", interleaved with cache and whatnot as needed.

Tim

[Terabyte2007]

Typically 4-5 GHz is all you will see unless you're modding the PC with super cooling, water cooling, overclocking, etc. One of the possibilities is to build up or essentially 3D cores. These would be stacks of cores and by keeping the clock speeds reasonable your could gain effective speed advantages there. Heat will always be your battle with silicon until we move away from this substrate to something entirely different you probably won't see big leaps in processor technology when it comes to speeds. But luckily, I don't think this will be an issue for much longer. 

[brucehoult]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

That was true in the 70s and early 80s. It hasn't been true since around the 80486/68040 era, and certainly not since anything Pentium/PowerPC. Except for the dog called Pentium 4, where they bumped the MHz for marketing reasons, but it wasn't really real.

[grumpydoc]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

That was true in the 70s and early 80s. It hasn't been true since around the 80486/68040 era, and certainly not since anything Pentium/PowerPC. Except for the dog called Pentium 4, where they bumped the MHz for marketing reasons, but it wasn't really real.

Indeed clocks these days are multiplied up for internal use.

[brucehoult]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

That was true in the 70s and early 80s. It hasn't been true since around the 80486/68040 era, and certainly not since anything Pentium/PowerPC. Except for the dog called Pentium 4, where they bumped the MHz for marketing reasons, but it wasn't really real.

Indeed clocks these days are multiplied up for internal use.

Well, yes, the external 100 MHz or 133 MHz or whatever clock is multiplied by 30 or 35 or 40 or 45 to get the internal (and advertised) clock speed, and it's even adjusted dynamically for power consumption, temperature, or load reasons.

[yada]

To add to David's and muxr's great responses:

If you want a high operating frequency, you need a small die so you can meet the timings. (Hint: signals travel at the speed of light at most, which becomes very noticeable at such clock frequencies; example: 5GHz=6cm wavelength, meaning a clock cycle's data can only travel 6cm before the next clock cycle begins)
If you increase the frequency, you increase the leakage and switching losses in the transistors, so smaller transistors generate more heat at the same frequency as bigger ones, causing issues with thermal management. (gate isolation is one of the issues)

I know we are getting to end of mores law as transistors approach the size of 10's of atoms, but why have clock speeds stopped at about 1.8 GHz and they just add more cores? Is it because of RF properties of the signals are causing too much loss? 1.8GHz doesn't seem that fast compared to the pace things have been progressing at.

I'm assuming you refer to CPUs used in mobile devices. As David pointed out, efficiency is an issue. I'd like to add to that and say that efficiency is an issue because of both the limited energy supply and the thermal management. Even if you had an unlimited power supply, the CPU ultimately warms the case (not a good heatsink) which you hold in your hand. There is no fan like in a desktop, so power dissipation has to be limited even when supplied with unlimited power, making efficiency paramount.

are you sure there is a 1:1 correlation in speed and wavelength? a 300MHz (1 meter wave length) signal will travel 6cm at the same speed as 6GHz. I think you are right for the wrong reasons.

[magetoo]

In addition to what's been said, another factor is that for desktop processors Intel has been the undisputed performance king for a long time with no real competition to speak of, and so has had little reason to waste money by trying to squeeze out more raw performance in their designs.  That might change now that AMD has a new generation of CPUs that can actually compete at the higher end.

And most of the peak computation load these days is probably done by GPUs anyway, not the CPU.

[brucehoult]

I would also add that current processors already operate at an energy density roughly equal to the surface of the sun

Hmm.

Skylake-K is according to Anandtech 122.4 m^2, which means with a 95W TDP that's 775 kW/m^2. Wiki says the Sun's luminosity is 3.828e26 W, and surface area is 6.09e18 m^2, so 62.8 MW/m^2.

So the Skylake energy density is almost two orders of magnitude lower.

Still, it's a lot.

Thus you see graphics cards with essentially 1000 Z80 processors on board

I don't think anything would be remotely like a Z80. At least the ones I'm familiar with are more like 1000 Cortex M4F processors, but with far more registers (and therefore longer instructions to encode them), but with maybe eight CPU cores sharing one program counter and instruction decode logic, and also operating as a "barrel processor" running maybe four threads in sequence (also processing the same instruction).

[Ampera]

Even though AMD is back into the performance game, Intel has a DAMN sight more money than AMD, and if they feel threatened they could steamroll AMD (no pun intended) within a year.

[T3sl4co1l]

AFAIK, GPUs are more about wide data than oodles of [independent] cores.  That is, you have, say, a thousand ALUs, and registers and pipelines and all that, but they're all driven by the same instructions from the same decoder and program counter.

Probably, they use a few levels of this approach, so it's not that the whole chip is doing just one series of instructions in lockstep, but a few in parallel, each of which is controlling a lot of parallel computations.

But I haven't done anything with them personally, so take this for what it is...

Tim

[wraper]

DDR5 has passed 7GHz speeds already.

Nope, divide that by 4 and get a real frequency.

[Muxr]

DDR5 has passed 7GHz speeds already.

Nope, divide that by 4 and get a real frequency.

Yup. It's a bit annoying when manufacturers count data rate as frequency.

[james_s]

It would be interesting to see what one could do with 1,000 Z80s, probably best tried in a large FPGA than using actual Z80 chips.

[grumpydoc]

It would be interesting to see what one could do with 1,000 Z80s, probably best tried in a large FPGA than using actual Z80 chips.

Not all that much

Modern processors are, conservatively 10's of 1000's of times faster than the Z80

Just on clock speed the fastest Z80 was 20MHz (the later, more highly integrated Z80181 and the EZ80's did go a bit faster) - modern CPU's are 3-4Ghz so there is a factor of 200 right there, 1000x if you compare a 4MHz Z80.

But we can factor roughly 6x more for the fact that it takes a Z80 about that many clocks on average to execute an instruction and at least 4x for working in 8 bit chunks not 32 - eg for a 32-bit memory to memory add (i.e operands in RAM, result to RAM) takes a Z80 about 80 clocks compared with 4 on an i386 so there's a 20x speed-up there. That brings us to 4000x before we consider more complex operations (how long does a 32-bit multiply take on a Z80?), super-scalar execution, 64 bit ops, SIMD/Vector ops, FPUs that can do 80 bit floating point operations in one clock and multiple cores per die. Oh and the miserly amount of RAM that a Z80 can address compared with a modern CPU.

1000 Z80's is worth maybe 5% tops of a core i7.

Edit2: OK, in an FPGA you could clock them a lot faster than "real" Z80s but even then......

It's still my favourite 8-bit micro though (6809 gets close however).

EDIT: changed 32-but multiply to 32-bit multiply. A 32 but multiply is what my teenage son does when doing his maths homework.

[Ampera]

TBH, my favorite micro is the 32-bit M68000. The register arrangement pleases me, and it's incredibly easy to program for with every instruction you could need in a basic CPU. That being said I have never really programmed much with it, but that's more out of being lazy than anything else.

The Z80 is a neat CPU, but it's IPC ratio is pretty poor compared to other micros of the time (Like the 6502)

Probably my least favourite has got to be the PPC. I first off do not like reduced instruction set CPUs, but to make matters worse, the execution by AIM was terrible. They did, granted, use IBM's POWER CPU which in it's own right is not a bad chip. Even still I don't hate the PPC and it found a lot of use in Macs for years, and even the 6th, 7th, and 8th generations of video game consoles all saw at least one PPC console. Even today, the PPC line is still used by the Amiga.

[james_s]

My favorite classic CPU is the 6502, partly because it was used in so many great systems and partly because it's such a great example of inexpensive minimalist hardware that people figured out all sorts of clever ways to make it do amazing things. It's the core of quite a few arcade games from the golden ere, it was used in the Commodore 64, the original Nintendo (in modified form) and countless other iconic systems of the late 70s through the mid 80s.

[grumpydoc]

TBH, my favorite micro is the 32-bit M68000. The register arrangement pleases me, and it's incredibly easy to program for with every instruction you could need in a basic CPU. That being said I have never really programmed much with it, but that's more out of being lazy than anything else.

+1

The Z80 is a neat CPU, but it's IPC ratio is pretty poor compared to other micros of the time (Like the 6502)

Maybe, the Z80 just needed more clocks to step through its internal state machine but I'm not sure it mattered much.

The 6502 can execute some instructions in one clock cycle whereas the Z80 takes 4 clocks to do the simple stuff and more to do anything complex. But the 6502 is hampered by its lack of registers, reliance on page zero and lack of any 16-bit operations - so in practice most instructions are at least two bytes/cycles.

So, whereas it looks at first sight that a 1MHz 6502 ought to be as fast as a 4MHz Z80 in practice the ratio was closer to 1:2 with a 4MHz Z80 having a slight edge on a 2MHz 6502 - when these were the fastest parts available that gave Zilog the edge. Also the Z80 moved on to higher clock speeds (eventually 20MHz for the CMOS part) whereas I don't think the 6502 got past 3MHz.

You can still buy new Z80's from mainstream distributors - I don't think that's true for the 6502.

[james_s]

The 6502 is still in production and available from mainstream distributors, clock speed up to at least 14 MHz but IIRC there are faster ones available.

http://www.mouser.com/ProductDetail/Western-Design-Center-WDC/W65C02S6TPG-14/?qs=sGAEpiMZZMtVFuKNr6IGvpdkwXR9vVB1

[Ampera]

TBH, my favorite micro is the 32-bit M68000. The register arrangement pleases me, and it's incredibly easy to program for with every instruction you could need in a basic CPU. That being said I have never really programmed much with it, but that's more out of being lazy than anything else.

+1

The Z80 is a neat CPU, but it's IPC ratio is pretty poor compared to other micros of the time (Like the 6502)

Maybe, the Z80 just needed more clocks to step through its internal state machine but I'm not sure it mattered much.

The 6502 can execute some instructions in one clock cycle whereas the Z80 takes 4 clocks to do the simple stuff and more to do anything complex. But the 6502 is hampered by its lack of registers, reliance on page zero and lack of any 16-bit operations - so in practice most instructions are at least two bytes/cycles.

So, whereas it looks at first sight that a 1MHz 6502 ought to be as fast as a 4MHz Z80 in practice the ratio was closer to 1:2 with a 4MHz Z80 having a slight edge on a 2MHz 6502 - when these were the fastest parts available that gave Zilog the edge. Also the Z80 moved on to higher clock speeds (eventually 20MHz for the CMOS part) whereas I don't think the 6502 got past 3MHz.

You can still buy new Z80's from mainstream distributors - I don't think that's true for the 6502.

Fair enough. Maybe poor wasn't the right word, but you still need a higher clock to get things done.


I will also punch the next person who calls a 68k a 16 bit CPU. If you call a 68k a 16 bit CPU you also have to call the 8088 an 8 bit CPU, and the P5 Pentiums 64 bit.

And for those who play video games, the Genesis is 16 bit only if the SNES is 8 bit.

[technix]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

That was true in the 70s and early 80s. It hasn't been true since around the 80486/68040 era, and certainly not since anything Pentium/PowerPC. Except for the dog called Pentium 4, where they bumped the MHz for marketing reasons, but it wasn't really real.

Indeed clocks these days are multiplied up for internal use.

Well, yes, the external 100 MHz or 133 MHz or whatever clock is multiplied by 30 or 35 or 40 or 45 to get the internal (and advertised) clock speed, and it's even adjusted dynamically for power consumption, temperature, or load reasons.

Even microcontrollers now have clock frequency multipliers. For example STM32F407ZGT6 can take an 8MHz crystal and generate a 168MHz clock out of it (8MHz / 8 * 336 / 2 = 168MHz.) this stands true for 8-bit cores too like ATtiny85 (8MHz * 8 / 4 = 16MHz) or PIC16F1455 (16MHz * 3 = 48MHz)

[Alex Eisenhut]

[Nusa]

68000 was a bit of a hybrid, bitwise. 32-bit internal bus, but 16-bit data bus.

The Sega CD had TWO 68000's and a Z80. How many bits is that? I developed games for it in the early 90's. The dev system's been sitting in a box since then. Might still work, but it was tempermental even when I was actively using it. Had to leave it powered up 24/7 to avoid random lockups.

[Ampera]

That's pretty cool. TBH I am actually interested in that. I don't have really any money at the moment, but in the future, would you be willing to sell it?

And yes, the 68k was a hybrid, but at it's heart, it had a 32-bit instruction set, internal data bus, and registers. The only 16 bit part to it was the data bus and ALU. Addressing was 24 bit.

[Carl_Smith]

It would be interesting to see what one could do with 1,000 Z80s, probably best tried in a large FPGA than using actual Z80 chips.

For years I've wanted to build a board with a crap ton of whatever is the cheapest 8 pin micro.  Like 1000 of them on a board.  Hook them all together somehow so they can work on some massive parallel processing application.   Could hook one output pin into an input of the next chip and have a 1000 step pipeline of tiny PICs or AVRs.  Or I could wire up a grid that could load some data across a row then step it down through the columns.

Thing is, it's a solution I don't have a problem for.  So I never did it.  And I know the Core i7 in my laptop would run circles around it.   But I still think about doing it...

[Alex Eisenhut]

Sounds like a bit slice processor?

https://en.wikipedia.org/wiki/Bit_slicing

[TiN]

I've overclocked Intel Celeron (Netburst arch) to 8.2GHz about 10 years ago .
Yes, that is with liquid nitrogen (-196c) cooling.  Intel Core i7 2600K CPUs were very common to do 4.9-5GHz 24/7/365 with decent cooling.
Yesterday released Core i7-7xxx X-series have clocks over 4GHz even for 10-core CPU. So yea, 1.8GHz was actual around 15 year ago with Athlon XP and Intel P4 era.

[grumpydoc]

The 6502 is still in production and available from mainstream distributors, clock speed up to at least 14 MHz but IIRC there are faster ones available.

http://www.mouser.com/ProductDetail/Western-Design-Center-WDC/W65C02S6TPG-14/?qs=sGAEpiMZZMtVFuKNr6IGvpdkwXR9vVB1

Ah, I stand corrected

I wonder what uses those. There again I wonder what uses new Z80s

[coppice]

The 6502 is still in production and available from mainstream distributors, clock speed up to at least 14 MHz but IIRC there are faster ones available.

http://www.mouser.com/ProductDetail/Western-Design-Center-WDC/W65C02S6TPG-14/?qs=sGAEpiMZZMtVFuKNr6IGvpdkwXR9vVB1

Ah, I stand corrected

I wonder what uses those. There again I wonder what uses new Z80s

I don't know who uses the 6502, but it seems an entire organisation is kept alive by it, since http://www.westerndesigncenter.com/wdc/ seems to handle nothing else.

The Z80 core is inside numerous things. I doubt many use the traditional 40 pin DIP Z80 now, but quite a few still use the core, sometimes running at hundreds of MHz.

[Ampera]

It's used in the 100+ USD Ti-83 and 84 line of calculators. I will never buy a single thing from Ti, and for as long as I can I will not touch one. Ti is ripping off hundreds of thousands, maybe even millions of college students. A group of people who are normally gutted for cash. 100+ dollars for a 1 dollar Z80 computer, 25 cent display, 10 penny keyboard, and couple cent case. It's highway robbery.

If you are in the market for a calculator, buy HP. They do a hell of a lot more than Ti, and for the same price. My HP Prime cost around 110 USD, and is 3 times as fast as even the most powerful Ti calculator on the market.

[james_s]

I'd bet it costs a lot more than you estimate to build those TI calculators. They may be overpriced but not to the degree you suggest.

You want an example of highway robbery just look at textbooks. They make trivial revisions consisting mostly of shuffling things around and call it a new edition which you then have to buy at exorbitant cost. Your expensive book is then almost worthless once a new edition comes out.

[Syntax_Error]

It's used in the 100+ USD Ti-83 and 84 line of calculators. I will never buy a single thing from Ti, and for as long as I can I will not touch one. Ti is ripping off hundreds of thousands, maybe even millions of college students. A group of people who are normally gutted for cash. 100+ dollars for a 1 dollar Z80 computer, 25 cent display, 10 penny keyboard, and couple cent case. It's highway robbery.

If you are in the market for a calculator, buy HP. They do a hell of a lot more than Ti, and for the same price. My HP Prime cost around 110 USD, and is 3 times as fast as even the most powerful Ti calculator on the market.

 

Ah, yes. All hail the almighty college student, the noblest of all demographics, entitled to free and cheap...everything. Again, 

[edavid]

68000 was a bit of a hybrid, bitwise. 32-bit internal bus, but 16-bit data bus.

That is not really correct, the 68000 had a 16 bit ALU and internal data bus.  This is obvious from the instruction timings, and has been confirmed by die photos.

[timb]

Just remember that the declared clock speed of a processor is the speed of the external clock only.   They routinely divide that clock a lot before it reaches the ALU.   Different parts of the CPU operate at sometimes vastly different clock speeds internally.

That was true in the 70s and early 80s. It hasn't been true since around the 80486/68040 era, and certainly not since anything Pentium/PowerPC. Except for the dog called Pentium 4, where they bumped the MHz for marketing reasons, but it wasn't really real.

Indeed clocks these days are multiplied up for internal use.

Well, yes, the external 100 MHz or 133 MHz or whatever clock is multiplied by 30 or 35 or 40 or 45 to get the internal (and advertised) clock speed, and it's even adjusted dynamically for power consumption, temperature, or load reasons.

Even microcontrollers now have clock frequency multipliers. For example STM32F407ZGT6 can take an 8MHz crystal and generate a 168MHz clock out of it (8MHz / 8 * 336 / 2 = 168MHz.) this stands true for 8-bit cores too like ATtiny85 (8MHz * 8 / 4 = 16MHz) or PIC16F1455 (16MHz * 3 = 48MHz)

Some even have full PLLs with freely configurable clock trees! The PSoC 5LP can take a 2-40MHz crystal/external clock (or the 3-24MHz Internal Main Oscillator) and connect it to the internal Fractional-N PLL, which can generate any frequency up to 80MHz.

The Bus Clock and Master clocks also have (optional) individual dividers and can be derived from any of the above mentioned sources (PLL, Crystal, External Clock or IMO).

On top of that there are a number of freely configurable clock dividers to use as sources for the various analog and logic block functions (ADCs, DACs, CapTouch, Timers, PWM, SPI, I2C, Logic Gates, etc). This means virtually any peripheral can have its own unique clock that's in phase with the master clock.

There is also an onboard frequency doubler, designed to take a 24MHz clock and convert it to 48MHz for the USB controller.

It's the most flexible (and easiest to use) clocking system I've come across on a MCU.

[Ampera]

I'd bet it costs a lot more than you estimate to build those TI calculators. They may be overpriced but not to the degree you suggest.

You want an example of highway robbery just look at textbooks. They make trivial revisions consisting mostly of shuffling things around and call it a new edition which you then have to buy at exorbitant cost. Your expensive book is then almost worthless once a new edition comes out.

Textbooks are a rip too, that's why everybody's buying a cheaper EBook from some guy who probably scanned it all in.

And no, I definitely think they are that cheap to make. What's even in it? There is no rechargable battery, the screen is barely a screen, the keyboard is as cheap as you can get with actual buttons, and if the RasPi foundation can make a powerful ARM computer on a board for five bucks, I doubt that Ti is doing a Z80 board for any more than a dollar.

[technix]

I'd bet it costs a lot more than you estimate to build those TI calculators. They may be overpriced but not to the degree you suggest.

You want an example of highway robbery just look at textbooks. They make trivial revisions consisting mostly of shuffling things around and call it a new edition which you then have to buy at exorbitant cost. Your expensive book is then almost worthless once a new edition comes out.

Textbooks are a rip too, that's why everybody's buying a cheaper EBook from some guy who probably scanned it all in.

And no, I definitely think they are that cheap to make. What's even in it? There is no rechargable battery, the screen is barely a screen, the keyboard is as cheap as you can get with actual buttons, and if the RasPi foundation can make a powerful ARM computer on a board for five bucks, I doubt that Ti is doing a Z80 board for any more than a dollar.

The real cost might be in the validations TI made for those calculators. To get an calculator allowed in a test takes a lot of validation. The TI 84 Plus SE is authorized to be used in SAT and AP tests. SAT does not allow the TI-89 Platinum but AP still allows it. The Voyage 200 is more advanced a calculator than TI-89 Platinum (both calculators have the same hardware design) yet the V200 is cheaper (at least in local stores here.) The reason behind this was that TI-89 Platinum required more validation than Voyage 200.

[Ampera]

The SAT allows all major HP calculators. It's the ACT that is strict, allowing only the HP48G as the most powerful HP Calculator. While I do have one, they are rare, and expensive.

HP is great for a class, but not for major testing, sadly. Some tests allow different calculators, and if your HP calculator is on that list, that's great. The 48G is often the lowest common denominator with testing. It's allowed almost everywhere the Ti-83/84 is.

The reason HP is banned everywhere is mostly down to the computer algebra system programmed into every modern HP calculator. The 48G does not have CAS (that being said it would be easy enough to implement) and is thus allowed on the ACT, provided you tape down the IR transceiver.

The HP Prime does have a testing mode, but it's not something most places want to bother with.

However, for those out of school, and those just in class, it's an amazing calculator. It's just too good for some testing.

[JPortici]

The PSoC 5LP

dsPICs can do that too.. and go higher in frequency
PIC32 too

and now that i think about it, every cortex i've ever used

[brabus]

I guess one of the reasons for the performance plateau is the lack of demand. Does anyone seriously need more desktop power? No. Today we just want the batteries in our smartphones to last longer.

Multi-GHz, Multi-Core SoC, thin as paper and barely lukewarm: this is the real modern days' miracle.

[Ampera]

That's half the story.

For the longest time AMD has been choking on it's own bad decisions. It finally got 14nm fabs and is pulling into it's former glory with the Zen architecture.

Before Zen, Intel had no real reason to compete. With what? CPUs that were years behind their tech, and even then slower? AMD was kept alive by the fact they had a good graphics division, and in some cases had better price to performance than Intel.

Now that the former reigning champion is back in the ring, we will most likely see some more competition as Intel uses it's almost endless wealth to do something against the cry of Ryzen.

It's almost like a scuffle before the end of the world. As Moore's law comes to a close, with the need of a total rethinking of how we produce CPUs, nobody knows what the hell is going to happen. AMD is at the liberty of the fabs they contract, while Intel has all of their production in house. So if Intel comes out with some new way of going smaller, or expanding in some other form, AMD will have to wait for companies like Global Foundries to catch up, which was one of the main reasons for their multiple year slump.

My opinion is from a strange perspective. I still use legacy machines, and so do a lot of people. The microcomputer revolution has been the brightest and flashiest event in human history, and people are still catching up. I'm not sure what I am trying to say, but maybe we will forget about going faster altogether. Maybe we will come back to enjoy the technology we already have. It is my hope, anyways.

What I'm really trying to say is I'm bummed I missed out on Quantum Link and I want them to bring it back (JK)

It's going to be an interesting few decades for those who can make it. We don't know where the human race is gonna go, but there is one thing we all do know. This has happened before, and we have always gone past it. People have said we are going to stop and return to old ways, but we never have, and from an educated point of view, I think we're gonna end up continuing forward the same way we have been.

[rdl]

I guess one of the reasons for the performance plateau is the lack of demand. Does anyone seriously need more desktop power? No. Today we just want the batteries in our smartphones to last longer.

Multi-GHz, Multi-Core SoC, thin as paper and barely lukewarm: this is the real modern days' miracle.

I sit at this desk and use this particular computer several hours a day. I built it in 2008 at a cost of almost $1000, but several parts have been downgraded since. It still does almost everything I need to do except run the more demanding games. I built a new computer for my lab bench almost 2 years ago. It is better than this machine in most ways, but cost less than $200 and uses less than 15 watts of power.

[james_s]

When I built my i7 desktop a couple years ago it was replacing a nearly 10 year old Pentium4 that still did about 90% of what I wanted to do with it. When I started playing with computers, one that was a year or two old was hopelessly obsolete, 10 years old would have been positively ancient. At this point I plan to keep using my i7 indefinitely, it's possible that something new and revolutionary will come out which will spur me to upgrade but lately it seems like software has peaked and has started to regress toward the lowest common denominator mobile platform. There is nothing compelling to upgrade to.

[David Hess]

The systems I regularly use are old enough that they are falling behind the supported software curve because Windows has deprecated so much hardware support.

I would accept higher power for higher performance if that was an option but Intel has been choosing smaller die area for lower cost.  This shows up in behavior like when you use the wider width vector extensions, the CPU clock speed is slowed down.  In some cases the loss of clock speed makes total performance lower than if the vector extension was not used.

Intel likes to pick benchmarks which take advantage of new processor features like instruction set extensions but I am interested in applications which do not take advantage of them.  Often this is because of Intel's market segmentation policies which discourage adoption of advanced features.

[senso]

There is also the funny you need a kabylake+Win10+Edge to watch 4k Netflix...
For the 1.8Ghz thing, my laptop with a 4720 hits 3.8Ghz single threaded all day, and a depressing 3.6Ghz with all 4 cores pegged, using a bugged micro-code I could OC it, but there is no point in that..
The bottleneck is enough RAM and slow HDD's for years, a core2quad with an SSD and some 50$ GPU is more than enough for internet, and even basic CAD work, anything office related, etc.

[james_s]

One thing I've never quite figured out is YouTube. I used to be able to stream fullscreen with no issues on a 1.4GHz laptop from ~2004. Then it got so slow eventually that it was unusable so I upgraded to a newer laptop which again streamed full screen just fine, then that became unbearably slow. Now with a core i7 laptop I can again stream full screen but I've noticed CPU usage has increased lately to the point that the fan revs up when it never used to. All that and the video doesn't look any better than it did back on the old 1.4GHz laptop. It's just a steady and dramatic increase in CPU usage for no perceptible improvement.

[Carl_Smith]

There is nothing compelling to upgrade to.

The one compelling upgrade everyone should consider is a solid state drive for your operating system.  The difference is amazing.  Programs load nearly instantly.  You'll be cursing your BIOS for taking 5 seconds to do it's thing because the OS only takes a few more seconds to load.  And the drives from good companies are very reliable now.

For most people - people who are not rendering video or worrying about frame rates in games - the apparent speed of your system is how quickly it reacts to your inputs, a lot of which are loading programs or files, and an SSD makes an amazing difference.

[Someone]

One thing I've never quite figured out is YouTube. I used to be able to stream fullscreen with no issues on a 1.4GHz laptop from ~2004. Then it got so slow eventually that it was unusable so I upgraded to a newer laptop which again streamed full screen just fine, then that became unbearably slow. Now with a core i7 laptop I can again stream full screen but I've noticed CPU usage has increased lately to the point that the fan revs up when it never used to. All that and the video doesn't look any better than it did back on the old 1.4GHz laptop. It's just a steady and dramatic increase in CPU usage for no perceptible improvement.

Its the partly the underlying codecs in use:
https://blogs.gnome.org/rbultje/2015/09/28/vp9-encodingdecoding-performance-vs-hevch-264/
But also that the video quality keeps improving with higher and higher delivered bit rates, the minimum bitrate stream available goes up year on year. You could play full screen video stream on an intel MMX processor, but it was at a much lower quality than you would see today.

[Axl]

For what it's worth:

Zilog ez80f91:
50 MHz
[seems like 4-6 stage] pipeline
Many instructions are single-cycle execution
Zero waitstate mem. access, using SDRAM
24-bit (data and address) bus
On-die EMAC coprocessor (100 Mb/s full duplex)
Etc.