When will we have CPUs that are 10x as fast as the current ones?

[Idontcare]

5690x at 4.6ghz.


Replying to posts on the internet. In 3D. With power gloves and cool visors. And generating important 'what-if' scenario cat videos.


The 'ME' demographic and my needs. It's a mobile desktop server morphic format that expands to meet the whims of Me within my holographic field of computation while costing less than an iDevice inside a variable whim-driven power envelope at zero cost with the on-chip fusion power generation block.

It would respond to the server morphic mode by generating indeterminate kilowatts of free power, yet at idle in mobile-morphic mode would use no power at all. In fact, while sitting idle in my Jetson Jeans pocket it would ensnare unsuspecting quarks and other high energy particles and store them in the onboard non-volatile 10 Shilentnobyte holographic cache for later analysis and flashy cat video props.

Basically, a simple, inexhaustible, fully overclockable computing platform for simple needs like mine.



:biggrin:

Ah, thank you for filling in the blanks so splendidly well! As I now understand the question, the answer itself is quite simple to provide. Its 42.

 

[Fjodor2001]

Step 1: Define "current ones"

Step 2: Define the "as fast as" conditions - what software, what workload, etc. 10x faster doing what?

Step 3: Define the imagined demographic that will be purchasing this 10x processor - is it a mobile product, a desktop, a server, what pricepoint, what formfactor, what power consumption, etc.

Fair enough. I'll be a bit more precise:

1. Top end mainstream CPUs. E.g. 4790K and it's successors.

2. General CPU benchmarks (e.g. Cinebench), not specialized ones. Both ST and MT performance would be interesting to measure and comment on, because I guess the answer to this question will differ much depending on that.

3. Desktop user (office work, gaming, video editing, etc). What specific SW they'll use is hard to know at this point, because that may have changed by then. Around $200-400 price point or so, at current market value. Desktop form factor. TDP range 45-135 W, to be flexible.

Hoping for a bit more specific and elaborative answer than "42" this time.

 

[Idontcare]

Fair enough. I'll be a bit more precise:

1. Top end mainstream CPUs. E.g. 4790K and it's successors.

2. General CPU benchmarks (e.g. Cinebench), not specialized ones. Both ST and MT performance would be interesting to measure and comment on, because I guess the answer to this question will differ much depending on that.

3. Desktop user (office work, gaming, video editing, etc). What specific SW they'll use is hard to know at this point, because that may have changed by then. Around $200-400 price point or so, at current market value. Desktop form factor. TDP range 45-135 W, to be flexible.

Hoping for a bit more specific and elaborative answer than "42" this time.

I see. Yeah the tricky part is the price. It isn't terribly difficult to make things that much better, unless you want it 10x better and cost the same price.

Once a business has that project requirement in place then you are looking at the development timeline itself becoming dominated by the pace of developing cost-reduction approaches and not about developing performance-enabling approaches.

Look at speculative processing, for example. A completely straightforward way to enable higher performance, like what you are looking for, but does so only by commensurately increasing power consumption and production cost.

But, as you've framed the question, basically looking for a timeline in which the high-end mainstream desktop SKU will deliver 10x performance at the same SKU ASP and power consumption, yeah I think it is going to take at least 15 years for any company to figure out how to do that without making it cost $3k or use 400W.

 

[Cerb]

Assuming single-threaded performance: when random access to DRAM takes 5ns.

 

[Idontcare]

Assuming single-threaded performance: when random access to DRAM takes 5ns.

Traces are simply too long for that to become a reality, even with photonic interconnects. Someday it will be possible via quantum entanglement, but not in the next 10yrs.

But the current trend is for companies to make the xtors smaller and smaller only to then use billions of them to create more and more levels of sram cache. So come 5nm or 3nm they'll probably have a few GB of sram cache on-die anyways, with sub 1ns access times, so the net affect will be what you are looking for I suspect.

 

[ClockHound]

...basically looking for a timeline in which the high-end mainstream desktop SKU will deliver 10x performance at the same SKU ASP and power consumption, yeah I think it is going to take at least 15 years for any company to figure out how to do that without making it cost $3k or use 400W.

Wait! Only 400W? 10x performance gain with only 4x the power! At the current cost? I'd even settle for 5x the performance for 2x the power. Can we start the petition now?

 

[Borealis7]

As we saw in the "Greatest Necro Thread of All Time(c)":
http://forums.anandtech.com/showthread.php?t=1730073

no one can tell for sure, and honestly we shouldn't even expect it - the reasons being more financial than technological.

what i think is going to happen in the not so distant future, is that we would be leasing our CPU time from server farms even in the domestic market. we wouldn't need to have a power hungry, hot & loud gaming machine at home, we would just have a monitor and a small relay box that goes to an "off-home" computing center which renders the frames for us. obviously, the input lag problem would first have to be solved in order to do that, but i leave that to the scientists.

 

[Idontcare]

Wait! Only 400W? 10x performance gain with only 4x the power! At the current cost? I'd even settle for 5x the performance for 2x the power. Can we start the petition now?

Ha ha

If you wanted it TODAY, then yes, I think a 10x processor could be developed with a targeted ASP of $3k and 400W power consumption envelope...but the pricepoint itself depends on there being enough people who actually will buy the thing.

If too few people are willing to buy it, then the price needs to go up such that the TAM at least equals the development cost.

But this is my primary concern. Who wants a processor that provides 10x the performance of a 4790? Not many, in fact not many people even want to pay for the performance of a 4790 itself, let alone one with 10x the performance.

 

[ClockHound]

Ha ha

Who wants a processor that provides 10x the performance of a 4790?

:wub: I do!!! :wub:

And admit it, you do too.

 

[Lonyo]

Good question. I imagine mainstream TDP range, around 45-135 W or so to be flexible.

That doesn't work.
We spent years increasing power to get more performance, peaking at 150w for mainstream processors.

Now mainstream is 84w and includes a GPU and various other non-CPU functions.
Intel could throw 1.5x as much power at their "mainstream" CPUs and make them 135w CPUs and bam, you have an instant increase in performance. Like AMD did with their stupid overclocked 200w FX processor.

Or you say OK, 84w is mainstream and has X performance, but you can get a Xeon Broadwell with lots of cores, and has 3x+ of the performance in a 145w envelope, instant "performance" increase.

54w is a mainstream dual core desktop. 145w is an 18 core Xeon.

 

[cytg111]

HWho wants a processor that provides 10x the performance of a 4790?

- what ?

 

[Idontcare]

:wub: I do!!! :wub:

And admit it, you do too.

Of course I do!

In fact I'd buy 6 of them to replace my current hex-quad-core crunching farm if given the opportunity, but between you, me, and maybe (absolute tops) 10k other people on planet Earth, the ASP for such a SKU would probably need to be north of $50k just for Intel to recover the amortization of the invested R&D to develop/debug/verify/validate such a processor.

And as a consumer my feeling here is that this is exactly the sort of business opportunity that AMD would throw themselves at, balls-to-the-wall, thinking it would be a door opener to unlocking future TAM. But Intel would look at the same opportunity and scoff, as Otellini so famously did towards Jobs and his pitiful iPhone dream, and conclude (rightfully so, mind you, in my opinion) that such a SKU would be a total flop, a derpdozeresque flop, and would not risk a single R&D dollar pursuing such a project.

At least not on a timeline that was more aggressive than 10yrs. 15 years though, probably economically reasonable. But I'd buy it anytime before then. Even now I keep eyeing those 18core V3's, they are nearly exactly what I'm interested in. Just not quite the clockspeed I want or need.

 

[Fjodor2001]

At least not on a timeline that was more aggressive than 10yrs. 15 years though, probably economically reasonable.

That would mean around 16% yearly performance increase, or roughly 20-24% per CPU generation. (1.16^15=9.26)

Is that realistic given the trend of performance increase that we've seen over the last few mainstream desktop CPU generations? It's more like ~8% per CPU generation these days.

So what technology do you see on the horizon that would bring back that level of yearly CPU performance increase? :hmm:

 

[Dave2150]

I'd say about 15 years or so. 30 core desktops @ 5ghz , with 3x the IPC will be normal by then.

The way things are going (Intel with no competition) we'll be very very lucky to have a hex core mainstream CPU from Intel in 15 years.

They want to keep the mainstream user on quadcores for as long as possible - as they make much more profit from spitting out the tiny 4core dies.

 

[ShintaiDK]

The way things are going (Intel with no competition) we'll be very very lucky to have a hex core mainstream CPU from Intel in 15 years.

They want to keep the mainstream user on quadcores for as long as possible - as they make much more profit from spitting out the tiny 4core dies.

Maybe you should make a chart of the die size of a quadcore over the years. Because its not what you expect it to be.

Just because Intel doesnt give you what you want doesnt mean they dont give the 99% crowd what they want.

And nobody is preventing you from buy a cheap 5820K or similar.

 

[Maxima1]

Maybe you should make a chart of the die size of a quadcore over the years. Because its not what you expect it to be.

Just because Intel doesnt give you what you want doesnt mean they dont give the 99% crowd what they want.

And nobody is preventing you from buy a cheap 5820K or similar.

That's really easy to meet, which is one reason why Intel has little power over the typical customer. But being easy to meet and having a subset of customers with little knowledge, AMD and Intel can sell castrated processors and the customers are still none the wiser.

 

[videogames101]

Barring a breakthrough on the tech side (which I don't rule out!), CPU architecture isn't going to get much faster. You can only pipeline so much before it costs you performance instead of increasing it, and you can't increase clockspeeds because of thermals.

 

[soresu]

I think that there are several possible areas of improvement in the future that will enable, if not 10x IPC, then certainly 10x performance per watt even without process shrinkage.

I know that it has been heralded since forever, but on chip optical interconnects are definitely in the works, both by IBM and Intel (to say nothing of independent academia). Optical interconnects have the potential to dramatically reduce the heat/power of processors, and then of course there are higher density, lower power, persistant cache memories (such as STT-MRAM, or MeRAM).

I saw a recent post about quantum wells being used for electron cooling, which could reduce the amount of heat in the chip again, although im not sure if reduced electron temperature could adversely affect speed in a complicated micro-architecture.

Recent studies on graphene spintronics suggest it may be a promising candidate for spintronic logic, which is another gateway to highly efficient logic circuitry.

Also, even if graphene cannot be used anytime soon for logic, its potential for heat exchange, and power delivery inside the chip(lower resistance, ballistic current) should be realised much sooner, with all the current advances in graphene production/synthesis.

I would not be surprised to see the processor design sector step back on process geometry to master these techniques, especially together with 3D stacking, as samsung has already done with NAND (40nm).

 

[Ajay]

Along the lines of what soresu posted - I think the asked for 10x increase in performance will require the introduction of disruptive technology/ies that cause an inflection point in the current throughput/year curve. Almost certainly, these technologies will debut in lab demos, then specialized large compute clusters and eventually make their way to desktops. I believe that we will see better than a 10x speed up, but I also think that it's going to take at 10+ years before someone puts it all together and creates a new type of CPU operating in a perfect environment in a lab. The big question will be how long it will take to get said CPU running well under the imperfect world that is the desktop environment and how quickly tool makers are able to design the physical tools required for HVM - IMHO.

 

[Dave2150]

Maybe you should make a chart of the die size of a quadcore over the years. Because its not what you expect it to be.

A quadcore die (4790k example) is much smaller than the die size of a hexcore (5820k for example).

They are much cheaper for intel to make - since a wafer can contain many more smaller dies than it can contain larger ones. Not rocket science, so should be easy for you to understand.

Just because Intel doesnt give you what you want doesnt mean they dont give the 99% crowd what they want.

And nobody is preventing you from buy a cheap 5820K or similar.

The 99% crowd of Intel's customers have no clue what CPU is in their tablet/phone/basic PC. No clue at all.

I could buy a 5820k, sure. So could you. So could anyone. Not sure what your point is exactly - I'm simply pointing out that it's in Intel's financial interest for the mainstream CPU's to be as small a die as possible, instead of the die size of the 5820k hex core.

Mainstream quadcore I7 die size = 177 mm². 5820k hexcore diesize = 356mm2. Hexcore die size is over DOUBLE the size of a quad core die.

Yes, I'm aware that the 5820k hex core dies could be smaller, since they contain more cores that are disabled (from the xeon range), though Intel obviously feel that this is the most profitable way for them to do it.

Long story short - Intel make far more money from limiting the mainstream to a quad core, instead of a hexcore.

 

[ShintaiDK]

A quadcore die (4790k example) is much smaller than the die size of a hexcore (5820k for example).

They are much cheaper for intel to make - since a wafer can contain many more smaller dies than it can contain larger ones. Not rocket science, so should be easy for you to understand.



The 99% crowd of Intel's customers have no clue what CPU is in their tablet/phone/basic PC. No clue at all.

I could buy a 5820k, sure. So could you. So could anyone. Not sure what your point is exactly - I'm simply pointing out that it's in Intel's financial interest for the mainstream CPU's to be as small a die as possible, instead of the die size of the 5820k hex core.

Mainstream quadcore I7 die size = 177 mm². 5820k hexcore diesize = 356mm2. Hexcore die size is over DOUBLE the size of a quad core die.

Yes, I'm aware that the 5820k hex core dies could be smaller, since they contain more cores that are disabled (from the xeon range), though Intel obviously feel that this is the most profitable way for them to do it.

Long story short - Intel make far more money from limiting the mainstream to a quad core, instead of a hexcore.

You are changing the goalpost.

One could also point out that a dualcore GT3 is 181mm2. Or a quadcore GT3 264mm2.

And did you forget the penalty? The 4790K runs 4Ghz base, 4.4Ghz turbo at 88W. How is the 5960X for example again? 8 cores at 3Ghz base, 3.5Ghz turbo? 5820K is 3.3Ghz base, 3.6Ghz turbo. Both at 140W.

It doesnt matter what the diesize is if you cant bin it.

And remember, there are no desktop CPUs. There are mobile and server.

 

[ClockHound]

And remember, there are no desktop CPUs. There are mobile and server.

Aha! Intel's version of BIG.little - Server and Mobile - They just don't have the marketing cajones to put them both in one chip. ;-)

 

[Dave2150]

You are changing the goalpost.

One could also point out that a dualcore GT3 is 181mm2. Or a quadcore GT3 264mm2.

And did you forget the penalty? The 4790K runs 4Ghz base, 4.4Ghz turbo at 88W. How is the 5960X for example again? 8 cores at 3Ghz base, 3.5Ghz turbo? 5820K is 3.3Ghz base, 3.6Ghz turbo. Both at 140W.

It doesnt matter what the diesize is if you cant bin it.

And remember, there are no desktop CPUs. There are mobile and server.

Long story short - Intel make far more money from limiting the mainstream to a quad core, instead of a hexcore.

 

[ShintaiDK]

Long story short - Intel make far more money from limiting the mainstream to a quad core, instead of a hexcore.

Wrong story.

And they are not limiting, they are giving the market exactly what it wants. Thats business 101.

You forget Intel is also getting 1000$ for those 356mm2 of your "hex" in the form of the 5960X. While they also only get 180$ in the form of an i5 44xx with 177mm2. Or they get down to 255$ for a 264mm2 R model. Not to mention 128MB EDRAM extra. And all that is without accounting for binning.

Doing selective compares just because you feel entitled to a mainstream hexcore way below the 400$ pricepoint doesnt make it right in any way.

 

[Dave2150]

You forget Intel is also getting 1000$ for those 356mm2 of your "hex" in the form of the 5960X.

Here's a little tip for you, since you don't seem to understand. Intel sell a few more mainstream quadcore CPU's than they sell 5960x's...........

Again, long story short, Intel make more money from selling small dies than they do selling large dies.

Doing selective compares just because you feel entitled to a mainstream hexcore way below the 400$ pricepoint doesnt make it right in any way.

Now you are fabricating nonsense. I don't feel entitled to anything, why you are questioning my personal purchase choices is bizarre - you're going off at a tangent, as you can't accept my opinion. My opinion is simple - Intel make more profit selling smaller dies than large ones. You disagree, that's fine, lets leave it at that.