Why arent CPU scaling with die shrinks?

[ComputerGuy2006]

I got an i7 970, its 45nm, 4 cores @ 2.8Ghz for ~$400. I have wanted to upgrade for a while, but it seems each generation Intel disappoints, forcing me to wait another year/generation until I can upgrade...

Broadwell and skylake are both 14nm (3.2 times smaller then 45nm). So in a *PERFECT* world with *PERFECT* scaling, we would have a 12 core CPU at 3GHZ for about $400....

Since we don't live in this perfect world, why are we so far behind? Why is an 8 core CPU (5960X) at 3GHZ cost over $1000?

If you compare the perfect world scenario as 100% scaling, to what we have, we today are at about 26% of a potential 100% scaling (when you factor core count and price).

What is going on? Why is Intel so far behind its potential hypothetical maximum numbers? We had quad cores at 65nm in 2006, yet almost a decade later with over 4x transistor shrink we are looking at over $1000 for an 8 core CPU?

 

[zir_blazer]

You're missing that on the Desktop space, the integrated GPU is around half of the actual die size, and there was a rather big performance jump in the last four generations. On the Server space, where CPUs still have nearly all the die for themselves, the scaling is around the levels you're talking. You have examples like the 18 Core Haswell-E.
Now, the shrinks and scaling will not drive prices down. It just means that they get more profit margins. Its the competence that drives prices down, and these days AMD can't be called that.

 

[Justinbaileyman]

That is the million dollar question isnt it.. I think it comes down to greed myself.I think both Intel and AMD are so busy pocketing the cash and price gouging us that instead of putting cash back into R&D we get stuck with minor upgrades.I mean this used to be about the love of the game not about making billions and trillions of buckaroo's.that is what I am guessing anyways. I mean do you really think it costs $1000 bucks to make a x5960 or $2000 for a E5-2670 v3?? Come on not even close.They need to stop catering to the super rich and get back to the old days of everyone can afford a kick ass pc at a reasonable price.then when the money starts rolling in then put huge bundles of cash into R&D so we can main the moore's law.

 

[LTC8K6]

A brand new Haswell-E i7-5820K 6 core is $380.00

Seems reasonable.

 

[ShintaiDK]

That is the million dollar question isnt it.. I think it comes down to greed myself.I think both Intel and AMD are so busy pocketing the cash and price gouging us that instead of putting cash back into R&D we get stuck with minor upgrades.I mean this used to be about the love of the game not about making billions and trillions of buckaroo's.that is what I am guessing anyways. I mean do you really think it costs $1000 bucks to make a x5960 or $2000 for a E5-2670 v3?? Come on not even close.They need to stop catering to the super rich and get back to the old days of everyone can afford a kick ass pc at a reasonable price.then when the money starts rolling in then put huge bundles of cash into R&D so we can main the moore's law.

You couldnt be more wrong.

 

[Hulk]

I got an i7 970, its 45nm, 4 cores @ 2.8Ghz for ~$400. I have wanted to upgrade for a while, but it seems each generation Intel disappoints, forcing me to wait another year/generation until I can upgrade...

Broadwell and skylake are both 14nm (3.2 times smaller then 45nm). So in a *PERFECT* world with *PERFECT* scaling, we would have a 12 core CPU at 3GHZ for about $400....

Since we don't live in this perfect world, why are we so far behind? Why is an 8 core CPU (5960X) at 3GHZ cost over $1000?

If you compare the perfect world scenario as 100% scaling, to what we have, we today are at about 26% of a potential 100% scaling (when you factor core count and price).

What is going on? Why is Intel so far behind its potential hypothetical maximum numbers? We had quad cores at 65nm in 2006, yet almost a decade later with over 4x transistor shrink we are looking at over $1000 for an 8 core CPU?

Think about what you are saying. You want Intel to take your 970 and shrink it to 14nm, make no architectural changes and plop three of them on one die. Even in this case there would be a lot of development money spent on the process development and changes to make the thing work.

Besides simply shrinking features, which as I noted above is very, very expensive, Intel also spends a boatload on new architecture so that the new part has new instructions, better IPC, and lower thermals and other characteristics specially designed for the new process technology.

Also iGPU die space has been growing with every new processor.

 

[LTC8K6]

$400 in 2008 is about $450 today

The 5820K will crush the 970, for a lot less money.

 

[stockwiz]

That is the million dollar question isnt it.. I think it comes down to greed myself.I think both Intel and AMD are so busy pocketing the cash and price gouging us that instead of putting cash back into R&D we get stuck with minor upgrades.I mean this used to be about the love of the game not about making billions and trillions of buckaroo's.that is what I am guessing anyways. I mean do you really think it costs $1000 bucks to make a x5960 or $2000 for a E5-2670 v3?? Come on not even close.They need to stop catering to the super rich and get back to the old days of everyone can afford a kick ass pc at a reasonable price.then when the money starts rolling in then put huge bundles of cash into R&D so we can main the moore's law.

That would be nice but it's really not up to us to decide how much profit is too much for a company. Nvidia does the same thing. So does Apple. So does the place I work at. Big profit margin on some of our products....

 

[ShintaiDK]

I got an i7 970, its 45nm, 4 cores @ 2.8Ghz for ~$400. I have wanted to upgrade for a while, but it seems each generation Intel disappoints, forcing me to wait another year/generation until I can upgrade...

Broadwell and skylake are both 14nm (3.2 times smaller then 45nm). So in a *PERFECT* world with *PERFECT* scaling, we would have a 12 core CPU at 3GHZ for about $400....

Since we don't live in this perfect world, why are we so far behind? Why is an 8 core CPU (5960X) at 3GHZ cost over $1000?

If you compare the perfect world scenario as 100% scaling, to what we have, we today are at about 26% of a potential 100% scaling (when you factor core count and price).

What is going on? Why is Intel so far behind its potential hypothetical maximum numbers? We had quad cores at 65nm in 2006, yet almost a decade later with over 4x transistor shrink we are looking at over $1000 for an 8 core CPU?

There are several issues.
With more cores you need more interconenct. You also need more memory bandwidth. Then there is the heat and power consumption. Take a look at the 18 core Haswell. It runs at what, 2.3Ghz base? Even the 8 core 5960X only got a baseclock of 3Ghz.

Then there is the integration part. Your i7 920 example didnt have PCIe ondie, it didnt have any IGP and so on.

Lastly and perhaps most important is software. Serial code is serial code and the more there is, the less benefit more cores have.

Not to mention you also wanted higher core performance along the way. Because 2.8Ghz Nehalem cores just doesnt cut it anymore.

 

[SteveGrabowski]

A few months ago I built a gaming system with a Xeon E3-1231v3 and a GTX 970 for less than $1000. I mean that's pretty reasonable to me, considering my gaming systems before have cost a lot more for similar performance tier for the time components. So I don't agree with this greed complaint. I don't ever remember it being cheaper to build systems than it is now.

 

[Fjodor2001]

Of course performance has improved the last few years too. But that's not the issue. The issue is that performance no longer improves at the pace it used to.

As the OP states: "Why arent CPU scaling with die shrinks?"

 

[ComputerGuy2006]

You're missing that on the Desktop space, the integrated GPU is around half of the actual die size, and there was a rather big performance jump in the last four generations. On the Server space, where CPUs still have nearly all the die for themselves, the scaling is around the levels you're talking. You have examples like the 18 Core Haswell-E.
Now, the shrinks and scaling will not drive prices down. It just means that they get more profit margins. Its the competence that drives prices down, and these days AMD can't be called that.

That has only scaled in core count, but not in price.

A brand new Haswell-E i7-5820K 6 core is $380.00

Seems reasonable.

I do not like the idea of buying old tech, the 'high end' is too far behind for me to pay any attention to it anymore. We had a tick at 14nm, and soon we will have a tock. The idea of getting a bit better IPC and only 2 extra cores is not worth buying an entirely new mobo/ram/cpu....

I think the latest rumors are the broadwell highend will be out in 2016, that is after a skylake launch...

Ideally id like to drop a good $2000 or so and not think about a PC upgrade for another 3-4 years (other then GPU upgrades). I think that is part of my problem, the mainstream intel CPU are always so disappointing, while the highend are always so behind. I never feel like there is a good time to upgrade.

Think about what you are saying. You want Intel to take your 970 and shrink it to 14nm, make no architectural changes and plop three of them on one die. Even in this case there would be a lot of development money spent on the process development and changes to make the thing work.

Besides simply shrinking features, which as I noted above is very, very expensive, Intel also spends a boatload on new architecture so that the new part has new instructions, better IPC, and lower thermals and other characteristics specially designed for the new process technology.

Also iGPU die space has been growing with every new processor.

I have a 930, not 970. I do have a basic understanding of CPU archictuture, but your post more or less seems to be implying the reason why things are so expensive is so Intel can spend all that money on RND... I don't buy that.

 

[Yuriman]

The way I see it, the best time to upgrade is when you need more performance...

 

[ComputerGuy2006]

The way I see it, the best time to upgrade is when you need more performance...

Need is subjective tho. In a way ive always *needed* more power. More ram for more VM images, more CPU power for faster encoding (also helps run many image at once), more GPU power for more FPS in games, more speed on my SSD, etc.

So for me upgrading has always been an ROI thing, and in the past few years Intel has never given me an chance to buy something new using this strategy. I feel I either need an obvious large IPC gain (like over 2x, since id be spending over $1k to replace a pc that works) or same IPC but more cores.

I kind of agree with Justinbaileyman. I think this is greed my Intel. They are slowing things down so they can milk the customer more and return more profits to their share holders. Without competition, what do they have to gain by releasing a very fast mainstream CPU? Absolutely nothing if people are willing to buy a slower CPU at the same prize as a hypothetical faster one...

 

[ShintaiDK]

I kind of agree with Justinbaileyman. I think this is greed my Intel. They are slowing things down so they can milk the customer more and return more profits to their share holders. Without competition, what do they have to gain by releasing a very fast mainstream CPU? Absolutely nothing if people are willing to buy a slower CPU at the same prize as a hypothetical faster one...

What do you base this on?

 

[ComputerGuy2006]

What do you base this on?

1# Intel has record profits, had over 11b profit last year.

2# They make it a point not to release a hex/octo core CPU for the mainstream segment.

3# How many generations have we gone now with the highest end CPU in the mainstream segment still using IGP? Ive read some rumors skylake might have a CPU without IGP, but who knows. Id love to know the statistic of people who use the IGP on a +$300 CPU....

4# And of course my OP. The progress just seems way to slow and expensive... Things have slowed down too much. At this rate it wont be until ~2022 at ~3nm that we see a 16 core at +$1000... And that is assuming the current pace continue and does not get even slower.

 

[Justinbaileyman]

What do you base this on?

I base my fact on its been 10-12 years now and we are still at 4c/8t count for the modern user. it was a add 2 core count every 2-4 years starting in 2000 first with dual core to quad core then to hexacore and now all the sudden we are backing down and taking a step back to quad cores.we should be at a hexacore count already for standard user base.why is this thinking wrong?its my money and the consumers decide were the quality and price point should be at not the companies.if that was the case it would be $100 bucks for a cup of milk.

 

[frozentundra123456]

Another one of these simplistic threads. No trend continues forever. Pretty much any technology improves rapidly in the beginning, and the improvements slow down as the "low hanging fruit" is picked, and ultimately physical limits come into play. Yes, you could add more cores, but that comes at a price of higher TDP and lower clockspeed to compensate. Personally, even if the price were the same, I would choose a 4ghz quad core over a 3ghz hex or octo core, simply because I have little use for the extra cores.

OP you are looking at only the last few years of improvements after the mature stage has been reached, and cpu progress has slowed down. Igp performance OTOH, has improved tremendously over that same period, partly because it is in the new, easy to improve phase, and also because it is a lot easier to just throw transistors at a parallel task like gpu than it is to improve cpu performance, which does not scale well for most tasks with more than 2 to 4 cores.

Personally, my first computer was a 233 mhz single core pentium that cost 2000.00. So I think computers today are a *tremendous* value. And with all the competition they are getting from ARM, I think it is absurd to think intel is deliberately withholding performance gains. Could they sell their cpus for less? Of course, but so could any other profitable company. That is why they are profitable.

 

[SPBHM]

if they give enthusiasts what they want, like a higher clocked Xeon D, people would be missing the awesome laptop IGP which is using the die space inside the enthusiast 84-95W TDP CPUs, Intel knows what is best for you, not moar coars, moar crappy IGPs!

 

[Arachnotronic]

Intel spent nearly $12B on R&D last year. If you think Intel isn't investing in future chips, then please tell me what it's spending that $12B on.

 

[frozentundra123456]

I base my fact on its been 10-12 years now and we are still at 4c/8t count for the modern user. it was a add 2 core count every 2-4 years starting in 2000 first with dual core to quad core then to hexacore and now all the sudden we are backing down and taking a step back to quad cores.we should be at a hexacore count already for standard user base.why is this thinking wrong?its my money and the consumers decide were the quality and price point should be at not the companies.if that was the case it would be $100 bucks for a cup of milk.

Your thinking is incorrect because the "standard user base", has no need for a hex core because they dont use software that needs it. In fact I would estimate that over 90 percent of users wont fully utilize a haswell or skylake quad. So why should Intel lose money by selling a bigger die for the same price or lose customers by selling a bigger die and charging more for it. And if you really want moar cores, you can always go AMD. That strategy has worked out well for them. In fact, I would argue that a big part of AMDs problem is that they misjudged the market and did not build what the vast majority really wants, a relatively small number of fast, efficient cores.

And it is definitely your money. Last time I checked, nobody was forcing you to buy a cpu, and anyone who thinks it is so easy and improvements should continue indefinitely is perfectly free to start their own company and try it out.

 

[CHADBOGA]

Terrible thread with a few posters who seem totally detached from reality, waxing on about their dreams. D:

 

[LTC8K6]

I think a hex core unlocked haswell-e at $380 has to be for the mainstream segment?

It's not bottom end, and it's not for the enthusiasts.

 

[ShintaiDK]

1# Intel has record profits, had over 11b profit last year.

2# They make it a point not to release a hex/octo core CPU for the mainstream segment.

3# How many generations have we gone now with the highest end CPU in the mainstream segment still using IGP? Ive read some rumors skylake might have a CPU without IGP, but who knows. Id love to know the statistic of people who use the IGP on a +$300 CPU....

4# And of course my OP. The progress just seems way to slow and expensive... Things have slowed down too much. At this rate it wont be until ~2022 at ~3nm that we see a 16 core at +$1000... And that is assuming the current pace continue and does not get even slower.

1. That part comes from the datacenter.

2. Because its not what the 99% crowd wants.

3. Mainstream is mainstream. Just buy the enthusiast if you want one without IGP. Chips like the 5820K isnt exactly pricy and you get 6C/12T and quadchannel.

4. Try graph the process with performance/watt. Again, what the 99% crowd wants.

 

[Enigmoid]

Intel has been delivering the consistent (but low) IPC gains each gen since conroe.

They have not increased core count on mainstream since Core 2 quad.

With 14nm its time for a hex core option on the mainstream platform. Two broadwell cores + L3 (and I assume a very similar size for Skylake) are ~20-24 mm^2. With the GT2 Skylake die being <120 mm^2 (assuming they do not go crazy with IGP relative to broadwell) its easy enough to add a mainstream hex on the desktop. 14 nm is already low power enough that hex cores could scale to 45W laptop SKUs as well (with aggressive turbo).

Skylake may be +15% over haswell but is disappointing its still 4 cores.