In what Intel CPU generation will 8 cores be introduced?

[Fjodor2001]

Yes and notice how the highest TDP 8C Xeon E5-2xxx part is only 95 watts, and the fastest 8c is only 2.3GHz.

Notice how the 2011 socket parts have higher TDP and faster clocks...

At 1 volt a 150 watt part is going to draw 150 amps! That is a lot of current, like about as much as your whole house uses. How much current do you think each of those tiny wires on a CPU can safely handle?

Have you ever seen the pinout on a modern CPU? A whole bunch of pins are devoted to POWER.

So now you're saying that the additional pins on Socket 2011 are due to higher TDP requirements instead? I.e. approximately 2011-1155=856 pins are devoted to power? Really?

And how come the Pentium D managed on Socket 775, while still having 130W TDP (Presler XE)? I know the power management was not the best on that one (understatement ), but still...

Are you interested in a 2.3GHz 8c machine? I'm not! I'll take a 2600k over that any day.

The problem with that reasoning is that you are stuck in present technology. What I wonder is in what future mainstream Intel CPU generation 8 cores will be introduced. Within 4-5 years, there will have been 2 node shrinks, so the TDP per core will be much lower. And the amount of transistors per die area will also be much higher. What should they do with all those transistors?

You seem to think that we will be stuck with 4C in the mainstream Intel CPUs forever? Do you ever think we'll see 8C mainstream Intel CPUs? Or has some "CPU God" decided that 4C is the holy core count that thou shall never exceed in mainstream CPUs...

 

[lakedude]

So now you're saying that the additional pins on Socket 2011 are due to higher TDP requirements instead? I.e. approximately 2011-1155=856 pins are devoted to power? Really?

Yeah, that isn't really what I said at all. I said power and data...

See:

"It is a challenge to keep 8 cores powered and fed with data so you need more "pins" and a better platform to support all the I/O an 8 core machine can generate."

I've not been able to find the pinouts but it is a safe bet that higher TDP sockets have more power pin than do the lower TDP sockets. I've seen this information somewhere before and was shocked at the huge amount of power pins until I did some simple ohms law calculations and came up with lots of amps needing lots of pins.

If I'm wrong, what is your explanation as to the reason the LGA 1356 parts all have lower TDP and are all slower? I think power and heat have something to do with it.

 

[Fjodor2001]

Ok I'll play.

As the nm of the process shrinks, less power will be needed so that will help as far as needing a ton of power pins.

However, as the core size shrinks removing the heat will be a bigger challenge. I'm pretty sure they are already making parts artificially big to spread out the hot spots. In other words if heat was not an issue they could make cores smaller.

I agree. That's a valid point. But it's also one that has been valid for previous node shrinks, and then they have been able to deal with it anyway. And actually, the latest CPU generations such as SB/IB have been some of the most "overclock friendly" CPUs ever, contradicting the fear of having a heat problem (perhaps some with IB, but it still overclocks well).

I don't see why things should be any different with the coming two node shrinks? Is there anything in particular pointing in that direction?

Speaking of which you need look no farther than Bulldozer to see why at this time in the desktop PC space an 8 core machine is not such a great idea. A FX-8150 is an awesome chip for the money if you can use all 8 cores but most folks would rather have an Intel SB/IB quad.

Well, the 8C Bulldozer is not a "true" 8C. Also, you're putting an 8C Bulldozer against a 4C(/8T) Ivy Bridge, which is not completely fair since they are completely different CPU architectures. The Intel cores simply has a better CPU architecture than AMD these days, with higher IPC etc.

Also, as mentioned in a previous post, most SW is optimized for the current mainstream Intel CPUs which accounts for 80-90% of the market. If 8C would be mainstream don't you think SW would adapt? See e.g. BF3 which already makes good use of 4C. Don't you think a future BF4 will make good use of 8C in 4 years, if widely available?

 

[BTRY B 529th FA BN]

Haven't read the entire thread but Intel introduced MIC? is it mic? but this has got to be the right approach seeing most software is slow to support multi-core processors.

I think the article is on the front page...

 

[ShintaiDK]

Also, as mentioned in a previous post, most SW is optimized for the current mainstream Intel CPUs which accounts for 80-90% of the market. If 8C would be mainstream don't you think SW would adapt? See e.g. BF3 which already makes good use of 4C. Don't you think a future BF4 will make good use of 8C in 4 years, if widely available?

 

[ShintaiDK]

Do you mean what you wrote here? I.e.:

"VRM integration, PCH(SATA, USB, NIC, Audio etc) integration, faster iGPU, possible memory integration and so on.
[...]
The cores themselves also grow in transistor terms."

In that case, you're arguments did not hold. See previous posts here and here which break your arguments.

No you didnt. And you never compared a Conroe core to an IB core in terms of transistor numbers. The cores grow in sizes and the extra diespace is used for caches, iGPU, soon VRM and so on. Something that holds greater value to 99.99% of all the users than adding 4 extra cores.

 

[AtenRa]

An 8 core 3.4GHz 130W TDP at 22nm is possible.

Assuming 22nm have the same (or close to) yields as 32nm process then an 8 core 22nm CPU could be sell at current 6 core prices and make tones of money for Intel. Actually it could make more money because the die size will be smaller.

It will also help to differentiate the 1155 to 2011 socket for enthusiast and power users.

It would also make excellent Server CPUs with more cores and lower power consumption.

 

[fuzzymath10]

You can get a 6-core i7 for around $500-600. That's pretty reasonable. The high end never scales well, and less than 2x a quad core i7 is pretty fair.

In 2002 a mid-high end P4 1.8A was $365 in Canada. That's about $450-500 today. Having fast computing is not a right.

 

[Fjodor2001]

Are you assuming we didn't know about Amdahl's Law? Of course you will not get a 100% performance improvement going from 4C to 8C, but nobody's expecting that either. So what exactly is your point?

 

[ShintaiDK]

Are you assuming we didn't know about Amdahl's Law? Of course you will not get a 100% performance improvement going from 4C to 8C, but nobody's expecting that either. So what exactly is your point?

Diminishing returns. You seem to think there will be plenty of software options, just we add enough cores. Not to mention the extra cost vs other things is easily worth it. Its just not gonna happen and its not how it is. Again, may I remind you of the Mitosis project.

You would be amazed if you for example knew how much serverside software thats essentially singlethreaded, but spawns multiple threads only due to the concurrent users benefit. That benefit aint on the desktop.

AMD and Intel both seem to think max 4 threads will dominate the future...for a really long time.

 

[Fjodor2001]

No you didnt. And you never compared a Conroe core to an IB core in terms of transistor numbers. The cores grow in sizes

I wrote:

"Conroe was 3 node shrinks ago (we have transitioned from 65->45->32->22 nm). I.e. Ivy Bridge can fit 8 times the amount of transistors on the same die area as Conroe. Are you saying each Ivy Bridge CPU core has 8 times the amount of transistors compared to a Conroe CPU core (or anywhere close to that)?"

To which you never responded. Are you now saying that each IB core occupies 8 times the amount of transistors compared to a Conroe CPU core (or anywhere close to that)?

and the extra diespace is used for caches, iGPU, soon VRM and so on. Something that holds greater value to 99.99% of all the users than adding 4 extra cores.

Yes, it's good that they are adding the VRM, Integrated memory controller, and similar on the CPU die. But it does not occupy that much space at all (which is what we are discussing), as can be seen in the picture here.

Regarding the caches I have said that it natural that they grow when the number of cores increase, but apart from that adding more cache currently does not improve performance that much since it has reached a point of diminishing returns.

So the only main parts left to use the 4 times amount of transistors what we can expect in 4-5 years is for the iGPU and CPU cores. But the iGPU is also starting to reach a point where it is sufficient for most users. And those that require more (mainly for games and video editing), will buy a discrete GFX card anyway.

So are you suggesting that we should have a CPU die where 80% of the area is occupied by the iGPU, when most people only will use 10% of its capacity? And not spend any of the increased number of transistors on increasing the number of CPU cores? Does that make sense?

 

[ShintaiDK]

I am not saying IB cores are 8 times bigger. Im just saying you can avoid to counter in that cores grow as they develop with each uarch. Meaning an IB core aint for example 1/8th of an Conroe core.

A C2D is 143mm2 and a quadcore IB is 160mm2. Or translated if calculate the shrink gains and you remove the iGPU etc. Then an IB core is over 2x the size of a C2D core. Meaning one node is wasted just for the core development. Another node for the addition of 2 cores. And the 3rd for adding iGPU, memory controller, PCIe etc.

In the same time the TDP has gone from 65W to 77W. Tho the platform power is greatly reduced due to the integration.

 

[Fjodor2001]

Diminishing returns. You seem to think there will be plenty of software options, just we add enough cores. Not to mention the extra cost vs other things is easily worth it. Its just not gonna happen and its not how it is. Again, may I remind you of the Mitosis project.

You would be amazed if you for example knew how much serverside software thats essentially singlethreaded, but spawns multiple threads only due to the concurrent users benefit. That benefit aint on the desktop.

Yes, there are problems with getting efficient SW utilization of multiple cores - it's nothing new. But Intel is staring to address that, e.g. with adding TSX in Haswell as was discussed here and here. It will not magically make all SW use multiple cores efficiently, but it is a step in the direction of making it easier to write SW that efficiently use multiple cores. As one of the articles said:

"Generally, Intel’s TSX should be helpful for improving the programmability and scalability for concurrent workloads. Even with a modest number of threads, locks can easily limit the benefits from additional cores. While that is not a problem for 2-4 core processors, it is a much bigger factor going forward."

I.e. the introduction of TSX in Haswell could likely be a sign that Intel is preparing for future mainstream CPUs with more than 4 cores.

AMD and Intel both seem to think max 4 threads will dominate the future...for a really long time.

Where's the proof of that? Have you obtained secret copies of AMD's and Intel's internal CPU design documents for the next 5 years?

And once again, if you are saying they will not increase the number of cores, what should they spend all the increased number of transistors on?

 

[ShintaiDK]

You keep looping. Im out of this.

 

[Fjodor2001]

I am not saying IB cores are 8 times bigger. Im just saying you can avoid to counter in that cores grow as they develop with each uarch. Meaning an IB core aint for example 1/8th of an Conroe core.

A C2D is 143mm2 and a quadcore IB is 160mm2. Or translated if calculate the shrink gains and you remove the iGPU etc. Then an IB core is over 2x the size of a C2D core. Meaning one node is wasted just for the core development. Another node for the addition of 2 cores. And the 3rd for adding iGPU, memory controller, PCIe etc.

In the same time the TDP has gone from 65W to 77W. Tho the platform power is greatly reduced due to the integration.

Ok, so over 3 node shrinks the CPU core has grown 2 times. But during that time the amount of transistors per die area has grown 8 times. So only 25% of the increased transistor count has been used by growing the CPU core size.

Now let's look forward 4-5 years when there after 2 node shrinks will be 400% the amount of transistors available on the same die area. There's not much more stuff such as VRM and Integrated memory controller to add on the die (and it doesn't occupy that much space either, as shown previously). If we assume each CPU core will grow at the same rate in the next 2 node shrinks as it has in the last 3 node shrinks, then 25% * 2 / 3 = 16% of the increased transistor count will be spent on growing the CPU core size. Then you have 400% * (100% - 16%) = 336% increase in transistor count left. How do you suggest that is spent between the iGPU and number of CPU cores? Everything on the iGPU, and none of it on more CPU cores?

 

[Ferzerp]

The die must shrink with each node shrink to maintain the current prices as each fab increases in cost.

However, you aren't really discussing anything with anyone. You already have your view and have no interest in doing anything other than saying "nuh uh" to anyone who doesn't agree with you, and like the other guy, that's the last I have to say.

 

[Fjodor2001]

The die must shrink with each node shrink to maintain the current prices as each fab increases in cost.

Not true. See how the die sizes have varied over time:

I.e. the dies size has been going both up and down over time.

However, you aren't really discussing anything with anyone. You already have your view and have no interest in doing anything other than saying "nuh uh" to anyone who doesn't agree with you, and like the other guy, that's the last I have to say.

The problem with you "and that last other guy [ShintaiDK]" is that all you two have to say is that "there will never be any 8C mainstream Intel CPU". Then you ignore any signs of the contrary, and refuse to provide any proof of your statements. Or you provide "proof", which then turn out to not hold when critically verified, for example as above.

And most important, none of you have still been able to answer the crucial question, which is how you think Intel intends to spend the increased transistor count if it will not be on more CPU cores? ShintaiDK gave it a try, but failed. If you want to be taken seriously, then you have to be able to answer that question. Just saying "there will never be any 8C mainstream Intel CPU" does not give you much credibility...

 

[lakedude]

I don't think some of you are being fair. We kinda sidetracked this tread and this was explained later on.

Fjodor2001 wants us to speculate on when 8c CPUs will be more common, like quads are today. We all understand that anyone with enough money can buy an 8c today.

So here goes:

An 8 core CPU is going to draw roughly 2x the power of a quad regardless of what year it is. This will cause an octo-core to produce 2x the amount of heat. Also most software is not making use of so many cores. An 8 core will also need a better platform for increased I/O. In a given power/heat envelope a quad will always be able to be clocked higher than an 8 core...

So there you have it!

Octo cores will be common when an 8 core will draw as much power as our current quads and when average software can make better use of more cores better than it can higher clocks. Power, heat and software are the obstacles.

When is this gonna happen? I sure don't know!

It is short sighted to assume this will never happen.

 

[AtenRa]

This is SB-E at 32nm, an 8-core die, 2,27B transistors at 435mm2
From this die we have the Core i7 39xx series 6-core CPUs.

And this is SB-E at 22nm, an 8-core die, 2,27B transistors at 220-240mm2

Just for reference, SB die is 216mm2 and Quad core SB-E (Core i7 3820) is at 294mm2.

I dont believe anyone should have any doubt that Intel cant release a 6 core IV-E at $300 and 8-core at $500 and up and make tons of profit both in High End desktop and Server.

This is not a question IF they can do it, this is a question If they want to do it

 

[jpiniero]

Intel's main focus on the mainstream cpus is mobile; and I would be surprised to see more than 4 cores until it's feasible to do there.

 

[intangir]

AMD and Intel both seem to think max 4 threads will dominate the future...for a really long time.

Where's the proof of that? Have you obtained secret copies of AMD's and Intel's internal CPU design documents for the next 5 years?

http://slashdot.org/story/06/07/27/1753255/intel---market-doesnt-need-eight-cores
"Two are enough for now, four will be mainstream in three years and eight is something the desktop market does not need. ... I want everybody to go from a frequency world to a number-of-cores-world. But especially in the client space, we have to be very careful with overloading the market with a number of cores and see what is useful."
--David Perlmutter, Executive Vice President and General Manager of Intel Corporation's Intel Architecture Group (IAG).

That was back in July 2006 when Conroe was just released. Has the situation changed at all since then?

And once again, if you are saying they will not increase the number of cores, what should they spend all the increased number of transistors on?

More AVX and GPU compute units, obviously, based on what everyone likes to hear about.

 

[exar333]

This is SB-E at 32nm, an 8-core die, 2,27B transistors at 435mm2
From this die we have the Core i7 39xx series 6-core CPUs.

And this is SB-E at 22nm, an 8-core die, 2,27B transistors at 220-240mm2

Just for reference, SB die is 216mm2 and Quad core SB-E (Core i7 3820) is at 294mm2.

I dont believe anyone should have any doubt that Intel cant release a 6 core IV-E at $300 and 8-core at $500 and up and make tons of profit both in High End desktop and Server.

This is not a question IF they can do it, this is a question If they want to do it

Great post. :thumbsup:

 

[BenchPress]

http://slashdot.org/story/06/07/27/1753255/intel---market-doesnt-need-eight-cores
"Two are enough for now, four will be mainstream in three years and eight is something the desktop market does not need. ... I want everybody to go from a frequency world to a number-of-cores-world. But especially in the client space, we have to be very careful with overloading the market with a number of cores and see what is useful."
--David Perlmutter, Executive Vice President and General Manager of Intel Corporation's Intel Architecture Group (IAG).

That was back in July 2006 when Conroe was just released. Has the situation changed at all since then?

Yes it has/will. Haswell. More specifically, TSX. There is one and only one reason to introduce hardware transactional memory and lock elision technology: to help scale the software to use more cores. Ergo, the number of cores will be increased soon after Haswell.

 

[Fjodor2001]

http://slashdot.org/story/06/07/27/1753255/intel---market-doesnt-need-eight-cores
"Two are enough for now, four will be mainstream in three years and eight is something the desktop market does not need. ... I want everybody to go from a frequency world to a number-of-cores-world. But especially in the client space, we have to be very careful with overloading the market with a number of cores and see what is useful."
--David Perlmutter, Executive Vice President and General Manager of Intel Corporation's Intel Architecture Group (IAG).

That was back in July 2006 when Conroe was just released. Has the situation changed at all since then?

That was 6 years ago. Since we are discussing future Intel CPU generations, looking forward ~4 years the statement will be 10 years old. Things change and evolve. Once upon a time someone said that nobody would ever need more than 640 kB RAM, remember.

So no, I don't think it's very valid for Intel's current CPU plans. See the comment by BenchPress which also points to Intel preparing for release of mainstream CPUs with more than 4 cores not far from now.

More AVX and GPU compute units, obviously, based on what everyone likes to hear about.

AVX and similar does not require that many transistors compared to how much the "transistor per die area" count has grown. See this post, which shows that the number of transistors per core has grown by 200% in the last 6 years, at the same time as the transistor count per die area has grown by 800%!

The GPU can of course be expanded though, and it can require a lot of transistors. But if all the increased transistor count over the next 4-5 years is spent on the GPU, we will be having processor chips consisting of roughly 80% GPU cores and 20% CPU cores. Is that realistic to expect?

 

[Fjodor2001]

I don't think some of you are being fair. We kinda sidetracked this tread and this was explained later on.

Fjodor2001 wants us to speculate on when 8c CPUs will be more common, like quads are today. We all understand that anyone with enough money can buy an 8c today.

So here goes:

An 8 core CPU is going to draw roughly 2x the power of a quad regardless of what year it is. This will cause an octo-core to produce 2x the amount of heat. Also most software is not making use of so many cores. An 8 core will also need a better platform for increased I/O. In a given power/heat envelope a quad will always be able to be clocked higher than an 8 core...

So there you have it!

Octo cores will be common when an 8 core will draw as much power as our current quads and when average software can make better use of more cores better than it can higher clocks. Power, heat and software are the obstacles.

When is this gonna happen? I sure don't know!

It is short sighted to assume this will never happen.

Thanks for bringing this thread back on track again!