Discussion Intel current and future Lakes & Rapids thread

[LTC8K6]

Just because they used it doesn't mean it's any good. Does it replicate CPU's? CPU's are what we cool. If plattens replicate those, they are accurate.

I don't recall even giving an opinion on it...

It was just an example of how it could be done, albeit an old example.

Lots of such heaters are available in a plethora of shapes, sizes, and types, probably including some with variable surface heating.

You could also possibly use Peltier devices.

 

[dullard]

One thing to ALWAYS remember.

Empirical data beats theory 100% of the time. If the real world says something different to your predictions, then your theory/model is wrong.

We can all speculate ad infinitum, on designing a simulator that's close to reality, but nothing is as good as the real thing. That's why we still get surprises when building stuff after millions of hrs in simulations.

This is about as simple of a task as can be: cool a small flat surface.

This isn't something that requires millions of hours in simulations. It isn't something that needs theory or modelling. You are way over complicating it. If the heat sink can't pull 150 W out of a metal block, then it can't pull 150 W out of a CPU. Plain and simple.

ehume can use a simple rig to compare the future of heat sinks. Can the cutting edge heat sink keep the NEXT, unreleased, CPU cool? That would be great information to know before we invest in a heat sink. You don't need a CPU to determine if the heat sink can cool a given load evenly.

 

[LTC8K6]

A couple of interesting reads from Frostytech.

http://www.frostytech.com/testmethod_mk2.cfm

http://www.frostytech.com/articleview.cfm?articleID=1915

 

[maddie]

This is about as simple of a task as can be: cool a small flat surface.

Ever think that your basic assumption is incorrect? You are assuming that the small flat surface is uniform or close to it.

This isn't something that requires millions of hours in simulations. It isn't something that needs theory or modelling. You are way over complicating it.
.

I meant that even simulations totalling millions of hrs can lead to false solutions if the model is wrong. Not saying and never said that this one needs millions of hrs.

If the heat sink can't pull 150 W out of a metal block, then it can't pull 150 W out of a CPU. Plain and simple.

This is where you you might be making an error. No one is saying that it can't pull the same heat out of both the test block and the CPU.

What we are saying is that the temperature needed for the two scenarios are different. The CPU in a lot of, or probably most cases, will have to have hotter regions than the test block average. A CPU has many regions that produce wildly varying amounts of heat. How in the world can you simulate this properly with a very rough scale heater block. You just have to use AVX code to show this.

In simpler terms, if the CPU average temp is X, then there are regions with X+ and X++. There will also be regions of X- and X- -.

A cooler that can pull 200W from a heater plate might cause a CPU of 125W to overheat due to hotter regions in the CPU that the heater plate temps might suggest.

Anyhow, this will be my last post as this thread is getting derailed.

 

[dullard]

Ever think that your basic assumption is incorrect? You are assuming that the small flat surface is uniform or close to it.

No, because computer simulations of heat transfer with heat sinks is a big portion of my career spanning decades.

What we are saying is that the temperature needed for the two scenarios are different. The CPU in a lot of, or probably most cases, will have to have hotter regions than the test block average. A CPU has many regions that produce wildly varying amounts of heat. How in the world can you simulate this properly with a very rough scale heater block. You just have to use AVX code to show this.

In a few minutes I gave you a way to simulate 6 different heat zones in a CPU sized block. With a few weeks, we could build something far more sophisticated.

 

[ehume]

. . .A cooler that can pull 200W from a heater plate might cause a CPU of 125W to overheat due to hotter regions in the CPU that the heater plate temps might suggest. . . .

And this is why simulations are just that. No matter how well simulated, and how many decades of experience one has with simulations, one must test things out with real-world applications. Why else do we have wind tunnels?

 

[dullard]

Why else do we have wind tunnels?

Building a fully functional airplane, putting a pilot inside who might die, just to test a minor design tweak is unethical and excessively expensive. Building a test rig for heat sinks with hot spots is neither unethical nor expensive. Nor is it very difficult to build in a way that accurately reflects CPUs.

I'm just trying to help you out. So that you can simulate expected performance months earlier.

 

[ehume]

Building a fully functional airplane, putting a pilot inside who might die, just to test a minor design tweak is unethical and excessively expensive. Building a test rig for heat sinks with hot spots is neither unethical nor expensive. Nor is it very difficult to build in a way that accurately reflects CPUs.

I'm just trying to help you out. So that you can simulate expected performance months earlier.

Suboptimal designs fly. Suboptimal designs cool. I'm happy that designers make simulations: it is easier to manipulate electrons than it is to bend metal. And platens are fine to test prototypes. But in the end platens are not CPU's; and we do not cool platens, we cool CPU's. This is why I abandoned Frostytech years ago, pay no attention to their results. Give me testing on a CPU. That I can relate to.

 

[Dayman1225]

Next update of HWInfo adds "improved support of Intel Icelake"

Link

 

[mikk]

ICL-U is still targeted for mid 2019, hopefully no further delays for this part.

 

[jpiniero]

ICL-U is still targeted for mid 2019, hopefully no further delays for this part.

Extremely unlikely. Comet Lake is the 2019 U part. Maybe a low volume Y part if they decide to do so.

 

[Dayman1225]

Extremely unlikely. Comet Lake is the 2019 U part. Maybe a low volume Y part if they decide to do so.

Nobody but PCWatch has mention a Comet Lake U, and if it is true it is still possible that it's ICL U (GT3e) 28w. Just have to wait and see.

 

[IntelUser2000]

Nobody but PCWatch has mention a Comet Lake U, and if it is true it is still possible that it's ICL U (GT3e) 28w. Just have to wait and see.

PCWatch has a very good track record.

Besides, according to Ashraf, Comet Lake U is a 28W part and ICL U coming in 2019 is a 15W part.

Though the logic on that seems somewhat flawed. It means Comet Lake is a) a decent change despite being on 14nm, b) Icelake U is artificially crippled c) Icelake is just not that good. You'd think a 15W 10nm part would be competitive with a 28W 14nm part whether in graphics or CPU.

 

[jpiniero]

Besides, according to Ashraf, Comet Lake U is a 28W part and ICL U coming in 2019 is a 15W part.

Dunno about that. It would make sense if Intel prepared 15W versions of both Comet and Ice; but given what we know about 10 nm it just seems so unlikely that Intel will fix it in time.

GT3e seems destined to be discontinued, if it hasn't already... unless Apple continues to want it.

 

[IntelUser2000]

Dunno about that. It would make sense if Intel prepared 15W versions of both Comet and Ice; but given what we know about 10 nm it just seems so unlikely that Intel will fix it in time.

What would be the point of that?

It reads like you are thinking high memory bandwidth parts are only for 28W. That's not true. 15W parts have eDRAM versions as well. That means having 15W Comet and Ice will be redundant.

Currently:
28W-eDRAM only
15W-w eDRAM and w/o eDRAM

(Icelake generation though is ripe for Intel replacing eDRAM with HBM variants)

 

[jpiniero]

You're right, there are 15W GT3e models, but they are so rare it may as well not exist. Hell, GT3e is rare enough as it is outside of Apple and the NUCs.

That means having 15W Comet and Ice will be redundant.

It has to be redundant, because they can't trust 10nm. That the EU counts likely differ will work it out on it's own. Icelake I believe has additional options for disabling EUs for yield reasons so Intel could in theory disable down to whatever Comet Lake's EU count is to keep them equal.

 

[IntelUser2000]

You're right, there are 15W GT3e models, but they are so rare it may as well not exist. Hell, GT3e is rare enough as it is outside of Apple and the NUCs.

You don't think the 28W chips are rare? They are even worse. Only Apple and Vaio uses them. At least the 15W GT3e chips are used as the highest end configuration from few manufacturers beyond Apple.

Let's speculate on another option. They decide eDRAM/HBM is only for Comet Lake. Right now the 48EU/eDRAM configuration gives 50% advantage over the 24EU only configuration at same TDP. Given that Icelake is Gen 11 and is supposed to be a big generational advancement, even without the eDRAM it might be rather competitive with Gen 9 and eDRAM. The lack of eDRAM can be made up by significantly improved compression techniques. Seeing how well AMD does without them, its certainly possible.

Then Comet Lake vs Icelake may become yet another HD 6000 vs 5500 where despite having double the EUs it performed barely above the 5500(0-10%). Better uarch on Icelake and 10nm may make 15W ICL U faster than 28W Comet Lake U.

Artificial crippling starts introducing problems when competition is strong.

 

[jpiniero]

Remember that Icelake (and possibly Comet Lake) supports LPDDR4(X) so in theory you would have 4266 available.

 

[dooon]

kontron Roadmap (source)

 

[yoram]

Some information on Icelake performance counters from Intel

https://dyninst.github.io/scalable_...ssing challenges in Icelake - Ahmad Yasin.pdf

 

[jpiniero]

https://pbs.twimg.com/media/Die8sXmU0AA7ekM.jpg

Another image of the Cannonlake package. And remember, basically 2/3rds of the CPU die is disabled. There's a ton of wasted space.

Does make me wonder if Intel could backport EMIB back to Icelake they could live with the crazy high defect rate because the individual dies are so small.

 

[dooon]

hmm... (source:disqus)

 

[NTMBK]

hmm... (source:disqus)

I like how Cooper Lake is still "in planning", despite the fact it's meant to come out before Icelake. A panic "we need something to ship" product if I ever saw one

 

[Dayman1225]

hmm... (source:disqus)

Better quality one from David Schor, it's real.

EDIT: Slide is located inside this PDF

 

[jpiniero]

I like how Cooper Lake is still "in planning", despite the fact it's meant to come out before Icelake. A panic "we need something to ship" product if I ever saw one

Well, the rumor is that Intel is going to use two dies fused together via EMIB for Cooper Lake and Icelake Server. This is how Cascade Lake-AP works.

So perhaps:

18x2 core Cooper Lake
28x2 core Cooper Lake
38x2 core Icelake Server (bet it's a paper launch though)

Be interesting how Intel spins it, especially given that Rome should be on the market for some time before this is made available. Either way margins will take a giant dump.