Intel Preparing Multiple Hexacore Coffee Lake CPUs

[MageTank]

Just now, Grinners said:

Heat IS inefficiency. 

 

By definition an increase in heat all else the same is a decrease in efficiency. 

 

Heat is wastage. 

This is the key part. Not all else is the same. If you are getting more performance at the cost of heat increasing (as more and more of your CPU is being used), heat becomes a byproduct of that efficiency. Again, using AVX as an example. Load up Linpack, and slowly alter either your AVX offset (if you have a board that allows it) or adjust your tRDWR/tWRRD timings. As the amount of FLOPS you get increase, so does your CPU's temperatures. Lowering the AVX offset, or weakening the timings associated with AVX (timings that alter bandwidth), your FLOPS are lower, and temperatures are lower as a result.

 

I am not trying to deceive anyone when I say this, but it should be common knowledge by now. If you are using MORE of your CPU, it should go without saying that it will run hotter than what it would if you were using less. Efficiency is dependent entirely on the context. Hopefully I've made that clear this time around. 

 

For those of you that do not have Linpack, or are too afraid to run it, you can test this without using FLOPS as a measurement. Simply compare the various "100% load" stress tests. While most of them still load your CPU to 100% usage, their difference in heat can be monumental. 

[Drak3]

Someone should talk about the difference between increasing the efficiencies of instruction sets (which increases heat with performance) and that of current use/leakage (which decreases heat with what it takes to maintain a certain clock).

 

I'd do it, but someone would probably call me obsessed, again.

[Bleedingyamato]

16 minutes ago, Drak3 said:

Someone should talk about the difference between increasing the efficiencies of instruction sets (which increases heat with performance) and that of current use/leakage (which decreases heat with what it takes to maintain a certain clock).

 

I'd do it, but someone would probably call me obsessed, again.

Is it being obsessed in cases where you're right though?  I don't think it's bad to clarify things when you're factually correct.  

[MageTank]

11 minutes ago, Drak3 said:

Someone should talk about the difference between increasing the efficiencies of instruction sets (which increases heat with performance) and that of current use/leakage (which decreases heat with what it takes to maintain a certain clock).

 

I'd do it, but someone would probably call me obsessed, again.

Nah, I'd call you passive aggressive this time around.

 

That being said, I'll elaborate once more about efficiency in this context. What @Grinners said is true, heat is inefficiency in most aspects. However, CPU's are different.  Hotter is more efficiency in the sense that more cycles used, less dead time = more heat. 

 

I'll pose a question, if it will help. Which is more efficient. A 50w CPU that can do 100GFLOP's and run super cool, or a 200w CPU that does 500GFLOP but runs super hot? The hard part, is conveying the context of efficiency, because efficiency can mean many things. In this case, it would be heat vs performance vs time. Just because something uses more power, doesn't automatically make it the less efficient of the two. Just because something is hotter, doesn't make it the less efficient option either. Sometimes, the heat can be considered a byproduct of the efficiency, as long as it has the relative performance to go with it. 

[arnavvr]

Hmm maybe I should list my 7700K on eBay... 5.2Ghz on 6C/12T sounds nice 

[Drak3]

1 minute ago, Bleedingyamato said:

Is it being obsessed in cases where you're right though?  I don't think it's bad to clarify things when you're factually correct.  

Another member thinks I'm obsessed with him. He also missed the purpose of my post.

[LooneyJuice]

3 minutes ago, MageTank said:

Nah, I'd call you passive aggressive this time around.

 

That being said, I'll elaborate once more about efficiency in this context. What @Grinners said is true, heat is inefficiency in most aspects. However, CPU's are different.  Hotter is more efficiency in the sense that more cycles used, less dead time = more heat. 

 

I'll pose a question, if it will help. Which is more efficient. A 50w CPU that can do 100GFLOP's and run super cool, or a 200w CPU that does 500GFLOP but runs super hot? The hard part, is conveying the context of efficiency, because efficiency can mean many things. In this case, it would be heat vs performance vs time. Just because something uses more power, doesn't automatically make it the less efficient of the two. Just because something is hotter, doesn't make it the less efficient option either. Sometimes, the heat can be considered a byproduct of the efficiency, as long as it has the relative performance to go with it. 

So you could, for example, use a procession of generational leaps within a family of processors as a performance index to gauge efficiency against heat output, yes?

[Bleedingyamato]

5 minutes ago, Drak3 said:

Another member thinks I'm obsessed with him. He also missed the purpose of my post.

Obsessed with him?  That's cute.  /s lol

 

 

In any case if you have accurate and helpful info to contribute don't let some jerk stop you.  ?

[MageTank]

3 minutes ago, Drak3 said:

Another member thinks I'm obsessed with him. He also missed the purpose of my post.

No, you are making passive aggressive posts out of anger for me telling you that you were being overly obsessive with me. It's not out of line for one to tell another they need to dial back their compulsive posting, is it not? We are all adults here.

 

1 minute ago, LooneyJuice said:

So you could, for example, use a procession of generational leaps within a family of processors as a performance index to gauge efficiency against heat output, yes?

Yes, it's part of how we extrapolate how well a "new" architecture will perform relative to the previous architecture it's based on. It's also how we are able to guess at potential clock speeds, assuming all other factors are the same (socket pinout, cache structure, instruction sets, etc). With the rumors of Intel re-designing the socket pinout to improve efficiency(don't ask me how), it could potentially throw off our numbers. 

 

1 minute ago, Bleedingyamato said:

Obsessed with him?  That's cute.  /s lol

 

 

In any case if you have accurate and helpful info to contribute don't let some jerk stop you.  ?

I've never once tried to stop him. When he makes posts relative to the topic, I have no issues at all. I do however, take issue with constantly being mentioned in random threads that have nothing to do with me, or people randomly quoting me to point out things that have nothing to do with my posts or the topics. It's creepy, lol. 

[Drak3]

Just now, Bleedingyamato said:

In any case if you have accurate and helpful info to contribute don't let some jerk stop you.  ?

I wouldn't necessarily call him a jerk.

 

Lies, I totes would.

 

But I've got the central idea out, and if someone wants me to elaborate, I will. But as it stands, there's not much reason for me to want to butt in.

[MageTank]

Just now, Drak3 said:

I wouldn't necessarily call him a jerk.

 

Lies, I totes would.

 

But I've got the central idea out, and if someone wants me to elaborate, I will. But as it stands, there's not much reason for me to want to butt in.

You are doing a bad job for your case if you brought it up yourself without anyone else asking you to do so. At this point, it's akin to begging for attention. I have no issue with you making posts, and I certainly do not have the right to tell you what you can or cannot say. I see no need to get over-dramatic for me simply asking you to dial back the random mentions/off topic posting. Kind of like what we are doing now, due to the posts having nothing to do with the topic at hand.

 

I hold no ill-will towards anyone on this forum (and contrary to popular belief, never have). I also do not shy away from any problems people may have with me. If there is something you would like to discuss, send me a PM. Otherwise, let's avoid derailing another thread together. 

[LooneyJuice]

2 minutes ago, MageTank said:

 

Yes, it's part of how we extrapolate how well a "new" architecture will perform relative to the previous architecture it's based on. It's also how we are able to guess at potential clock speeds, assuming all other factors are the same (socket pinout, cache structure, instruction sets, etc). With the rumors of Intel re-designing the socket pinout to improve efficiency(don't ask me how), it could potentially throw off our numbers. 

 

I mean, I know it sounds like a bit of a "duh" question, but it would be a cool actual metric to have to graphically demonstrate efficiency gains over the course of a few generations to someone like in this case. Instead of merely TDP and frequency, which admittedly doesn't help anyone.

[MageTank]

Just now, LooneyJuice said:

I mean, I know it sounds like a bit of a "duh" question, but it would be a cool actual metric to have to graphically demonstrate efficiency gains over the course of a few generations to someone like in this case. Instead of merely TDP and frequency, which admittedly doesn't help anyone.

I probably could do that. I have a 7700k in my rig, a 6700k on my desk, and a 4790k in my brothers PC downstairs. Would take a little time to set it all up, but it's not a bad idea. Heat vs Power vs FLOPS. Sounds like a good time.

[LooneyJuice]

Just now, MageTank said:

I probably could do that. I have a 7700k in my rig, a 6700k on my desk, and a 4790k in my brothers PC downstairs. Would take a little time to set it all up, but it's not a bad idea. Heat vs Power vs FLOPS. Sounds like a good time.

FLOPS per Watt, yeah. I guess it could be as simple as that.

[MageTank]

Just now, LooneyJuice said:

FLOPS per Watt, yeah. I guess it could be as simple as that.

Did you have something else in mind? I am up for testing anything. 

[AlwaysFSX]

I just want 12 fast cores that won't break the bank or turn my room in to a sauna..

 

15 minutes ago, MageTank said:

I've never once tried to stop him .. being .. creepy .. . 

The truth comes out eventually.

[LooneyJuice]

Just now, MageTank said:

Did you have something else in mind? I am up for testing anything. 

Not necessarily, other than completely saturating every chip and measuring its relative performance against heat output. You're way more savvy at that stuff. Additionally, I only really have this oldie Sandy machine, so it's not like I can provide much useful methodology input currently.

 

It would be cool to have some kind of chronology though. And I'm guessing it would have to be done strictly on a stock specification basis, as every chip ran out of the box.

[ARikozuM]

1 minute ago, LooneyJuice said:

It would be cool to have some kind of chronology though. And I'm guessing it would have to be done strictly on a stock specification basis, as every chip ran out of the box.

Totally agreed on this, [unless we can get all of the tested components to the same frequency].

[MageTank]

Just now, AlwaysFSX said:

I just want... a sauna..

You and me both, brother.

 

Just now, LooneyJuice said:

Not necessarily, other than completely saturating every chip and measuring its relative performance against heat output. You're way more savvy at that stuff. Additionally, I only really have this oldie Sandy machine, so it's not like I can provide much useful methodology input currently.

 

It would be cool to have some kind of chronology though. And I'm guessing it would have to be done strictly on a stock specification basis, as every chip ran out of the box.

The plan is to completely utilize the CPU cores, cache subsystem, and memory controller using Linpack MKL, and recording the FLOPS we get. While doing so, I'll also monitor power consumption from the CPU package, along with reported temperatures. First, I would like to delid my brother's 4790k so that I can keep it all fair, as my 6700k and 7700k all have liquid metal on the die. The hard part, will be trying to figure out the impact Haswell's FIVR will have on the results. I guess i'll figure that out once I get the results. 

[LooneyJuice]

Just now, MageTank said:

You and me both, brother.

 

The plan is to completely utilize the CPU cores, cache subsystem, and memory controller using Linpack MKL, and recording the FLOPS we get. While doing so, I'll also monitor power consumption from the CPU package, along with reported temperatures. First, I would like to delid my brother's 4790k so that I can keep it all fair, as my 6700k and 7700k all have liquid metal on the die. The hard part, will be trying to figure out the impact Haswell's FIVR will have on the results. I guess i'll figure that out once I get the results. 

Well, this is a dumb idea, but would it really matter? It's the heat transfer to the IHS that was bottlenecked, but after equilibrium, they still output the same Wattage in heat, yes? Or is this a thermodynamics Faux Pas? Can't rightly say, a bit taxed on sleep.

[MageTank]

Just now, LooneyJuice said:

Well, this is a dumb idea, but would it really matter? It's the heat transfer to the IHS that was bottlenecked, but after equilibrium, they still output the same Wattage in heat, yes? Or is this a thermodynamics Faux Pas? Can't rightly say, a bit taxed on sleep.

I am not an expert by any means, but thermal runaway dictates that the hotter something gets, the more power it uses, the more power it uses, the hotter it gets. An engineer can certainly correct me if I am wrong, but that's my basis behind wanting to delid to keep the variables as low as possible.

 

I can also run them all at 4ghz, if that helps. Or, overclock them all to 4.5 (the 7700k's stock speed) and see if that makes any difference. 

[LooneyJuice]

15 minutes ago, MageTank said:

I am not an expert by any means, but thermal runaway dictates that the hotter something gets, the more power it uses, the more power it uses, the hotter it gets. An engineer can certainly correct me if I am wrong, but that's my basis behind wanting to delid to keep the variables as low as possible.

 

I can also run them all at 4ghz, if that helps. Or, overclock them all to 4.5 (the 7700k's stock speed) and see if that makes any difference. 

Yeah regarding thermal runaway, sure, that's definitely a thing. But I was thinking more along the lines of how much of a difference would it make running stock, since stock wouldn't be that insanely thermally taxing.

 

Additionally, I may be creating a pointless dilemma for myself, but the reason I initially suggested out of the box performance was that they were launched as such, and rated at those specs on, I'm assuming, more or less repeatable testing. Meaning that they did take into account the generational improvements and leeway they'd have with frequencies to maintain their target performance/TDP envelope for every launch. Now, that also implies that the numbers may be muddled by Intel's policies at the time (sandbagging, architectural stopgap measures etc, manufacturing), seeing as Sandy f.ex had an insane amount of headroom, and at 4.7, a 2600k is about as fast (for most intents and purposes) as a stock 4790k. That being said, a 4790k doesn't nearly have that much headroom. 

 

So, I think the above does make a case for fixed (albeit realistic) clocks. Something everyone would be comfortable accommodating, running, testing, and providing results with. What I really need is a hook for this reason. I wouldn't want it to be arbitrary. There must be a succinct reason to mess with the field and the potential FLOPS per chip in order to lock frequencies. 

 

I already have a hook for the "out of the box" idea. And that's basically denoting a performance index as Intel proposed, presented, rolled out and wanted to present their chips and comparing chronologically within the same family, which would also be pertinent for these new releases. The only reason I can think of for locking frequencies is "just 'cause", or, "The history of the Core Architecture according to the 4GHz benchmark".

[MageTank]

Just now, LooneyJuice said:

Yeah regarding thermal runaway, sure, that's definitely a thing. But I was thinking more along the lines of how much of a difference would it make running stock, since stock wouldn't be that insanely thermally taxing.

 

Additionally, I may be creating a pointless dilemma for myself, but the reason I initially suggested out of the box performance was that they were launched as such, and rated at those specs on, I'm assuming, more or less repeatable testing. Meaning that they did take into account the generational improvements and leeway they'd have with frequencies to maintain their target performance/TDP envelope for every launch. Now, that also implies that the numbers may be muddled by Intel's policies at the time (sandbagging, architectural stopgap measures etc, manufacturing), seeing as Sandy f.ex had an insane amount of headroom, and at 4.7, a 2600k is about as fast (for most intents and purposes) as a stock 4790k. That being said, a 4790k doesn't nearly have that much headroom. 

 

So, I think the above does make a case for fixed (albeit realistic) clocks. Something everyone would be comfortable accommodating, running, testing, and providing results with. What I really need is a hook for this reason. I wouldn't want it to be arbitrary. There must be a succinct reason to mess with the field and the potential FLOPS per chip in order to lock frequencies. 

 

I already have a hook for the "out of the box" idea. And that's basically denoting a performance index as Intel proposed, presented and rolled out and wanted to present their chips and comparing chronologically within the same family, which would also be pertinent for these new releases. The only reason I can think of for locking frequencies is "just 'cause", or, "The history of the Core Architecture according to the 4GHz benchmark".

Linpack MKL is terrifying, even at stock, lol. Delidded; liquid metal on the die, at stock on this 7700k, I can still hit 75C on MKL if I have enough bandwidth to feed AVX. The fact that it also stresses cache simultaneously, makes it all the more terrifying. 

 

As for testing their out of the box performance, I can do that, but then it raises another question. Do I disable advanced turbo, and use their default boost tables? Or do I allow them to take advantage of my motherboards all-core turbo functionalities? Certain boards allow you to run your highest single-core turbo on all cores simultaneously. If that is the case, the 4790k would run all 4 cores at 4.4ghz, the 6700k at 4.2ghz, and the 7700k at 4.5ghz. Might impact the results a little. 

[LooneyJuice]

4 minutes ago, MageTank said:

Linpack MKL is terrifying, even at stock, lol. Delidded; liquid metal on the die, at stock on this 7700k, I can still hit 75C on MKL if I have enough bandwidth to feed AVX. The fact that it also stresses cache simultaneously, makes it all the more terrifying. 

 

As for testing their out of the box performance, I can do that, but then it raises another question. Do I disable advanced turbo, and use their default boost tables? Or do I allow them to take advantage of my motherboards all-core turbo functionalities? Certain boards allow you to run your highest single-core turbo on all cores simultaneously. If that is the case, the 4790k would run all 4 cores at 4.4ghz, the 6700k at 4.2ghz, and the 7700k at 4.5ghz. Might impact the results a little. 

Yeah I understand regarding Linpack. I've never, ever run a load like that as I'm a bit weary of my CPU running long in the tooth. But I remember it can rip stuff to shreds even at stock. 

 

Regarding the latter, I think default boost tables is the way to go. The efficiency of their boost tables IMO would count towards their efficiency index, it's another thing that I do think should be embedded in the metric. I know a lot of boards do a better job, but that's the point. Not everyone would have the better board to represent the CPU in a better light.

[MageTank]

Just now, LooneyJuice said:

Yeah I understand regarding Linpack. I've never, ever run a load like that as I'm a bit weary of my CPU running long in the tooth. But I remember it can rip stuff to shreds even at stock. 

 

Regarding the latter, I think default boost tables is the way to go. The efficiency of their boost tables IMO would count towards their efficiency index, it's another thing that I do think should be embedded in the metric. I know a lot of boards do a better job, but that's the point. Not everyone would have the better board to represent the CPU in a better light.

I like that idea. It will also allow us to test once and for all the difference between Speedshift 1.0 and 2.0. Kaby feels faster, even though it's fundamentally the same as Skylake. I am not the only one that noticed this as well, as @done12many2 pointed it out as well. For people that keep their system running full bore at all times, it probably won't make the slightest difference to them, but changing power states and getting your clocks instantaneously is certainly beneficial for mobile users, or ITX users (me having previously been one).