You spend $300 on a processor and they skimp on something that would only cost them a cent? Would it not be better to simply up the cost of the processor to... I don't know, $300.01? Seriously. If you want to save a cent, fine, but do it on something that doesn't affect your user base to such a degree. Come on AMD, please get competitive again...
Yeah, it seems a bit weird to gimp the OC potential of these just to save on the thermal paste.
On the other hand - it may very well be that there's a special overclocker's edition on the horizon with more carefully selected units and a better heatspreader/paste...
The K SKU's are supposed to be the special overclocker's edition tho... Coming out with a Devil's Lake a year later is still a cash grab at best. Won't stop me from buying a 6700K next week, I'm fine with 4.6-4.7, but it's mildly insulting nonetheless if that's how it pans out.
Actually, you don't know if it would cost them a cent. The pure material cost may be about that, but there are other questions to be considered:
1. Whats the overall process time for the two methods? Do you need to leave one solution lying around for an hour before the CPU can go to the next process step, while the simpler solution can just continue without delay? Does the application process of the paste itself take much longer for the better solution?
2. Are there any impacts on the work environment? Is the better paste / soldering creating fumes that need to be removed and filtered by a more costly ventilation system?
3. What are the failure rates of paste application? Will the application of the better paste with the existing machines destroy a few CPUs?
Overall, I think (I hope?) that the lower material cost of the paste is the least likely reason for the chosen design.
Of course, the justification may not be technical at all. Maybe Intel once again just wants to sell you an optimized overclocking solution next year.
Longevity of the TIM is something else Intel needs to worry about. It's all good and well to delid a CPU and use high performance compound that needs reapplication every other year, but out-of-the-box they need to last 5-10 years minimum.
All of these can be represented dollars. Add this amount to the CPU price , apply your standard profit margin and hey presto, much better CPU at higher price that people are willing to buy!
So let it be $5. Or $10. Or $20. You seriously think people wouldnt pay that extra on K, S or T CPUs where its NEEDED? Please... They are paying strangers on forums to delid their CPUs $20 to $40. Why wouldnt they pay that to Intel to get it stock?
I remember last time with Devil's Canyon, they weren't able to replicate the processes that resulted in enthusiast-friendly pre-Haswell chips. They had to emulate those techniques with something new that performed close enough.
So it wasn't just about spending that extra penny. There are other costs such as additional equipment and materials (and training) that wouldn't normally be necessary.
That said, Intel overcame those limitations with Devil's Canyon and produced a satisfactory part. They probably could do it again.
Their total costs for Skylake could be as low as 10$. Really that's about the lowest realistic costs. It's more likely about 2x that and could be even plenty more if yields are not that good. There isn't much info about their costs and yields but there is some info about the costs at foundries so if Intel is not a lot less efficient one can make some guesses.
Yes i am talking cost so BOM, it's pretty much the same thing here. Their financial model is similar to other semi companies except they have much higher margins enabled by the lack of competition. Lets say a standard model for a semi company would be 50% gross margins,20% R&D, 10% others and 20% operating margins. In 2014 Intel had 55,87B revenue,gross margins were at about 64%, R&D close to 20%.. Selling ,general and admin was a bit high at 15% and operating income was some 27.5% so plenty high. And that's overall, including the many billions spent on getting rid of tablet chips. In PC and server their numbers are much better. Not to mention that volumes are going down , die sizes are going down, yet they manage to keep margins way high.
Actually, that's not how big publicly traded companies work. They optimise their selling price to make the most money possible (obviously). Now, if the price has already been optimised, then adding even a single cent would be suboptimal, leading to less cumulative revenue.
If, however, they are able to reduce their total manufacturing cost per unit by a single cent, then they can either adjust their selling price (if the reduced cost has made it sub optimal for the metric it was optimised for) or keep it as-is and just make a few million more at the end of the day.
You need to remember: the shareholders that invest in a company like Intel do not care about the advancement of personal computation; most don't even own at least a 0.4 HorsePower (300 Watts) computer. It is exactly the same mechanic that keeps single-threaded performace/clock rates/powerfullness at basically the same place year after year, just viewed from a different perspective.
It also seems like something that could bite them in the ass in the near future. If the thermal material is that bad it will break down even at stock speeds fairly quickly, possibly within warranty for systems running in poor environments (warehouses, factory floors, etc) where the ambient temp is high and the system will be prone to constant thermal throttling.
The K CPUs are already 20 - 40$ more expensive, just for the "service" of not locking them. People are fine with that, as the unlocked CPUs provide a lot more value to some. But wouldn't it be cool if you actually got anything better from Intel for that additional money? Apply liquid metal for the ultimate heat transfer (it's safe under the IHS) and charge 10$ more. Many people would be really happy to buy this.
The top of the die being closer to the mobo than the top of the heat spreader will prevent a lot of heatsink mounting designs from being able to make contact; and you're back in the world of 15 years ago where a botched heat sink install can crush the (edge of the) core and ruin the chip.
Well, it will no longer spread the heat; so don't do this if you have cooler with a direct heat-pipe design, your heat source may end up between two heat-pipes.
Also you take a bigger risk of damaging the CPU while installing the cooler. Nevertheless, some people have done exactly what you are talking about, and MSI have even designed a "Die Guard" support plate for exactly this purpose:
A "heat spreader" is just copper. It doesn't spread the heat any better than the copper base of a good heat sink, irregardless of heat pipe placing. The issue of mechanical protection is real, though.
Based on Ian's graph, these die are getting smaller and smaller and yet their TDPs aren't getting lower. So there's a heat density problem. I wonder if the IHS helps mitigate that.
On the flip side, if the IHS is off, then you can see exactly where the die is and place your thermal paste appropriately on these increasingly asymmetrical chips (5775C?). And you know for sure that there's no shitty material between the die and the IHS. You have more control over overclocking success.
And then there's always the durability argument. The IHS protects the fragile die.
I wonder how true this is. Sure the TDP remains the same... but these onboard GPUs are getting monstrous in size, and grow 20-50% each generation. My bet is that the CPU TDP is continually shrinking, and then they fill up the remaining TDP budget with as much GPU as they can. For those of us who rarely (if ever) touch the GPU on these chips it should not be a huge concern.
Sure it could, but will it? Client workloads have plateaued, but Intel churns out a predictable 5ish% performance bump every year.
It's always nice to have legroom for those unusual circumstances when your cpu needs to "race to sleep", but it's a diminishing returns kind if thing because the use case is already pretty well satisfied.
That's true for the 6700K and any E series CPU. The regular Skylakes will have 65 W TDPs at most, which is easy to cool and about the same power density as previous designs.
BTW: this issue of "where to apply thermal paste" sounds really strange. "Back in the day" you tried not to make a small dot of thermal paste exactly where the chip is, you tried to cover the entire surface with a thin layer. If Ian sees 15+°C too high temperatures with his pea dot method, I would say the pea was simply too small rather than in wrong spot.
Without the heat spreader you wont be able to close the lid on the LGA socket, or at least it wont clamp down on the processor correctly. The LGA socket actually pushes down on the heatspreader to compress the CPU against the LGA pins on the mobo, so without that in place the socket will be pressing against the PCB on the CPU which is probably pretty fragile -- overall not something I'd want to try!
Intel is Devil. With this cpu Intel says: hey! my HP process is fine......still it can be better. Next year, with the "refresh" Intel will give us the same cpu on the "same" process, with AVX2 enabled and likely some internal buses running at full speed, moreover there will be a far better TIM than actual and.......Intel: hey!!! the process finally is fine !!! look!! the cpu is cooler !! it overclocks. to 5.2Ghz!!!! do the upgrade please!!!!
Honestly a good Zen cpu is the only medicine to stop this crap monopolistic way to act.
You mentioned zen, but I don't think that'll help. Amd quoted like a 40% improvement over bulldozer in power thread performance, but there's a graph that shows skylake beating bulldozer by like 80%. I can't recall the graph, might've been in the tech report review. Needless to say, it's not looking good for amd.
I think AMD said 40% faster than Excavator, which is bdver4. Was Skylake beating Bulldozer, Piledriver or Steamroller in that graph, and at what test?
With each generation of the Bulldozer architecture improving performance by about 7-10%, Excavator would already be 30% or so faster than Bulldozer; another 40% on top would get that 80%, but Excavator's performance is untested at this point so speculation is all we have.
Sorry i meant AVX 512 :). Still the crap habit of putting an inexpensive TIM in a $350 cpu is here to stay i believe. A cpu aimed to entusiasts has to be soldered IMO.
In the following recent thread involving AMD AM2 processors there were at least two people complaining of increased processor temps---> http://forums.anandtech.com/showthread.php?t=24383... Apparently beginning with AM2, AMD used TIM which could dry out....increasing temps.
This, of course, is several years before Intel started doing the same (which started with IVB in 2012). With that mentioned, I wonder if we see the same thing (TIM drying out) happen with Intel processors over time?
My estimate for Broadwell was that they could do a quad with no GPU at about 60mm2- we have Core M die size and some die shots for Broadwell so easier to do the math. Skyalake does seem to be bigger than expected, i'm assuming that's on the GPU side, if it was Broadwell i would have expected it to be bellow 100mm2 with the same number of EUs but we really need a die shot to be sure about what's going on. So if Zen is competitive we could finally get a lot more cores at sane prices plus cheap and fast quads. Sure the GloFo process is a bit bigger but not everything needs to go 4x for 16 cores. If they do 16 cores in 250mm2 with 70ish yiels, that's 200 good dies per wafer so maybe 50$ per chip if a wafer costs about 10k$. From 50$ their cost to 300$ in retail would be okish - Intel has it's own fabs so their costs could be half but no good way to guess their costs.However they do make some claims about how their costs scale with each nodes. Would be hard to factor in the die size of what migrated from chipsets to the die in the last decade or so for a more accurate look at how Intel's dies are shrinking with the lack of competition but it would be interesting to plot perf and clock for clock perf, as well as retail prices and ofc include AMD. A A10-7850K Kaveri is 245mm2 on 28nm and retails for 129$m so how much is a fair price for Skylake really? Intel started to rip us off with Nehalem or better said Gulfy. What's the perf increase since? Maybe 2x over Nehalem depending on task? But that means clock for clock is just some 30+% gain while die size went from 263mm2 to about 60mm2 (excluding the GPU) and that doesn't even factor in what migrated from the chipset to the SoC. PS My estimate for Broadwell was some 164mm2, derived using the Core M info available so it's likely not all that accurate but should be close enough.
Somehow I don't think Dhrystone ALU, Memory Bandwidth, HyperPi or wPrime is in any way shape or form indicative of real-world encoding/transcoding. Perhaps the benchmarks we ran at AnandTech for the Skylake review would be a better initial start... :D
I'd be more interested in what the die-size implications are for future "EP" chips. Removing 50% of that die to remove the iGP leaves an incredibly small chip. In other words an Octi-core Broadwell-E would be only a little larger than the current die size... I would hope this means consumers will finally get beefed up or higher core counts with Skylake-E.
There is more hope from AMD and ARM. e could get a lot more cores that just 8. AMD could go with 12-16 cores at 300$ while for ARM a quad A72 cluster with 2MB L2 is some 8mm2 so they could offer 4x the cores Intel is offering in about the same die. Remains to be seen how high A72 could clock if you give it TDP room and no HT but on the other hand no need for Intel like margins either. 125mm2 with 32 A72 cores at 150$ would be interesting but nobody is pushing ARM on desktop so we won't get that. We might get some server chips like that but with a big price premium.
You're forgetting the extra L3 cache, the additional ring buses, the extra DRAM channels, the extra PCIe lanes, and the QPI interface. There's a lot of uncore on the -E processors.
Yeah look at the Sandy Bridge 4C/GT2 vs SnB-E 6 core (which is actually an 8 core die) -- 216 vs 435 sq mm -- so I would put 8c skylake die at about 225-250mm, although I bet they do a 10-core die, then like a 16-18 core die and then the bit 22-24 core die.
I did say larger, but not quite as much as you imply. The 6700K has 8MB L3 in this die shot already. So doubling it to an 8-core makes 16MB, very close to the 20MB in Haswell-E. So only add a little bit more space for the 4MB difference.
Swapping DMI 3.0 for QPI shouldn't change much either. And for the uncore, when we double the 6700K we automatically get the quad-channel memory controllers so we aren't adding extra to make up for that. Same for the extra PCIe lanes since we'd end up with 32 of them, almost as much as Haswell-E already.
Intel would need to add very little from a straight up doubled 6700K. If I recall correctly on some of these quad chips the iGP was up to 66% of the die too...
It would not surprise me at all to see a 'refresh' i7K chip down the line. They made bank with the i7 2700K after perfecting the process with the i7 2600K even though it was essentially the same chip with a higher stock clock.
It used to be normal to bring out some higher clocked version of the same chip on the same process every few months. It was called progress rather than "rip off" or "refresh" and didn't require new generation names or platforms.
That is incredible in terms of technology. Die sizes are so much more reasonable now, we have Quad Core with IGP at the same levels as Dual Core's with no IGP back with the Core 2 generation.
They can probably go Hexa Core in the mainstream and Deca Core in the high end. If there was enough competition lol.
If we assume same scaling from Haswell 8C to Skywell 8C as Haswell GT2 4C to Skywell GT2 4C it'll be 122.4/177*356 = 246 mm^2, so octo-core would be totally okay for a mainstream chip. On 28nm AMD and nVidia are making 600 mm^2 enthusiast graphics chips, Intel makes 662mm^2 Xeons on 22nm for the server market and they could surely make an ultra-enthusiast platform from that. But why would they do that when they can sell a 122mm^2 Skylake for $350 and a 356 mm^2 Haswell for $999? I'm sure both products have >80% gross margin, when their average is 65%.
What a rip off. Wow. $3 per square millimeter of silicon. Has Intel ever gouged its customers this badly? This is only marginally lager than a smartphone SoC.
Yeah, I will taking being "gouged" for a $300 part from Intel that we all know is superior to anything that AMD will have available next year rather than getting a "good deal" on a $900 FX-9590 space-heater that was criminally overpriced the day that AMD launched it.
So Intel's selling quite clearly substandard products. Worse, they've reduced the quality of their products so that they can then charge more for the ones that work properly. I'm disappointed that more isn't made of this in the article. Pretty soft reporting to be honest.
If LIsa Su honestly believed that a satanic sacrificial ritual where you were dissected alive would give AMD a "substandard" product along the lines of Skylake, she'd already be sharpening the knives and drawing the pentagram right now.
Admittedly Kabini includes the southbridge on-die, but when you consider that a full-bore Skylake part that AMD probably won't be able to match with Zen in 2016 is only 15% larger than an Atom-competitor from AMD, it shows just how big of a gap there really is right now.
Is Intel using the heat spreader as a mechanism for controlling the overclockability of processors as a way to intentionally gimp them? Using a car analogy, is this like the putting restrictor plates in engines on a new car model, then improving the "performance" of a subsequent car model simply by removing the restrictor plate? And in the process charging for a new "upgraded" model that really is just de-gimped?
Actually, to reply to my own post ... what exactly is the point of a CPU heat spreader anyway? Since you're going to be putting a heat sink onto the CPU, wouldn't a heat spreader only have value if it somehow spread heat more effectively than the heatsink surface?
Is it the case that whatever material the Intel heat spreader is made if is significantly better than whatever material the heat sink is made of at "spreading" heat? I doubt it, but maybe someone knows better than I do here.
And isn't it the case that the "heat spreader" is only better than a bare heat sink if the thermal interface material between the CPU die and the heat spreader is at least as good as would be used between the die and the heat sink?
Is it possible that the "heat spreader" is nothing more than a means for Intel to specifically decrease the cooling performance of the processor (by using inferior TIM), so as to reduce performance of a given CPU, in order to allow them to improve the performance of a subsequent generation of that CPU just by improving the TIM?
Someone, please tell me why "heat spreaders" are better than just good direct interfaces between CPU die and heat sink.
I can see one value of "heat spreaders", and that is as a protection layer to prevent the die from being damaged by direct contact with a heat sink. But I don't remember that being a particularly big problem pre-heat-spreaders ... am I wrong?
As I mentioned ... we had bare dies for decades, and I don't remember damage to the CPU when installing the cooler being a significant problem. Is your experience different?
It pulls double duty. It takes the heat from the concentrated area, and it also protects the die to avoid crushing the die with a heatsink: "Because of the sheer size of the Pentium 4’s core Intel employed an integrated heat spreader in order to take the concentrated heat being produced and spread it over a larger surface area. This makes it able to dissipate heat in a more effective manner" http://www.anandtech.com/show/661/8
Because it actually spreads the heat, making it easier to remove. Ask the guys with their water cooling who tried to cool it without heatspreader and actually got higher temps than with a heatspreader and LM.
I can give you five reasons for the heat spreader on Intel CPUs.
1. With the move to LGA, the metal clamp causes the CPU core to become recessed, so a spreader to raise the heatsink contact height works better.
2. CPU die size has gotten smaller, which equals less contact area for heatpipe type heatsinks. In which case, a heatspreader makes sense.
3. Clamping force of the heatsink itself has increased, bolting to the motherboard, instead of plastic tabs on the socket from P3/Athlon XP era.
4. Ease of installation for system integrators. Having the spreader means no particular care needs to be taken to protect the core from being cracked by careless heatsink installation.
5. Heatsink weight and size has gone up significantly, compared to what was available in the past. Which means there is a lot more chances of cracking a bare core during transportation with a massive heatsink.
Just curious though, what is the heat spreader made of that makes it conduct heat so much better than a heatsink would? Keep in mind that the heatspreader makes the same area of contact with the CPU die that a heatsink would.
Its made from copper. But thats not the point. The thinness of it causes the heat to spread across its actual surface much quicker before it can be transferred to the thermal paste and heatsink. That effectively increases the surface which can be used to transport heat to the heatsink. As I said, ask the guys that tried to cool a bare die with water cooling. They got worse temperatures than with a die exactly because of that effect. A huge heatsink right on the tiny die simply cant take heat up as efficiently, because most of its surface isnt used.
Making a heat conductor thin does not make it conduct better in the plane. Make it too thin and heat conductivity even drops, although you have to approach the phonon wavelengths for that (<1 µm, depending on the material).
You are right that the heat conduction for the system: heat spreader -> TIM -> Heat Sink is better than: die -> TIM -> Heat Sink
But that comparison is not representative of what happens during CPU cooling. the first case has to be: die -> TIM -> heat spreader -> TIM -> Heat Sink Given similar TIMs the transfer from the die to the heat spreader is just as bad as from the die to the heat sink. And with the heat spreader there's another poorly conducting interface to the heat sink. I don't know what those water cooling guys did, but a lot can go wrong in such a test apart from basic physics.
Go put a pan on your hot plate and see how fast it heats up. Then place a 10" thick block of the same material on it and see how fast that heats up. Its simple physics. People have tried several times to cool a die without HS, and have always had worse temps. Check the delidding threads on any known forums.
Are you trying to say that the pan heats up faster than the block? So I guess you're agreeing with me that the heat spreader is worse than the bare heatsink would be directly touching the TIM of the CPU? Because clearly the block taking longer to heat up means that it's absorbing more heat.
I think that the correct test would be to put the block on the pan and see how long it takes to heat up vs. putting the block directly on the hot plate. The block will heat up faster on its own. QED.
Well, see, here's the thing. Not all heatsinks are made equally. There are two types of heatpipe heatsinks. Ones with the direct contact heatpipes, and ones with a base plate soldered to the heatpipes.
Direct heatpipe contact can give the best results, but there is a catch to this, and it's that heatpipes don't seem to transmit heat to other heatpipes as easily. Which means that, if not all heatpipes are making full contact to the core, then maximum efficiency of heat distribution to the cooling fins cannot be reached.
And that's where a solid base plate comes in. Solid copper does a better job of distributing heat from the core to multiple heatpipes, which gives better heat distribution to the cooling fins and this increases cooling efficiency.
Now, heatsinks that don't use heatpipes, those typically don't benefit from a heat spreader, since they tend to have very thick, solid centers, which spread the heat out to cooling fins. Although, these types of heatsinks are typically inferior to heatpipe heatsinks.
These differences are, probably, more evident in video cards these days, where, if all the heatpipes don't make contact with the core, then a base plate is used, but if they do, then the direct contact method is used.
No, it's like putting seat belts, airbags and all that stuff into a car. Then they'll built an expensive racing version where they switch to expensive leight-weight seats etc. which would be a bad deal for regular users.
@Ian: "If the cost of the interface is reduced by 0.1 cents, then that's a significant saving on millions of processors. Devil's Canyon was a small subset of sales, so spending that extra for that specific crowd could be seen as beneficial to Intel's perspective by overclockers."
While this may be true for the exact TIM in use and the binning, that is most certainly not why they are using TIM instead of solder. The problem with using solder with such small dies is the issue of die cracking from repeated expansion stress. It's less of an issue on bigger dies, which is why the LGA2011 chips still have a soldered IHS.
Intel is destroying the Earth. More heat = less efficient = more leakage = more power needed = more carbon emission. Fuck you Intel, your indecent greed is well over 9000 now.
Take this from an IPC+ certified PCB Designer: More likely than not, there is NO WAY that is a 5 layer PCB. That just begs for warpage among other manufacturing issues.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
86 Comments
Back to Article
Raniz - Monday, August 10, 2015 - link
"Die size aside, Skylake also has a substantially thinner package than Skylake"The last "Skylake" should probably be Broadwell
Ian Cutress - Monday, August 10, 2015 - link
Thanks for the catch! It's a comparison to Haswell/Devil's Canyon.Wardrop - Monday, August 10, 2015 - link
You spend $300 on a processor and they skimp on something that would only cost them a cent? Would it not be better to simply up the cost of the processor to... I don't know, $300.01? Seriously. If you want to save a cent, fine, but do it on something that doesn't affect your user base to such a degree. Come on AMD, please get competitive again...Raniz - Monday, August 10, 2015 - link
Yeah, it seems a bit weird to gimp the OC potential of these just to save on the thermal paste.On the other hand - it may very well be that there's a special overclocker's edition on the horizon with more carefully selected units and a better heatspreader/paste...
Impulses - Monday, August 10, 2015 - link
The K SKU's are supposed to be the special overclocker's edition tho... Coming out with a Devil's Lake a year later is still a cash grab at best. Won't stop me from buying a 6700K next week, I'm fine with 4.6-4.7, but it's mildly insulting nonetheless if that's how it pans out.ShieTar - Monday, August 10, 2015 - link
Actually, you don't know if it would cost them a cent. The pure material cost may be about that, but there are other questions to be considered:1. Whats the overall process time for the two methods? Do you need to leave one solution lying around for an hour before the CPU can go to the next process step, while the simpler solution can just continue without delay? Does the application process of the paste itself take much longer for the better solution?
2. Are there any impacts on the work environment? Is the better paste / soldering creating fumes that need to be removed and filtered by a more costly ventilation system?
3. What are the failure rates of paste application? Will the application of the better paste with the existing machines destroy a few CPUs?
Overall, I think (I hope?) that the lower material cost of the paste is the least likely reason for the chosen design.
Of course, the justification may not be technical at all. Maybe Intel once again just wants to sell you an optimized overclocking solution next year.
Gigaplex - Monday, August 10, 2015 - link
Longevity of the TIM is something else Intel needs to worry about. It's all good and well to delid a CPU and use high performance compound that needs reapplication every other year, but out-of-the-box they need to last 5-10 years minimum.Arnulf - Monday, August 10, 2015 - link
All of these can be represented dollars. Add this amount to the CPU price , apply your standard profit margin and hey presto, much better CPU at higher price that people are willing to buy!Beaver M. - Monday, August 10, 2015 - link
So let it be $5. Or $10. Or $20. You seriously think people wouldnt pay that extra on K, S or T CPUs where its NEEDED? Please... They are paying strangers on forums to delid their CPUs $20 to $40. Why wouldnt they pay that to Intel to get it stock?ImSpartacus - Monday, August 10, 2015 - link
I remember last time with Devil's Canyon, they weren't able to replicate the processes that resulted in enthusiast-friendly pre-Haswell chips. They had to emulate those techniques with something new that performed close enough.So it wasn't just about spending that extra penny. There are other costs such as additional equipment and materials (and training) that wouldn't normally be necessary.
That said, Intel overcame those limitations with Devil's Canyon and produced a satisfactory part. They probably could do it again.
Beaver M. - Monday, August 10, 2015 - link
What? People still delidded it, because LM would still offer a huge improvement in temps.jjj - Monday, August 10, 2015 - link
Their total costs for Skylake could be as low as 10$. Really that's about the lowest realistic costs. It's more likely about 2x that and could be even plenty more if yields are not that good. There isn't much info about their costs and yields but there is some info about the costs at foundries so if Intel is not a lot less efficient one can make some guesses.ImSpartacus - Monday, August 10, 2015 - link
Are you talking bom?Intel has lots of overhead, you know.
jjj - Monday, August 10, 2015 - link
Yes i am talking cost so BOM, it's pretty much the same thing here.Their financial model is similar to other semi companies except they have much higher margins enabled by the lack of competition. Lets say a standard model for a semi company would be 50% gross margins,20% R&D, 10% others and 20% operating margins. In 2014 Intel had 55,87B revenue,gross margins were at about 64%, R&D close to 20%.. Selling ,general and admin was a bit high at 15% and operating income was some 27.5% so plenty high. And that's overall, including the many billions spent on getting rid of tablet chips. In PC and server their numbers are much better. Not to mention that volumes are going down , die sizes are going down, yet they manage to keep margins way high.
Xenonite - Monday, August 10, 2015 - link
Actually, that's not how big publicly traded companies work. They optimise their selling price to make the most money possible (obviously). Now, if the price has already been optimised, then adding even a single cent would be suboptimal, leading to less cumulative revenue.If, however, they are able to reduce their total manufacturing cost per unit by a single cent, then they can either adjust their selling price (if the reduced cost has made it sub optimal for the metric it was optimised for) or keep it as-is and just make a few million more at the end of the day.
You need to remember: the shareholders that invest in a company like Intel do not care about the advancement of personal computation; most don't even own at least a 0.4 HorsePower (300 Watts) computer. It is exactly the same mechanic that keeps single-threaded performace/clock rates/powerfullness at basically the same place year after year, just viewed from a different perspective.
Samus - Monday, August 10, 2015 - link
It also seems like something that could bite them in the ass in the near future. If the thermal material is that bad it will break down even at stock speeds fairly quickly, possibly within warranty for systems running in poor environments (warehouses, factory floors, etc) where the ambient temp is high and the system will be prone to constant thermal throttling.MrSpadge - Monday, August 10, 2015 - link
The K CPUs are already 20 - 40$ more expensive, just for the "service" of not locking them. People are fine with that, as the unlocked CPUs provide a lot more value to some. But wouldn't it be cool if you actually got anything better from Intel for that additional money? Apply liquid metal for the ultimate heat transfer (it's safe under the IHS) and charge 10$ more. Many people would be really happy to buy this.grrrgrrr - Monday, August 17, 2015 - link
Judging from die size 6700K should have been mid-ranged, obviously not worth spending money on that.Wardrop - Monday, August 10, 2015 - link
Question, is it unwise to take off the heat spreader and simply leave it off? What are the considerations in doing that?Gothmoth - Monday, August 10, 2015 - link
that´s actually a great idea! try it!!Ryan Smith - Monday, August 10, 2015 - link
The risk is cracking your CPU die. It's not guaranteed, but these CPUs are capped with an IHS for a reason.DanNeely - Monday, August 10, 2015 - link
The top of the die being closer to the mobo than the top of the heat spreader will prevent a lot of heatsink mounting designs from being able to make contact; and you're back in the world of 15 years ago where a botched heat sink install can crush the (edge of the) core and ruin the chip.ShieTar - Monday, August 10, 2015 - link
Well, it will no longer spread the heat; so don't do this if you have cooler with a direct heat-pipe design, your heat source may end up between two heat-pipes.Also you take a bigger risk of damaging the CPU while installing the cooler. Nevertheless, some people have done exactly what you are talking about, and MSI have even designed a "Die Guard" support plate for exactly this purpose:
http://www.madshrimps.be/articles/article/1000656/...
MrSpadge - Tuesday, August 11, 2015 - link
A "heat spreader" is just copper. It doesn't spread the heat any better than the copper base of a good heat sink, irregardless of heat pipe placing. The issue of mechanical protection is real, though.ImSpartacus - Monday, August 10, 2015 - link
Based on Ian's graph, these die are getting smaller and smaller and yet their TDPs aren't getting lower. So there's a heat density problem. I wonder if the IHS helps mitigate that.On the flip side, if the IHS is off, then you can see exactly where the die is and place your thermal paste appropriately on these increasingly asymmetrical chips (5775C?). And you know for sure that there's no shitty material between the die and the IHS. You have more control over overclocking success.
And then there's always the durability argument. The IHS protects the fragile die.
CaedenV - Monday, August 10, 2015 - link
I wonder how true this is. Sure the TDP remains the same... but these onboard GPUs are getting monstrous in size, and grow 20-50% each generation. My bet is that the CPU TDP is continually shrinking, and then they fill up the remaining TDP budget with as much GPU as they can. For those of us who rarely (if ever) touch the GPU on these chips it should not be a huge concern.Gigaplex - Monday, August 10, 2015 - link
Benchmarks showing power usage of just the CPU portion says otherwise. The CPU component alone can consume the entire TDP.ImSpartacus - Monday, August 10, 2015 - link
Sure it could, but will it? Client workloads have plateaued, but Intel churns out a predictable 5ish% performance bump every year.It's always nice to have legroom for those unusual circumstances when your cpu needs to "race to sleep", but it's a diminishing returns kind if thing because the use case is already pretty well satisfied.
ImSpartacus - Monday, August 10, 2015 - link
You're probably right.Intel sends to care about gpus now.
MrSpadge - Tuesday, August 11, 2015 - link
That's true for the 6700K and any E series CPU. The regular Skylakes will have 65 W TDPs at most, which is easy to cool and about the same power density as previous designs.BTW: this issue of "where to apply thermal paste" sounds really strange. "Back in the day" you tried not to make a small dot of thermal paste exactly where the chip is, you tried to cover the entire surface with a thin layer. If Ian sees 15+°C too high temperatures with his pea dot method, I would say the pea was simply too small rather than in wrong spot.
extide - Monday, August 10, 2015 - link
Without the heat spreader you wont be able to close the lid on the LGA socket, or at least it wont clamp down on the processor correctly. The LGA socket actually pushes down on the heatspreader to compress the CPU against the LGA pins on the mobo, so without that in place the socket will be pressing against the PCB on the CPU which is probably pretty fragile -- overall not something I'd want to try!Gondalf - Monday, August 10, 2015 - link
Intel is Devil. With this cpu Intel says: hey! my HP process is fine......still it can be better. Next year, with the "refresh" Intel will give us the same cpu on the "same" process, with AVX2 enabled and likely some internal buses running at full speed, moreover there will be a far better TIM than actual and.......Intel: hey!!! the process finally is fine !!! look!! the cpu is cooler !! it overclocks. to 5.2Ghz!!!! do the upgrade please!!!!Honestly a good Zen cpu is the only medicine to stop this crap monopolistic way to act.
ShieTar - Monday, August 10, 2015 - link
Yeah, because we don't get this kind of process-optimisation updates with GPUs, where we have a functioning duopoly. Or mobile phones.ImSpartacus - Monday, August 10, 2015 - link
Gpus are worse because most enthusiasts actually need more gpu performance. I can't say that I've lusted for more cpu performance in a long time.ImSpartacus - Monday, August 10, 2015 - link
You mentioned zen, but I don't think that'll help. Amd quoted like a 40% improvement over bulldozer in power thread performance, but there's a graph that shows skylake beating bulldozer by like 80%. I can't recall the graph, might've been in the tech report review. Needless to say, it's not looking good for amd.silverblue - Monday, August 10, 2015 - link
I think AMD said 40% faster than Excavator, which is bdver4. Was Skylake beating Bulldozer, Piledriver or Steamroller in that graph, and at what test?With each generation of the Bulldozer architecture improving performance by about 7-10%, Excavator would already be 30% or so faster than Bulldozer; another 40% on top would get that 80%, but Excavator's performance is untested at this point so speculation is all we have.
Xenonite - Monday, August 10, 2015 - link
What are you talking about?AVX2 has been enabled on Intel CPUs for 2 generations already.
Gondalf - Monday, August 10, 2015 - link
Sorry i meant AVX 512 :).Still the crap habit of putting an inexpensive TIM in a $350 cpu is here to stay i believe. A cpu aimed to entusiasts has to be soldered IMO.
Computer Bottleneck - Monday, August 10, 2015 - link
In the following recent thread involving AMD AM2 processors there were at least two people complaining of increased processor temps---> http://forums.anandtech.com/showthread.php?t=24383... Apparently beginning with AM2, AMD used TIM which could dry out....increasing temps.This, of course, is several years before Intel started doing the same (which started with IVB in 2012). With that mentioned, I wonder if we see the same thing (TIM drying out) happen with Intel processors over time?
Pissedoffyouth - Monday, August 10, 2015 - link
AMD solder the IHS to the die. So you can't remove it.patrickjp93 - Monday, August 10, 2015 - link
No they don't. Kaveri isn't soldered. Neither are Trinity or Richland if I remember correctly.jjj - Monday, August 10, 2015 - link
My estimate for Broadwell was that they could do a quad with no GPU at about 60mm2- we have Core M die size and some die shots for Broadwell so easier to do the math. Skyalake does seem to be bigger than expected, i'm assuming that's on the GPU side, if it was Broadwell i would have expected it to be bellow 100mm2 with the same number of EUs but we really need a die shot to be sure about what's going on.So if Zen is competitive we could finally get a lot more cores at sane prices plus cheap and fast quads. Sure the GloFo process is a bit bigger but not everything needs to go 4x for 16 cores. If they do 16 cores in 250mm2 with 70ish yiels, that's 200 good dies per wafer so maybe 50$ per chip if a wafer costs about 10k$. From 50$ their cost to 300$ in retail would be okish - Intel has it's own fabs so their costs could be half but no good way to guess their costs.However they do make some claims about how their costs scale with each nodes.
Would be hard to factor in the die size of what migrated from chipsets to the die in the last decade or so for a more accurate look at how Intel's dies are shrinking with the lack of competition but it would be interesting to plot perf and clock for clock perf, as well as retail prices and ofc include AMD.
A A10-7850K Kaveri is 245mm2 on 28nm and retails for 129$m so how much is a fair price for Skylake really?
Intel started to rip us off with Nehalem or better said Gulfy. What's the perf increase since? Maybe 2x over Nehalem depending on task? But that means clock for clock is just some 30+% gain while die size went from 263mm2 to about 60mm2 (excluding the GPU) and that doesn't even factor in what migrated from the chipset to the SoC.
PS My estimate for Broadwell was some 164mm2, derived using the Core M info available so it's likely not all that accurate but should be close enough.
bug77 - Monday, August 10, 2015 - link
Oh, you're gonna love these: http://www.hardocp.com/article/2015/08/05/intel_sk...Basically, about 10% improvement across four generations if you're doing encoding/transcoding.
Ian Cutress - Monday, August 10, 2015 - link
Somehow I don't think Dhrystone ALU, Memory Bandwidth, HyperPi or wPrime is in any way shape or form indicative of real-world encoding/transcoding. Perhaps the benchmarks we ran at AnandTech for the Skylake review would be a better initial start... :Dhttp://www.anandtech.com/show/9483/intel-skylake-r...
http://www.anandtech.com/bench/CPU/1052
Kougar - Monday, August 10, 2015 - link
I'd be more interested in what the die-size implications are for future "EP" chips. Removing 50% of that die to remove the iGP leaves an incredibly small chip. In other words an Octi-core Broadwell-E would be only a little larger than the current die size... I would hope this means consumers will finally get beefed up or higher core counts with Skylake-E.jjj - Monday, August 10, 2015 - link
There is more hope from AMD and ARM. e could get a lot more cores that just 8. AMD could go with 12-16 cores at 300$ while for ARM a quad A72 cluster with 2MB L2 is some 8mm2 so they could offer 4x the cores Intel is offering in about the same die. Remains to be seen how high A72 could clock if you give it TDP room and no HT but on the other hand no need for Intel like margins either. 125mm2 with 32 A72 cores at 150$ would be interesting but nobody is pushing ARM on desktop so we won't get that. We might get some server chips like that but with a big price premium.ViRGE - Monday, August 10, 2015 - link
You're forgetting the extra L3 cache, the additional ring buses, the extra DRAM channels, the extra PCIe lanes, and the QPI interface. There's a lot of uncore on the -E processors.extide - Monday, August 10, 2015 - link
Yeah look at the Sandy Bridge 4C/GT2 vs SnB-E 6 core (which is actually an 8 core die) -- 216 vs 435 sq mm -- so I would put 8c skylake die at about 225-250mm, although I bet they do a 10-core die, then like a 16-18 core die and then the bit 22-24 core die.extide - Monday, August 10, 2015 - link
the big* 22-24 core die, I mean.Kougar - Monday, August 10, 2015 - link
I did say larger, but not quite as much as you imply. The 6700K has 8MB L3 in this die shot already. So doubling it to an 8-core makes 16MB, very close to the 20MB in Haswell-E. So only add a little bit more space for the 4MB difference.Swapping DMI 3.0 for QPI shouldn't change much either. And for the uncore, when we double the 6700K we automatically get the quad-channel memory controllers so we aren't adding extra to make up for that. Same for the extra PCIe lanes since we'd end up with 32 of them, almost as much as Haswell-E already.
Intel would need to add very little from a straight up doubled 6700K. If I recall correctly on some of these quad chips the iGP was up to 66% of the die too...
CaedenV - Monday, August 10, 2015 - link
It would not surprise me at all to see a 'refresh' i7K chip down the line. They made bank with the i7 2700K after perfecting the process with the i7 2600K even though it was essentially the same chip with a higher stock clock.MrSpadge - Tuesday, August 11, 2015 - link
It used to be normal to bring out some higher clocked version of the same chip on the same process every few months. It was called progress rather than "rip off" or "refresh" and didn't require new generation names or platforms.Harry Lloyd - Monday, August 10, 2015 - link
Core i5-750 - 296 mm2 - 196 $Core i5-6400 - 122 mm2 - ~170 $
Six years of progress!
coldpower27 - Monday, August 10, 2015 - link
That is incredible in terms of technology. Die sizes are so much more reasonable now, we have Quad Core with IGP at the same levels as Dual Core's with no IGP back with the Core 2 generation.They can probably go Hexa Core in the mainstream and Deca Core in the high end. If there was enough competition lol.
Kjella - Monday, August 10, 2015 - link
If we assume same scaling from Haswell 8C to Skywell 8C as Haswell GT2 4C to Skywell GT2 4C it'll be 122.4/177*356 = 246 mm^2, so octo-core would be totally okay for a mainstream chip. On 28nm AMD and nVidia are making 600 mm^2 enthusiast graphics chips, Intel makes 662mm^2 Xeons on 22nm for the server market and they could surely make an ultra-enthusiast platform from that. But why would they do that when they can sell a 122mm^2 Skylake for $350 and a 356 mm^2 Haswell for $999? I'm sure both products have >80% gross margin, when their average is 65%.Harry Lloyd - Wednesday, August 12, 2015 - link
That is the point. NVIDIA is charging 330 $ for a card with a 400 mm2 GPU, and Intel is charging 240 $ for a 122 mm2 CPU. Intel are thieves.Shadowmaster625 - Monday, August 10, 2015 - link
What a rip off. Wow. $3 per square millimeter of silicon. Has Intel ever gouged its customers this badly? This is only marginally lager than a smartphone SoC.CajunArson - Monday, August 10, 2015 - link
Yeah, I will taking being "gouged" for a $300 part from Intel that we all know is superior to anything that AMD will have available next year rather than getting a "good deal" on a $900 FX-9590 space-heater that was criminally overpriced the day that AMD launched it.MrSpadge - Tuesday, August 11, 2015 - link
You're paying for Intels IP, advanced process and future research, not only for the production costs of the chip.boozed - Monday, August 10, 2015 - link
So Intel's selling quite clearly substandard products. Worse, they've reduced the quality of their products so that they can then charge more for the ones that work properly. I'm disappointed that more isn't made of this in the article. Pretty soft reporting to be honest.CajunArson - Monday, August 10, 2015 - link
If LIsa Su honestly believed that a satanic sacrificial ritual where you were dissected alive would give AMD a "substandard" product along the lines of Skylake, she'd already be sharpening the knives and drawing the pentagram right now.boozed - Monday, August 10, 2015 - link
What does AMD have to do with this?Gigaplex - Tuesday, August 11, 2015 - link
Because if Intel are so far ahead of the competition, clearly their products aren't all that substandard after all.CajunArson - Monday, August 10, 2015 - link
You left something out of that review.. Kabini's die size, which is about 107mm^2 (See here: http://www.anandtech.com/show/6977/a-closer-look-a...Admittedly Kabini includes the southbridge on-die, but when you consider that a full-bore Skylake part that AMD probably won't be able to match with Zen in 2016 is only 15% larger than an Atom-competitor from AMD, it shows just how big of a gap there really is right now.
tynopik - Monday, August 10, 2015 - link
"silicon dye"Ian Cutress - Monday, August 10, 2015 - link
Corrected ! :)extide - Monday, August 10, 2015 - link
Wow, so 4cores + GT2 is SMALLER than Conroe, which was JUST 2 cores, no uncore, no igp, etc. It's pretty much in the size range of mobile SOC's !bji - Monday, August 10, 2015 - link
Is Intel using the heat spreader as a mechanism for controlling the overclockability of processors as a way to intentionally gimp them? Using a car analogy, is this like the putting restrictor plates in engines on a new car model, then improving the "performance" of a subsequent car model simply by removing the restrictor plate? And in the process charging for a new "upgraded" model that really is just de-gimped?bji - Monday, August 10, 2015 - link
Actually, to reply to my own post ... what exactly is the point of a CPU heat spreader anyway? Since you're going to be putting a heat sink onto the CPU, wouldn't a heat spreader only have value if it somehow spread heat more effectively than the heatsink surface?Is it the case that whatever material the Intel heat spreader is made if is significantly better than whatever material the heat sink is made of at "spreading" heat? I doubt it, but maybe someone knows better than I do here.
And isn't it the case that the "heat spreader" is only better than a bare heat sink if the thermal interface material between the CPU die and the heat spreader is at least as good as would be used between the die and the heat sink?
Is it possible that the "heat spreader" is nothing more than a means for Intel to specifically decrease the cooling performance of the processor (by using inferior TIM), so as to reduce performance of a given CPU, in order to allow them to improve the performance of a subsequent generation of that CPU just by improving the TIM?
Someone, please tell me why "heat spreaders" are better than just good direct interfaces between CPU die and heat sink.
I can see one value of "heat spreaders", and that is as a protection layer to prevent the die from being damaged by direct contact with a heat sink. But I don't remember that being a particularly big problem pre-heat-spreaders ... am I wrong?
Brett Howse - Monday, August 10, 2015 - link
They protect the die from damage when putting the cooler onbji - Monday, August 10, 2015 - link
As I mentioned ... we had bare dies for decades, and I don't remember damage to the CPU when installing the cooler being a significant problem. Is your experience different?bji - Monday, August 10, 2015 - link
Also why not call it a "CPU protector" then, instead of "heat spreader"?Brett Howse - Monday, August 10, 2015 - link
It pulls double duty. It takes the heat from the concentrated area, and it also protects the die to avoid crushing the die with a heatsink:"Because of the sheer size of the Pentium 4’s core Intel employed an integrated heat spreader in order to take the concentrated heat being produced and spread it over a larger surface area. This makes it able to dissipate heat in a more effective manner"
http://www.anandtech.com/show/661/8
Beaver M. - Monday, August 10, 2015 - link
Because it actually spreads the heat, making it easier to remove. Ask the guys with their water cooling who tried to cool it without heatspreader and actually got higher temps than with a heatspreader and LM.meacupla - Monday, August 10, 2015 - link
I can give you five reasons for the heat spreader on Intel CPUs.1. With the move to LGA, the metal clamp causes the CPU core to become recessed, so a spreader to raise the heatsink contact height works better.
2. CPU die size has gotten smaller, which equals less contact area for heatpipe type heatsinks. In which case, a heatspreader makes sense.
3. Clamping force of the heatsink itself has increased, bolting to the motherboard, instead of plastic tabs on the socket from P3/Athlon XP era.
4. Ease of installation for system integrators. Having the spreader means no particular care needs to be taken to protect the core from being cracked by careless heatsink installation.
5. Heatsink weight and size has gone up significantly, compared to what was available in the past. Which means there is a lot more chances of cracking a bare core during transportation with a massive heatsink.
bji - Monday, August 10, 2015 - link
Thank you for your reply, much appreciated.Just curious though, what is the heat spreader made of that makes it conduct heat so much better than a heatsink would? Keep in mind that the heatspreader makes the same area of contact with the CPU die that a heatsink would.
Beaver M. - Tuesday, August 11, 2015 - link
Its made from copper. But thats not the point. The thinness of it causes the heat to spread across its actual surface much quicker before it can be transferred to the thermal paste and heatsink. That effectively increases the surface which can be used to transport heat to the heatsink.As I said, ask the guys that tried to cool a bare die with water cooling. They got worse temperatures than with a die exactly because of that effect. A huge heatsink right on the tiny die simply cant take heat up as efficiently, because most of its surface isnt used.
MrSpadge - Tuesday, August 11, 2015 - link
Making a heat conductor thin does not make it conduct better in the plane. Make it too thin and heat conductivity even drops, although you have to approach the phonon wavelengths for that (<1 µm, depending on the material).You are right that the heat conduction for the system:
heat spreader -> TIM -> Heat Sink
is better than:
die -> TIM -> Heat Sink
But that comparison is not representative of what happens during CPU cooling. the first case has to be:
die -> TIM -> heat spreader -> TIM -> Heat Sink
Given similar TIMs the transfer from the die to the heat spreader is just as bad as from the die to the heat sink. And with the heat spreader there's another poorly conducting interface to the heat sink. I don't know what those water cooling guys did, but a lot can go wrong in such a test apart from basic physics.
Beaver M. - Tuesday, August 11, 2015 - link
Go put a pan on your hot plate and see how fast it heats up.Then place a 10" thick block of the same material on it and see how fast that heats up.
Its simple physics. People have tried several times to cool a die without HS, and have always had worse temps. Check the delidding threads on any known forums.
bji - Tuesday, September 1, 2015 - link
Are you trying to say that the pan heats up faster than the block? So I guess you're agreeing with me that the heat spreader is worse than the bare heatsink would be directly touching the TIM of the CPU? Because clearly the block taking longer to heat up means that it's absorbing more heat.I think that the correct test would be to put the block on the pan and see how long it takes to heat up vs. putting the block directly on the hot plate. The block will heat up faster on its own. QED.
meacupla - Tuesday, August 11, 2015 - link
Well, see, here's the thing.Not all heatsinks are made equally.
There are two types of heatpipe heatsinks. Ones with the direct contact heatpipes, and ones with a base plate soldered to the heatpipes.
Direct heatpipe contact can give the best results, but there is a catch to this, and it's that heatpipes don't seem to transmit heat to other heatpipes as easily. Which means that, if not all heatpipes are making full contact to the core, then maximum efficiency of heat distribution to the cooling fins cannot be reached.
And that's where a solid base plate comes in. Solid copper does a better job of distributing heat from the core to multiple heatpipes, which gives better heat distribution to the cooling fins and this increases cooling efficiency.
Now, heatsinks that don't use heatpipes, those typically don't benefit from a heat spreader, since they tend to have very thick, solid centers, which spread the heat out to cooling fins. Although, these types of heatsinks are typically inferior to heatpipe heatsinks.
These differences are, probably, more evident in video cards these days, where, if all the heatpipes don't make contact with the core, then a base plate is used, but if they do, then the direct contact method is used.
MrSpadge - Tuesday, August 11, 2015 - link
No, it's like putting seat belts, airbags and all that stuff into a car. Then they'll built an expensive racing version where they switch to expensive leight-weight seats etc. which would be a bad deal for regular users.ZeDestructor - Monday, August 10, 2015 - link
@Ian: "If the cost of the interface is reduced by 0.1 cents, then that's a significant saving on millions of processors. Devil's Canyon was a small subset of sales, so spending that extra for that specific crowd could be seen as beneficial to Intel's perspective by overclockers."While this may be true for the exact TIM in use and the binning, that is most certainly not why they are using TIM instead of solder. The problem with using solder with such small dies is the issue of die cracking from repeated expansion stress. It's less of an issue on bigger dies, which is why the LGA2011 chips still have a soldered IHS.
There's a few papers out there on the subject.
Dug - Thursday, August 13, 2015 - link
Someone could make a nice little side business selling these with corrected thermal paste.I would pay for it done, as I don't have the dexterity anymore to do it myself.
Ramon Zarat - Thursday, August 13, 2015 - link
Intel is destroying the Earth. More heat = less efficient = more leakage = more power needed = more carbon emission. Fuck you Intel, your indecent greed is well over 9000 now.estarkey7 - Friday, August 14, 2015 - link
Take this from an IPC+ certified PCB Designer: More likely than not, there is NO WAY that is a 5 layer PCB. That just begs for warpage among other manufacturing issues.