PC Watch, a Japanese computer hardware website, acquired at least one Skylake i7-6700K and removed the heatspreader. With access to the bare die, they took some photos and tested a few thermal compound replacements, which quantifies how good (or bad) Intel's default thermal grease is. As evidenced by the launch of Ivy Bridge and, later, Devil's Canyon, the choice of thermal interface between the die and the lid can make a fairly large difference in temperatures and overclocking.
Image Credit: PC Watch
They chose the vice method for the same reason that Morry chose this method in his i7-4770k delid article last year. This basically uses a slight amount of torque and external pressure or shock to pop the lid off the processor. Despite how it looks, this is considered to be less traumatic than using a razer blade to cut the seal, because human hands are not the most precise instruments and a slight miss could damage the PCB. PC Watch, apparently, needed to use a wrench to get enough torque on the vice, which is transferred to the processor as pressure.
Image Credit: PC Watch
Of course, Intel could always offer enthusiasts with choices in thermal compounds before they put the lid on, which would be safest. How about that, Intel?
Image Credit: PC Watch
With the lid off, PC Watch mentioned that the thermal compound seems to be roughly the same as Devil's Canyon, which is quite good. They also noticed that the PCB is significantly more thin than Haswell, dropping in thickness from about 1.1mm to about 0.8mm. For some benchmarks, they tested it with the stock interface, an aftermarket solution called Prolimatech PK-3, and a liquid metal alloy called Coollaboratory Liquid Pro.
Image Credit: PC Watch
At 4.0 GHz, PK-3 dropped the temperature by about 4 degrees Celsius, while Liquid Metal knocked it down 16 degrees. At 4.6 GHz, PK-3 continued to give a delta of about 4 degrees, while Liquid Metal widened its gap to 20 degrees. It reduced an 88 C temperature to 68 C!
Image Credit: PC Watch
There are obviously limitations to how practical this is. If you were concerned about thermal wear on your die, you probably wouldn't forcibly remove its heatspreader from its PCB to acquire it. That would be like performing surgery on yourself to remove your own appendix, which wasn't inflamed, just in case. Also, from an overclocking standpoint, heat doesn't scale with frequency. Twenty degrees is a huge gap, but even a hundred MHz could eat it up, depending on your die.
It's still interesting for those who try, though.
Delidded
Delidded
You’re missing the point
You’re missing the point about de-lidding, Scott. You improve thermal headroom quite a bit, allowing you to push more volts and get further overclocks. 20C reduction is nothing to sneeze at and everyone should stick their middle fingers at Intel for not spending a few cents more per processor to get them cooled properly. 12 Billion dollar income per quarter should give them just enough money to pull that off.
Not to mention that after a proper de-lid, instead of having to resort to a high end AIO to keep things cool at the top end, you can grab something cheap like a Hyper 212+ and get the same effectiveness if not more than an AIO would have before the de-lid.
I think I addressed those
I think I addressed those points sufficiently, although I framed it as "Hey, Intel could make a separate SKU for to see how well it'll sell".
I would still like to make the point that this change is unlikely to give you an extreme boost in performance, like a GHz or something. It might not even be enough to give you any boost at all. It probably will, but I don't want people using my advice, possibly wrecking hardware, only to find out they lost the silicon lottery.
So obviously you did all of
So obviously you did all of your testing by popping off the IHS, lapping the bare die with the higher perf materials, then put the lid back, dropped your cooling solution back on, then performed your testing. As you’ve already voided the warrant and tainted the part for any future testing under normal circumstances, would you consider doing a followup piece evaluating overclockability mounting your cooling solution on the bare die utilizing a bracket like MSI’s Delid Die Guard or EKWB’s EK-Supremacy PreciseMount kit?
I would love to see some 5.5+GHz without resorting to LN2, as well as simply verifying that the previously mentioned kits continue to fit properly on the new chip.
Wasn’t me who did it. It was
Wasn't me who did it. It was PC Watch, a Japanese computer hardware site. Totally not going to risk a $350 processor that I paid for myself. lol.
Well i delided my 4770K and I
Well i delided my 4770K and I do not recommend to use the MSI guard. Seems like its not always working and you crack your CPU. Putting the lid back on will get likely the same performance and I do recommend deliding with a vice, its easy and save if you do it carefully. My CPU dropped by 20C.
I had a very bad experience
I had a very bad experience with the MSI die guard, do not recommend using it at all. Delidded a perfectly good 4770K w/o issue, put on the die guard, cracked the core and killed the mb as well. Not a fun day.
Writeup about the delidding debacle is here:
https://pcper.com/reviews/Editorial/Delidding-your-Intel-Haswell-CPU
Morry
I wouldn’t blame the die
I wouldn’t blame the die guard. It’s a simple device designed to hold the CPU in place while you mount the cooler. The reason it cracked is because of an uneven amount of pressure from the CPU cooler/block, which Is the users fault. In fact, it is strongly recommended to re-use the IHS as it evens the pressure and reduces die fracture probability.
I’ve de-lided at least two dozen processors dating back to Ivy bridge. I’ve killed one using the bare die mounting solution which is very easy to do.
Everyone should stick their
Everyone should stick their middle fingers to Intel’s ass not because they’re using ultra-shitty thermal paste for their stock, but because of the mere fact that they’re using thermal paste AT ALL instead of the GODLIKE direct soldering. Bow before your Sandy Bridge gods, Ivy/Failswell/Skyfake kiddie peasants! MUAAAAHAHAHAHAHAHAHAHAHAHAHAHAAAAAAH!!!11
More like pie in the Skylake,
More like pie in the Skylake, and no complete specifications until Intel is ready. I’ll bet that the graphics will have to get a little help from AMD or Nvidia to get those AAA’s above 30 on ultra, and why spend the money, get a Sandybridge, or Ivybridge or a little newer and there is not enough change to justify the cost if it’s not time for a complete system replacement. Intel is so slow with the meaningful improvements that its older SKU do just fine in the competition with the newer ones.
Plenty of time to wait for AMD’s systems on an interposer, and that silicon interposer can host a lot more traces to much higher bandwidth memory than even GDDR5, let alone DDR4 memory. The Interposer is a game changer possibly with a coherent fabric and maybe some cache coherency past level 3 cache and onto an interposer based extended cache, to go along with the HBM memory stacks. The more Intel drags its feet the more time AMD has to innovate, and Zen time is getting closer, along with the systems on an interposer with those fat GPU dies, and HBM.
Quite frankly, not even Ivy.
Quite frankly, not even Ivy. i5 2500K (if you can still find it, since they’re getting phased out like crazy) or i5 2550K are still absoeffinglutely more than enough, you don’t even have to get i7 2600K or i7 2700K because Hyper Threading plays menial function in all of this.
you showed the boringest
you showed the boringest photos of the article …
uh are you okay over there?
wat
wat
The sky in open on this one.
The sky in open on this one.
^^The sky is open on this
^^The sky is open on this one.^^
^^^The sky is open on this
^^^The sky is open on this one.^^^
^^^^The sky is open on this
^^^^The sky is open on this one.^^^^
Just an example:
Ambient day
Just an example:
Ambient day time temps today: 80 to 85 degrees(F).
3570K on H80: idle temps@ 38 to 43 degrees(C).
2500K on CNPS9900MAX: 31 to 34 degrees(C).
Just sayin’…..
Sandy i7 2600K-god here,
Sandy i7 2600K-god here, still going very strong. Screw all ‘dem horse sperm lovers.
My opinion about delidding:
My opinion about delidding: Thinking that delidding will give you more clocks is usually a dream. If you’re willing to delid you must already have D14 or equivalent. You’re pumping high voltages already.
At least with Haswell, which everybody complained the temps of. I ran OCN’s Haswell OC Thread and I logged everybody’s overclocks myself. Stop running the latest version of Prime and your CPU will thank you. By the time the temps were a problem the voltage was already more than suspicious.
Having said that, as an enthusiast at heart, I am willing to pay for lower temps for the sake of it. SiliconLottery offers delidding service for 50 smackerinos. Don’t go bare die and you should be fine.
Deli-dding doesn’t
Deli-dding doesn’t automatically give you more clocks, obviously. It just allows you to change default stock horse sperm for something much better (GODLIKE MX-4 from Arctic Cooling in my case, for example), and thus you’ll be getting much lower temps, which, in itself, would open a much bigger OverClocking headroom for your stone. The “lottery” still plays it’s part either way, of course, sure, but with much better thermal paste inside the stone you’re bound to get much better result anyway. Personally, I LOATHE the whole “no more direct soldering, just horse sperm since Ivy”-shtick of Intel’s, but there are still ways to countermeasure that, so…if you were inexperienced (or stupid) enough to buy something besides directly soldered stones back in the days – at least you can still fix that somewhat by changing the godawful stock horse sperm for much better thermal paste or similar solution.
Better result as in, better
Better result as in, better temps, that is undeniable. Better result as in, on average you get higher clocks, that’s very debatable. The lottery and whether you have the Prime95-or-die mentality are the two factors that determine your overclock. (Excluding extremes like running a stock cooler, etc.)
Skylake might be different, and I don’t have a sample to test. But I can say for Haswell.
Lower temps = higher
Lower temps = higher OverClock if voltage allows. What is there to not understand? As for “Prime-or-die”…for me personally it was always more of a “HyperPi/7-Zip/AIDA” thing.
AMD just filed a patent
AMD just filed a patent application for FPGAs added to the HBM stacks between the bottom die logic on the HBM stack and the memory stacks above. There will be no more worry about what new DX12, or Vulkan revisions are not implemented in hardware when any new graphics API feature set can be programmed into the FPGA’s on the HBM stacks, they’ll be able to add other extended features to those FPGAs on the stacks like Ray Tracing, and compression algorithms, or even more texture, SP units and Other decoding logic. Lots of interesting innovations for gaming coming from the exascale and HPC work that AMD is undertaking. Zen and the future ability to directly dispatch FP work to the GPU from the CPU is on the horizon from future HSA revisions, and I’ll wager that there will be GPU logic to chain those 32, or 64 bit FP units on the GPU into much larger than 512 bit effective registers for some workloads, in addition to what can be done of those FPGAs.
No way in hell would they be
No way in hell would they be able to add full Ray Tracing management in there. That shit would melt the stone, otherwise.
Where did you infer Full Ray
Where did you infer Full Ray tracing for gaming, and a lot of ray tracing can now be accelerated on the GPU, instead of requiring an expensive CPU for the rendering. Hell the PowerVR wizard has some built-in ray tracing circuity and it’s enough to add to the overall lighting and reflections of parts of a scene. What really heats things up is trying to do ray tracing on the CPU with all those processor threads an limited FPUs going at full throttle for hours for graphics workloads, not gaming, but even gaming/lighting can be improved by a little ray tracing on the GPU/FPGAs. GPUs with ACE(Asynchronous Compute Engines) and the thousands of vector FP compute units are what is needed for ray tracing workloads. Ray Tracing on the CPU is the biggest time thief of them all, but those ACEs are going to put an end to the need for powerful CPUs for ray tracing rendering. ACEs that can context switch and do logical branching are enabling ray interaction calculations to be done on the GPU with its massive ranks of vector FP units. No one said real time Ray Tracing for games, but that is not out of the realm of possibilities when the computing workloads are spread to the GPU, and on the FPGAs attached to the interposer, and gaming is likely to see some ray tracing enhanced frames but not fully ray traced scenes like is done on the PowerVR wizard. Not everybody is just about gaming graphics, there are other graphics uses, and a lot of gaming creation like creating the textures to bake on the mashes for gaming is done with high quality ray traced renderings, and backgrounds, Etc.
Lots of GPU cores on the GPU and some FPGAs on the same stacks as the HBM will usher in the era of distributed HSA computing on the interposer, and that includes much faster than PCIe data speeds via coherent fabrics also. Just go and read about the AMD exascale system for HPC/supercomputers and see some of the technology that will work its way down into the consumer side of things.
That is sure to get Intel moving a little faster, rather than milking the status quo!
I was talking about GPU part
I was talking about GPU part of the implementation, obviously.
Intel just cannot into integrated graphics, as of right now. And won’t be able into it for quite some years to come. Absolute majority of full-blown dedicated discrete GPUs can’t calculate Ray Tracing for shit, and we’re talking about very tiny on-CPU-integrated video here. As many crap people give to AMD and APUs throughout the globe these days, one sheer fact still stays solidly strong – AMD’s integrated graphics are WAAAY ahead of what Intel’s been doing for the last eight years. And then there’s the incoming of Zen…
And yet at the 2014 SIGGRAPH
And yet at the 2014 SIGGRAPH AMD demonstrated Ray Tracing accelerated on its GPUs, and those ACE on AMDs GPUs have branching and other functionality that can allow for Ray tracing and ray interaction calculations on the ACEs for its GPUs, the Ray Tracing demo was done with AMDs Fire Pro graphics cards. I’m very excited about Carrizo, and the potential of doing the ray tracing on Carrizo graphics, and even the current IPC level of Carrizo’s cores are sufficient for my graphics workloads in Blender, especially after Blender 2.75’s inclusion of cycles rendering for AMD GCN based SKUs. CPUs and namely the IPC performance of the CPUs core is not as big a deal for me once the ray tracing calculations move to the GPU, and that is the great news of AMDs ACEs on its GPUs with all graphics workloads moving to the GPU, the CPU is not as important of a factor as it once was. Now if I can just find a laptop with a full wattage(35W) Carrizo part, and a 1080p screen, and add to that a discrete GPU, and I no longer need an expensive quad core i7, for rendering workloads.
>The CPU is not as important
>The CPU is not as important of a factor as it once was.
There’s just one slight (absolutely insignificant, uh-huh) problem with this: you don’t “just calculate Ray Tracing” with your CPU. You do many things with your CPU. And if your CPU could do Ray Tracing good, but suck with other tasks, nobody would like it, at least in mass consumer segment. That’s where the main problem of APUs is. They’re OUTSTANDINGLY GREAT stones when it comes down to graphics themselves, but they suck quite much sheer PPC-wise. They’re just not good at raw (non-graphical) computing itself, because their PPC is usually three to five times lower than that of the similar offerings of Intel’s, in the same price range. APUs are truly great if you’re building a tiny emulation machine or media center, but when it comes down to heavy calculation tasks – they suck major donkey balls. Considering latest financial reports, Zen is AMD’s last chance, essentially. They really need to lift up PPC for their stone DRAMATICALLY to stay any relevant, and if Zen turns out to be very underwhelming, AMD is basically finished as a hardware-making company. Radeons alone (no matter how well HBM-based cards might sell in the near future) can’t save their ass, this is now very clear as the most sunniest day ever. Zen is AMD’s last hope. If Zen fails – it’s either complete buy-out by some large conglomerate monster (like Samsung or IBM) or absolute bankruptcy and fade into obscurity for eternity.
AMD could survive just fine
AMD could survive just fine on its GPU, and custom APU business alone, and Zen does not have to have any better IPCs than say SandyBridge, or IvyBridge, for AMDs HSA aware APUs to offload the work onto its GPU, or even FPGAs on the HBM stacks, along with the memory stacks and controller logic chip, and Intel will be in for some serious computational competition. We are talking about an entirely different processing paradigm and future Zen based APUs able to directly dispatch FP instructions to the GPU, and that includes the heterogeneous processing across all and any of the processors on an interposer, FPGAs and others as well. We are talking about tightly coupled and coherent processing with all these devices cache memory broadcasting their cache’s update status for all the processing systems to work on unified memory/processing tasks without things getting unbalanced. Those HPC APUs that AMD are working on are much more than just the sum of the individual processing parts, or simply the matter of besting Intel in IPCs on the CPU core for core, they bring the ability to seamlessly distribute the work across all computing devices on the interposer.
>AMD could survive just fine
>AMD could survive just fine on their GPUs/APUs alone
>Zen does not need to have better PPC than Sandy Bridge
Are you even for real? Both Faildozer and Vishera get PWN’ed massively by Sandy, not even mentioning more “fresher” stones. If AMD won’t get it shit back together soon, they’re friggin’ toast. Just because Phenom II X4 955 (and also it’s higher iterations) was an exceptionally successful stone, doesn’t mean that we should sit on it for ten years more to come, especially when Intel keeps on tick-tocking newer and better (even if slightly, still BETTER. That is a FACT) products each and every year all the way since Core2Duo.
>For AMD’s HSA-aware APUs to
I hope that you do know and understand that for HSA to work, IT’S IMPLEMENTATION SHOULD BE SUPPORTED BY SOFTWARE DEVELOPERS THEMSELVES FIRST? No matter how good HSA can be in both theory and practice – if there’s no software/apps which can utilize it partially OR fully – it’s useless. When (and ONLY when) HSA becomes a globally mainstream thing, then we’ll talk about it. Until then (if that time even comes at all), it’s pointless to discuss this. It’s like PhysX all ov er again, and I don’t like such empty ramblings which might lead to nowhere. We’ll see, but until then – NOPE. Stop.
Why do you keep using the
Why do you keep using the word “stones”? What does that even mean?
Remember that Intel just
Remember that Intel just bought Altera, they have been making FPGAs integrated with Xeons already, have Knights Landing, XPoint, on chip fabric and silicon photonics and the list goes on.
However, i think Fujitsu will hit USEFUL exascale before anyone else, as they have the current best architecture. Theyve had SPARC XIfx and Primehpc FX100 deployed since 2014 and thats 10x faster than K and probably equal to or better than Knights Landing, which hasnt been released yet.
Tadashi Watanabe is probably gonna make an exascale K type computer before he retires, if he ever plans to.
Lets not forget PEZYs Exascaler which is at the top of the Green500 or NEC who has their third Earth Simulator up and running and is also the most efficient in the Graph500. Their successor to SX-ACE is also at least 10x its sustained performance.
All this discussion about Nvidolia, Intel and AMD leaves out the entire country that makes the best supercomputers: Japan.
Yes But Intel has no GPU IP
Yes But Intel has no GPU IP worth mentioning, and the world market for supercomputers is not going to accept a single supplier for its systems. There is already OpenPower and Nvidia, and even AMD is capable of taking on a Power license from open power and I’m sure that would make IBM very happy, as IBM loves its supplies of parts(CPU, APU, GPU, whatever) parts to be second sourced and not under the control of one supplier. Altera is not the only maker of FPGAs and AMD has some possibilities for complete HSA aware APU type systems with CPUs designed on High Density Low power using design libraries that will be a must for the exascale systems coming online in the future. Remember the calculations are going to be mostly done on the GPU, on the vector(thousands of vector FP) units and not require the hotter running CPU cores with their high power usage lower density design libraries designs. AMD with Carrizo, and high density design libraries applied to CPUs, and has something IP wise that will be very effective in getting even more CPU functionality packed in a smaller space using less power, and Even the Server Xeons are not clocked as high as the consumer SKUs, the power budgets are very strict for exascale computing. AMD has the ability to produce its HSA type of HPC APUs with the GPU closely integrated with its CPUs and both CPU/GPU taped out on high density libraries for lower power usage and higher density packing of processing power for exascale systems, so Intel, and even the openpower competition need to get the power usage down moving towards exascale computing. Sure SPARC is out there and not just for Oracle systems, SPARC is out there and Power is out there also for anyone to license and use, for the Japanese on the Chinese to make supercomputers with those designs.
Edit: Japanese on the
Edit: Japanese on the Chinese
To : Japanese or the Chinese
If you look at a lot of the
If you look at a lot of the issues with certain exascale problems, i dont think HSA will be the solution. That just lets CPUs address the same memory as GPUs. Didnt Nvidia have something similar with GK110s GPU Direct already?
The problems with exascale computers is that all those small, low power nodes have to communicate with each other. When you have that many nodes, memory and interconnect bandwidth means they have to wait to do work.
That means giving the cores enough memory bandwidth and making on chip fabric and silicon photonics with extremely high bandwidth and low latency so the system isnt bottlenecked like petascale computers would be if scaled up.
Im not saying HSA is useless, but its not the technology thats going to scale up to a useful exaflop. If you look at current supercomputers that use a hybrid architecture, theyre not that efficient. Some things can use the CPU cores but theyre usually just there to feed the GPUs.
The most computationally efficient and high bandwidth systems on the Top500 are custom CPU architectures like the one in K and the very new XIfx Primehpc FX100 systems. When NEC releases their Aurora architecture(not the Knights Hill supercomputer with the same name) i expect it to show massive Graph500 efficiency since SX-ACE does with normal DDR3.
The systems with high Graph500 performance have a focus on CPU cores with massive memory bandwidth per core and the overall systems have high bandwidth interconnects. Thats whats required for exascale.
Also, tons of really low power cores also poses another problem, which is errors and component failures with so many nodes. ECC and checkpointing become more complex and difficult as you increase node count.
http://www.hpcwire.com/2015/08/06/reading-list-fault-tolerance-techniques-for-hpc/
Is Skylake limited AVX clock
Is Skylake limited AVX clock to TDP like Haswell-EP?
Skylake with the stock interface at 4.6GHz on running prime95 v28.5 Small FFTs is much cooler than 4790k at 4.6GHz.
I don’t think using direct
I don’t think using direct die heatsink on Skylake is wise because of the thinner PCB could get bend when force is pushing on the center. I hope some company can sell those stand like IHS without the middle plate, just the stand so CPU retention can hold the CPU down while able to mount direct die cooler. Distributing the push down force to prevent PCB damage while get direct die cooling efficiency.
It’s much safer to use a
It’s much safer to use a razor. I used an exacto knife to only cut a little off one corner, then used a thin plastic card to remove the rest. This reduced the chance of damaging the pcb to almost nothing.
Delided with liquid ultra dropped my top end by 22c. This allowed me to get to 4.8 when previously my chip would not post at that speed. Regardless of voltage. And it never even breaks 60 now on water. Ever.