Intel "Coffee Lake" (2017/2018, 14 nm)

[DmitryKo]

I'm not really sure why you have to dispute official datasheets - is it so hard to believe that 6-core Kaby Lake can have a peak current of ~140 A when 4-core processors are specified for 100 A?


Datasheets for server/workstation and HEDT processors have a few more details, such as duration of peak currents, permitted voltage drop and maximum package power - look up "DC Specifications" section. (Note that these CPUs do not have integrated graphics, so voltages/currents are for the processor cores.)

Skylake-X (LGA2066)
https://www.intel.com/content/www/u...rs/core/6th-gen-x-series-datasheet-vol-1.html
6-, 8-, 10-, 12-core processors (TDP 140 W) with Integrated VR 13.0
Nominal voltage (VCCIN) 1.8 V
Sustained current (ICCIN_TDC) 73 A
Can draw maximum current (ICCIN_MAX) of 190 A (with transient tolerance -0.19 V and maximum power 297 W) for up to 2 ms

Broadwell-EP (LGA2011-v3)
https://www.intel.com/content/www/us/en/processors/xeon/xeon-e5-v4-datasheet-vol-1.html
4-,6- and 8-core processors (TDP 140 W) with Integrated VR 12.5
Nominal voltage (VCCIN) 1.82 V
Sustained current (ICCIN_TDC) 80 A
Can draw maximum current (ICCIN_MAX) of 170 A (with transient tolerance -0.179 V and maximum power 280 W) for up to 4 ms

Haswell-E (LGA2011-v3)
https://www.intel.com/content/www/us/en/processors/core/core-i7-lga2011-3-datasheet-vol-1.html
6- and 8-core processors (TDP 140 W) with Integrated VR 12.5
Nominal voltage (VCCIN) 1.8 V
Sustained current (ICCIN_TDC) 82 A
Can draw maximum current (ICCIN_MAX) of 175 A (with transient tolerance -0.2 V and maximum power 270 W) for up to 4 ms

Haswell-EX (LGA2011-1)
https://www.intel.com/content/www/us/en/processors/xeon/xeon-e7-v3-datasheet-vol-1.html
4-, 6-, 8- core processors (TDP 115/140 W) with integrated EVRD 12.5
Operating voltage (VCC) 1.5 - 1.8 V
Sustained current (ICC_TDC) 71/86 A
Can draw maximum current (ICC_MAX) of 165/201 A (with transient tolerance -0.132/-0.16 V) for up to 2 ms, followed by a fall-down to (ICC_TDC+ICC_MAX)/2 for 14 ms, for a total of 10 s
Heat sink should withstand 1.2xTDP

Ivy Bridge-EX (LGA2011-1)
https://www.intel.com/content/www/us/en/processors/xeon/xeon-e7-v2-datasheet-vol-1.html
6-,8-,10-,12- core processors (TDP 105/130 W) with external VR 12.0
Operating voltage (VCC VID) 0.6 - 1.25 V
Sustained current (ICC_TDC) 115/140 A
Can draw maximum current (ICC_MAX) of 165/190 A (with transient tolerance -0.092/-0.152 V) for up to 3 ms. followed by a fall-down to (ICC_TDC+ICC_MAX)/2 for 9 ms, for a total of 10 s
Heat sink should withstand 1.2xTDP

Sandy Bridge-E (LGA2011)
https://www.intel.com/content/www/us/en/processors/core/4th-gen-core-i7-lga2011-datasheet-vol-1.html
4- and 6-core processors (TDP 140 W) with external VR 12.0
Operating voltage (VCC VID) 0.65 - 1.35 V
Sustained current (ICC_TDC) 135 A
Can draw maximum current (ICC _MAX) of 165 A for 5 s, followed for a fall-down to ICC_TDC in 20 s

---
TDC (Thermal Design Current) is "the sustained (DC equivalent) current that the processor is capable of drawing indefinitely and should be used for the voltage regulator thermal assessment."

TDP (Thermal Design Power) is "the average power, in watts, the processor dissipates when operating at Base Frequency with all cores active under an Intel-defined, high-complexity workload."

 

[green.pixel]

No. I hope these Intel non-hyperthreaded CPUs would die already.

Why? There will be 6 cores/threads and a decent board for like 300 dollars/euros. R5 1600(X) is a better deal all-around and available right now, but still it's not bad.

 

[sebbbi]

Another year, another round of CPUs is released, and 2600k lives to die another day.

2600k is still usable for gaming and office apps, but if you do anything that scales to multiple threads, 2600k clearly shows its age. See here: https://www.anandtech.com/bench/product/287?vs=2025. In multithreaded benchmarks 8700k beats 2600k consistently by 2x-3x. That is a big deal if you use software like this. Waiting is not fun. Even in some games 8700k is clearly faster. For example 8700k performance is 56% better on Ashes in Singularity.

I was pondering about replacing our (3 people office + my home computer) Skylake 6700k processors with Coffee Lake 8700k. Single threaded perf is 5%-10% better and multithreaded is 50%-60% better. Way better than most Intel iterative updates recently (MT perf matters to me). See here: https://www.anandtech.com/bench/product/1543?vs=2025. With benchmark results like these I would have already done it, but Intel decided to make 8700k not socket compatible with 6700k/7700k. We have expensive high end motherboards (with multi-way SLI + lots of stuff) and I don't really want to replace those. Also replacing only the CPU is not much work, but taking out the motherboard means basically rebuilding the whole computer again (we have water coolers, etc). And if you start thinking about a whole new workstation, then you can't ignore i9 and Threadripper... That 16-core Threadripper is do damn cost-effective for tasks that slow down our iteration time (rebuilding Unreal Engine code base, rebuilding thousands of shaders, compressing textures, mesh -> distance field volume conversion, etc).

There's lots of people who bought Skylake 6600/6600K/6700/6700K. The previous chip, Broadwell was dead on desktops. Then Ryzen arrived and changed everything. Now consumer 8-core chips are available with high multithreaded performance, but you are stuck with your Skylake. I have heard many Skylake owners drooling over Ryzen. Intel would have made the decision easy for their customers if they kept socket compatibility. 8700k upgrade would have been 350$ in total and super simple. 8700K offers up to 2x multithreaded perf versus 6600K. Would have been the best CPU-only upgrade ever for those people drooling over Ryzen's multithreaded perf increase in consumer space. I feel that Intel missed this opportunity.

 

[Kyyla]

Why? There will be 6 cores/threads and a decent board for like 300 dollars/euros. R5 1600(X) is a better deal all-around and available right now, but still it's not bad.

There is no reason for it to not have 12 threads other than Intel shitting in the customers mouth.

Other than that I agree it's a great chip.

 

[entity279]

There is no reason for it to not have 12 threads other than Intel shitting in the customers mouth.

Other than that I agree it's a great chip.

Might help your binning? Non HT chips use less power and hence it's easier for them to fit in a power envelope

 

[DmitryKo]

https://www.anandtech.com/bench/product/287?vs=2025. In multithreaded benchmarks 8700k beats 2600k consistently by 2x-3x.
I was pondering about replacing ... Skylake 6700k processors with Coffee Lake 8700k. Single threaded perf is 5%-10% better and multithreaded is 50%-60% better. https://www.anandtech.com/bench/product/1543?vs=2025.
I have heard many Skylake owners drooling over Ryzen.

It don't think anyone disagreed that more cores and higher turbo frequences make for a better overall performance. It's about the cost of geting on this bandwadon. Many 2600K (LGA1155) owners skipped the upgrade to 6700K (7700K/X), since not everyone is a content creator running MT-friendly workloads, https://www.anandtech.com/bench/product/287?vs=1543 - so at this point they would need to replace their entire system anyway, no matter if they upgrade to Rizen/Threadpiper or Coffee Lake/Cannonlake/Skylake-X.

I would have already done it, but Intel decided to make 8700k not socket compatible with 6700k/7700k. We have expensive high end motherboards (with multi-way SLI + lots of stuff) and I don't really want to replace those.
Intel would have made the decision easy for their customers if they kept socket compatibility

10 nm Cannonlake and Ice Lake would probably have less demanding power requirements in comparison to 14 nm processors, and there are indications that Intel didn't really plan to offer 6-core processors on the LGA1151 socket during the Kaby Lake Refresh cycle - so, in some sense, you should really blame AMD for the incompatibility.

 

[xEx]

It would be just easier to use a new socket. Using the same socket but not compatible it's inconvenient but I think it was because of the rush but still adding a pin wouldn't take that long...

 

[sebbbi]

10 nm Cannonlake and Ice Lake would probably have less demanding power requirements in comparison to 14 nm processors, and there are indications that Intel didn't really plan to offer 6-core processors on the LGA1151 socket during the Kaby Lake Refresh cycle - so, in some sense, you should really blame AMD for the incompatibility.

8700k is only 4W more than 7700k. Intel could have simply clocked it 100 MHz lower (4.7 GHz -> 4.6 GHz turbo) to hit 91W if power delivery was the problem. It would have still had the fastest single core perf of any CPU (100 MHz higher turbo than 7700k) and had around 50% higher MT perf than 7700k. Maybe they badly wanted to beat Ryzen in MT perf... but even at these clock rates, Ryzen is slightly ahead in most MT benchmarks.

 

[DmitryKo]

It's not about thermal design current (ICC_TDC) - these processors can draw significantly more current (ICC_MAX) at maximum power, with short 4 ms spikes for up to 10 s, and that's why Intel added more VCC (+5V rail) and VSS (ground rail) contacts, and also relocated the processor detect (SKTOCC#) pin so that new CPUs wouldn't start in old motherboards and vice versa...

 

[Grall]

Are those pins really necessary tho? People have OCd sky and kaby lake CPUs to mad levels and I've never heard of any issues with the CPU welding itself to the socket...

 

[entity279]

https://www.bit-tech.net/features/tech/motherboards/asus-interview-andrew-wu-rog-motherboard-pm/1/

He says that if Intel provided the right firmware, they could make mbs supporting Coffe lake on Z270

 

[DmitryKo]

bit-tech: So if you wanted and Intel let you, you could make Z270 compatible?
Andrew: Yes, but you also require an upgrade from the ME [Management Engine] and a BIOS update. Intel somehow has locked the compatibility.

Yeah, 'somehow'. Again, PROC_DETECT#/SKTOCC# (processor detected/socket occupied) signal is connected to a different pin - VRM module will not even start supplying +5V power to the socket if processor presence is not detected. BIOS firmware cannot really run until you actually power up the processor.

 

[entity279]

That pin was "reserved" in KBL. So you'd need to explain further as the argument you make above cannot thus stand.

 

[DmitryKo]

RSVD pins are not connected. SKTOCC# pin should be "pulled down directly (0 Ohms) on the processor package to the ground".

 

[entity279]

Maybe, can't contradict you on this one. But this would just mean that existing KBL motherboards won't work with Coffe.

Nothing stops a new revision therefore from accepting both processors.

 

[DmitryKo]

this would just mean that existing KBL motherboards won't work with Coffe

No, the old location has become RSVD on CFL.

Nothing stops a new revision therefore from accepting both processors

A new revision of 200-series that would accept KBL and CFL: won't happen, MB vendors will not really appreciate even more confusion for end users.

As for 300-series, we shall see, looks highly unlikely to me.

 

[entity279]

But we've reached the point where we can see that CFL incompatibility with z270 is a choice rather than a necesity, a choice that could have been aborted at multiple points in time, for example by launching a z270 rev 1 socket instead of z370. Of course that now with z370 being launched, it makes 0 sense to create the KBL+CFL z270 soket.

____

Anyway, with regards to pins being supplied with voltages, being left unconnected and what not. Moterbords are stupidly hardwiring their soket instead of employing some means of programable logic in pin-voltage assignments? At least something as complex as LGA 2066 would benefit from the latter

 

[DmitryKo]

e can see that CFL incompatibility with z270 is a choice rather than a necesity

Again, Coffee Lake 6-core CPUs can draw more peak current than old motherboards and 4-core processors were specified and tested for.

Package TDP package stays the same because of agressive power-management strategies in Intel processors, which actively limits peak frequencies. CPU power draw does exceed TDP values - it does all the time, although momentarily because Turbo Boost makes sure these spikes do not last long, so the average consumption stays within TDP limits.


Updating the older 200-series chipset to support CFL, and not requiring new chipsets and motherboards that fit the new requirements, would only result in confusion for both end users and retailers alike.

Moterbords are stupidly hardwiring their soket instead of employing some means of programable logic in pin-voltage assignments? At least something as complex as LGA 2066 would benefit from the latter

Programmable logic cannot switch 200 A currents.

LGA 2066 works by overprovisioning the socket - you first come with a pin layout that supports highest-end many-core processors with enourmous current draw, multiple independent memory channels, and inter-socket QPI links. Then you work it down to ensure that this pinout can also support less-featured processors.

 

[CarstenS]

It's not like intel was totally surprised that they were at some point shipping six-core CPUs and with Z270 not bein that old, if they wanted, they could've made a cleaner break there: Leave the pinout (mostly) alone, up the specs for the socket, keep backwards compatibility and at the same time be ready for just another CPU launch instead of a (platform) no-use intermediate, which is not even a good fit for non-K-CPUs right now, because all the boards will be vastly overengineered, not to mention overpriced.

 

[Arnold Beckenbauer]

Intel 8th Generation and 9th Generation Processor Lists Leaked: Coffee Lake Refresh?

What our sources do confirm is that all the i7-H and i7-HK processors will be based on Coffee Lake-H hardware, rather than Kaby Lake Refresh. This might be due to the 45W nature of the processors, and it is expected that the i3/i5/i7 naming will follow desktop Coffee Lake core counts, namely that Core i3 will be quad-core, Core i5 will be six-core and Core i7 parts will be six-core with hyperthreading.