Intel Skylake / Kaby Lake

[frozentundra123456]

Threadripper says hi.

So does the threadripper thread.

 

[Dave3000]

I'm thinking about upgrading from an i7-4930k to an i7-7820x. I mainly game with my PC and only one game I frequently play is really hammering my CPU and that game is X-Plane 11. I'm also considering the Corsair H115i liquid CPU cooler. Will that CPU cooler be enough to prevent the i7-7820x from throttling to base clock from it's all core stock turbo clock during stress testing such as Prime 95 or Orthos?

 

[TheGiant]

I'm thinking about upgrading from an i7-4930k to an i7-7820x. I mainly game with my PC and only one game I frequently play is really hammering my CPU and that game is X-Plane 11. I'm also considering the Corsair H115i liquid CPU cooler. Will that CPU cooler be enough to prevent the i7-7820x from throttling to base clock from it's all core stock turbo clock during stress testing such as Prime 95 or Orthos?

I think if you game wait a month or 2 for CoffeeLake desktop 6C. Imagine current KBL performance with 6C at 5GHz, yay!

 

[TheGiant]

CFD you say? You're most likely to be memory latency or bandwidth bound, or even both.

So what do you recommend?
Buy the SKL Xeon with 6channel? Or TR or EPYC even with 8ch?

 

[tamz_msc]

So what do you recommend?
Buy the SKL Xeon with 6channel? Or TR or EPYC even with 8ch?

That's a difficult question to answer, since only you know about the specifics of your code. What I know is that CFD can be memory latency and/or bandwidth dependent in many cases, and not much dependent on peak FP performance.

Here is a short 15 min talk that you may find useful:

 

[Wyrm]

So what do you recommend?
Buy the SKL Xeon with 6channel? Or TR or EPYC even with 8ch?

There are too many factors involved to give you any meaningful advice. I suspect you will not be happy with tamz_msc's answer but it is really the best one can give you, given minimal information about your software. For example, here is just a list of factors that need to be taken into account:

• Does your problem fit in 128GB? If not you need a server processor because SKL-X does not support more than 128GB memory. Your software will end up thrashing to disk and running painfully slow.
• Do you live next to a nuclear reactor or at a high altitude (e.g. Colorado)? If yes you need ECC memory support, hence a server processor like a Xeon would be prudent to consider.
• Is your software well optimized (data structures are SSE friendly)? If yes SKL-X may be a good choice.
• Do you rely on external libraries? If yes you need to read their documentation to see what architectures they support. They may not even support SSE/AVX.
• What does profiling of your CFD software show? What are the top 2 tasks on that profile measured by wall clock time? You need to choose a processor for those tasks. For example, the profiling may show that your code is dominated by a sparse linear solver and a CFD meshing algorithm. The former doesn't scale well with multiple threads (both direct sparse solvers and multigrid preconditioned solvers don't scale well past 4 cores). Meshing algorithms do scale well to any number of cores. In that case, you'll need both in which case SKL-X is a great choice because of its Turbo Max 3.
• Are you willing to tolerate high power dissipation with proper cooling? If not Xeon is a better choice because of its lower clocks to keep it cooler than SKL-X. But the price is that it doesn't run single-threaded portion of your code fast enough.
• Are you willing/capable of overclocking your memory? Given the info I've seen SKL-X supports higher bandwidth memory than Ryzen. It may change with Threadripper, so I'd wait to see what TR motherboards can give you.
• And so on.

Above all else, please keep in mind that Threadripper is rumored to ship at the end of July. If you can wait it may be wise to see what AMD can deliver. There are also SKL-X with higher core counts and TDP of 160W shipping in August. In addition, extra time will give the motherboard vendors an opportunity to improve their VRM cooling and ship better motherboards. There are just too many variables and multiple great CPU choices suddenly showing up. I'd put it this way: 2017 is the worst year to advice anyone about what CPU they should buy.

 

[wildhorse2k]

@aigomorla I think the 7960X will be very limited by power and temps for any useful amount of overclocking. The 7920X, might be an interesting from the aspect of having a challenge in terms of overclocking and performance - with VRMs and CPU under custom H2O (or better!).

I think the 7980XE will be best if you can afford it. OCed 7900X is extremely energy inefficient at 4.5ghz+. Yes you get extra free performance but at cost of very high power draw like 400w from what de3bauer found. Noise from fans would probably be the main reason for me not to Oc 7900x much.

The 7980xe will have low base clocks and thus be much more energy efficient. 7920x will have the same bad latencies (like 7980xe) and only 2 more cores than 7900x. Ocing 7980xe will be mainly about adjusting turbo clocks so that you get almost 7900x performance at few thread apps and perhaps 200mhz for all turbo clock as then heat will become an issue. The good thing about 7980xe is that 100% of samples will be capable of achieving the same target clocks.

Question is whether IMC of 7980xe will be as good as 7900x and whether mesh will be capable of running at 3.2Ghz.

 

[beginner99]

I think if you game wait a month or 2 for CoffeeLake desktop 6C. Imagine current KBL performance with 6C at 5GHz, yay!

Yeah. Skylake-X makes little sense if you mostly use it for gaming. Waiting for Coffelake or buying ryzen now are much better alternatives.

 

[frozentundra123456]

Yeah. Skylake-X makes little sense if you mostly use it for gaming. Waiting for Coffelake or buying ryzen now are much better alternatives.

For primarily gaming in this price bracket, I would not recommend Ryzen. The only Ryzen chip that makes sense for gaming is the 1600X which gives very good performance for the price. The 7700k (and even BW-E) are still faster overall for gaming than Ryzen. Coffee Lake should be even better for highly threaded games while giving away little in single thread performance.

 

[Bassman2003]

No.

The VRMs won't heat up as much, but the CPU will still get very hot due to (a) heat spots, (b) bad conductivity of the thermal interface between die and the so-called heat spreader, (c) still very high power to be dissipated. And it may even throttle occasionally.

For gaming with a 4.5 GHz CPU, use the mainstream platform, not X299 whose CPUs are derived from the scalable but higher latency server CPU design.

For rendering, don't use a 4.5 GHz CPU, use CPUs at 3.5 GHz or less. Or at least configure the CPU to run at 4.5 GHz only under lowly threaded loads, and keep it at stock clocks under multithreaded loads.

I am a video editor and have been waiting to upgrade my 4790k CPU. Initially, the 7900x looked like a great fit but I am concerned about heat (and fan noise). My hope is to have the 7900x running at 4.5GHz for all cores when I encode and when I am editing for timeline playback. Right now I have a Corsair H110 AIO. I am planning on buying from Silicon Lottery in hopes to get the temps down.

Am I unrealistic to expect the 7900x to run at 4.5GHz with my cooler and keep encoding temps below 70C?

Should I be looking at the 7820x for 4.5GHz and cooler temps?

While it would be easy to say "run encoding at 3.5GHz" but CPU upgrades are about saving time in my world. So I would want as fast a clock speed I can get away with!

Thanks for your input.

 

[wildhorse2k]

I think the main advantage of high frequency OC skylake-x is capable is single/few thread performance and not all core performance as that comes at a great cost of TDP and noise. If you need better all core performance you should seriously consider paying $200 or for 2 more cores.

Anybody considering silicon lottery has the extra $200 to buy 2 extra cores to get performance without extra noise.

 

[StefanR5R]

I am a video editor and have been waiting to upgrade my 4790k CPU. Initially, the 7900x looked like a great fit but I am concerned about heat (and fan noise). My hope is to have the 7900x running at 4.5GHz for all cores when I encode and when I am editing for timeline playback. Right now I have a Corsair H110 AIO.

I don't have a Skylake-X myself. But based on the reviews I have read so far, my understanding is: Under full multicore load, the processor will lower its clock automatically to stay in the package power limit or/and to stay below temperature limits. You can raise either or both in the BIOS and then might be able to encode at 4.5 GHz on all cores if you run the cooler at full tilt and have good airflow in the case to cool the mainboard appropriately too. But only if your encoding software is not heavy on AVX.

If your encoder makes good use of AVX, then simply dial in a suitable AVX clock offset. Meaning, automatically and temporarily clock those cores down which detect an AVX workload. I used this feature on Broadwell-E and it is great. Despite the lower AVX clock, your encoding will still go plenty fast due to its very high utilization of the hardware available to it. And your non-AVX workloads will not be affected by the AVX clock offset.

I am planning on buying from Silicon Lottery in hopes to get the temps down.

Do you mean to buy a delidded processor? If so, this will get core temps down dramatically. But it means either direct-die cooling, which the cooler base and mounting system need to support, or the potential need to delid + relid again if/ when the 3rd party thermal interface has degraded.

Am I unrealistic to expect the 7900x to run at 4.5GHz with my cooler and keep encoding temps below 70C?

The H110 is merely a slim 280 mm radiator, and that is not very much if we talk about power dissipation far beyond 150 W. You will definitely need high fan speeds.

On the other hand, if the CPU is not delidded, then core temperatures will quickly and often reach the configured thermal limit, independently of radiator size and fan speeds, simply because of the high thermal resistance of Intel's stock TIM (a layer of paste much much thicker than a soldered interface, and thicker than with properly relidded CPUs).

And as noted it depends a lot on whether or not the encoder uses AVX.

Should I be looking at the 7820x for 4.5GHz and cooler temps?

Going with 7820X instead of 7900X means several things:

• 8 instead of 10 cores: Under full load at same clock simply means somewhat lower power draw but somewhat longer encoding times. How much so depends on how well the encoder scales with core count. If the encoder scales ideally, then of course the 7820X would draw 80 % of the power but take 125 of the time (again, if both processors were compared at same clock).
  
• Rather than 2 AVX-256 ports and 1 AVX-512 port per core on the 7900X, you'd get just the 2 AVX-256 units on the 7820X. Whether or not your encoder would use the additional AVX port of the 7900X depends on the encoder. It is more likely that it does not. Maybe the software vendor can tell you. If it is prepared to use the additional AVX port of the 7900X, then this would mean increased speed (but by far less than double because of imperfect scaling) at increased power draw of course.
• Only 28 instead of 40 PCIe lanes, which is possibly irrelevant for your use case.

Related to the first bullet point: If 7820X and 7900X were compared not at same clocks but at same power budget, and the encoder scales well with core count, then the 7900X would be faster. This is because you could run its 125 % cores at more than 80 % the clock compared to 7820X.

While it would be easy to say "run encoding at 3.5GHz" but CPU upgrades are about saving time in my world. So I would want as fast a clock speed I can get away with!

Encoding on all cores, even if no AVX is involved, will likely mean that clocks oscillate between stock all-core turbo and base clock, unless you lift both package power limit and temperature limit, and cool heavily. Much depends on your software, and some on how well you can (and are willing to) optimize/ oversize the cooling.

Again, these are just my conclusions based on the reviews I have read so far. My own overclocking experince is limited to Sandy Bride, Ivy Bridge-E, and Broadwell-E. But behavior of Skylake-X when overclocked appears to be highly similar to Broadwell-E but with the frequency wall of the transistors shifted past the limits imposed by power and temperature. (Edit: But temperatures regressed relative to BDW-E due to thermal paste instead of solder.)

------------

BTW, Tom's Hardware will apparently publish another socket 2066 related review this week, based on additional testing of the 7900X in their German lab (probably not yet with 7820X results) [source, in German]. AFAIU this review will include analysis at different package power levels, and also investigate topics more or less loosely related to those raised by der8auer, but from a different perspective.

 

[Edrick]

I'm thinking about upgrading from an i7-4930k to an i7-7820x. I mainly game with my PC and only one game I frequently play is really hammering my CPU and that game is X-Plane 11. I'm also considering the Corsair H115i liquid CPU cooler. Will that CPU cooler be enough to prevent the i7-7820x from throttling to base clock from it's all core stock turbo clock during stress testing such as Prime 95 or Orthos?

Yes it will be. My H100i easily prevents my 7820X from throttling at 4.2Ghz on all cores.

 

[IndyColtsFan]

There seems to be a lack of solid reviews with an overclocked 7820k and reports on temps, gaming performance, etc. Has anyone seen one? Or even better, any reviews of a delidded 7900k with the TIM replaced with the replacement Silicon Lottery uses?

 

[Bassman2003]

@StefanR5R - thanks for your reply. It sounds like the 7900x is my best choice even if I can not overclock it as much as I might want. I am not building until mid August so I will keep my eye on things.

 

[StefanR5R]

There seems to be a lack of solid reviews with an overclocked 7820k and reports on temps, gaming performance, etc.

A brief review in German (no overclocking):
https://www.golem.de/news/core-i7-7...ls-8c-und-6c-konter-verpufft-1707-128772.html

 

[TheGiant]

There are too many factors involved to give you any meaningful advice. I suspect you will not be happy with tamz_msc's answer but it is really the best one can give you, given minimal information about your software. For example, here is just a list of factors that need to be taken into account:
.

Well, my current solver machine is Broadwell 14C Xeon with 256GB of ECC memory. I am not going to buy a new one just making research of what is important. However if some of the new products will be cheaper and comparable in performance I will sell this one, still has a nice value.

I am not even thinking about buying something without ECC for the Solver machine.

As you are saying not everything is well multithreaded.My desktop i5-6600K oced to 4.3GHz (with DDR4 3200) is sometimes significantly faster for some types of calculations.
For fortran I am using a intel compiler. I tried to test today some calculation to be optimized only for SSE2 and then for AVX2 and the AVX2 version was nearly 60% faster. So my optimization should be better to use as much as I can from current CPU.
I also tried to run the Xeon with only 2CH memory and for most calculations it was nearly 2 x slower- so the memory bandwitch seems very important.

SKL Xeon with 6CH DDR4 2667 should provide a very nice performance boost.

 

[formulav8]

Imagine current KBL performance with 6C at 5GHz, yay!

And a new motherboard. KL persons got robbed. Dead end platform after a few months. Maybe Intel will do something to fix that, especially since its the same socket?

 

[Ajay]

And a new motherboard. KL persons got robbed. Dead end platform after a few months. Maybe Intel will do something to fix that, especially since its the same socket?

Depends on what the Z390 brings and whether it will be compatible with Ice Lake - way to early to tell. The Z370 is just a tweaked Z270. Both of those are gifts to the motherboard makers to keep the market robust.

 

[Atari2600]

So what do you recommend?
Buy the SKL Xeon with 6channel? Or TR or EPYC even with 8ch?

Ask whoever you are getting your software off (Altair/Ansys/CD-adapco etc).

They will have data on the different architectures.

You may want to have a canary test* in your request email to ensure you are not talking to a sales idiot that knows f**k all.


*for example:

"We are looking at buying a new system for our CFD simulations, would <insert knowingly deficient/old configuration> this be the best for the job, or would you recommend something else instead?"


I know Ansys are, or at least were, part of AMD's development program, so should have good data on EPYC.

 

[Atari2600]

For example, the profiling may show that your code is dominated by a sparse linear solver and a CFD meshing algorithm. The former doesn't scale well with multiple threads (both direct sparse solvers and multigrid preconditioned solvers don't scale well past 4 cores).

Eh? I thought moving the sparse solver to the GPU was one of the more successful examples of scaling!

Example: https://link.springer.com/article/10.1007/s11227-012-0825-3

Meshing algorithms do scale well to any number of cores. In that case, you'll need both in which case SKL-X is a great choice because of its Turbo Max 3.

And in my previous experience, most time in the meshing software was limited to a few cores, despite some of the sub-processes scaling up rather well.

 

[TheF34RChannel]

Depends on what the Z390 brings and whether it will be compatible with Ice Lake - way to early to tell. The Z370 is just a tweaked Z270. Both of those are gifts to the motherboard makers to keep the market robust.

If Z370 is a tweaked Z270, then Z390 is as well since it only differs in the latter having integrated Wi-Fi and USB 3.1.

I do believe that both chipsets will be compatible with Ice Lake because Intel always does two chips per chipset. However, wasn't the 400 series listed for that CPU family?

 

[Topweasel]

If Z370 is a tweaked Z270, then Z390 is as well since it only differs in the latter having integrated Wi-Fi and USB 3.1.

I do believe that both chipsets will be compatible with Ice Lake because Intel always does two chips per chipset. However, wasn't the 400 series listed for that CPU family?

It's two chips and chipsets per socket on the Consumer line and 2 chips per socket/chipset on the HEDT line. So there would be a 400 series but that would in theory work with both Coffee Lake and Ice Lake.

 

[TheF34RChannel]

It's two chips and chipsets per socket on the Consumer line and 2 chips per socket/chipset on the HEDT line. So there would be a 400 series but that would in theory work with both Coffee Lake and Ice Lake.

Indeed. My suspicion arose because the 2nd chip on this socket was previously the now canceled Cannonlake desktop for the 400 series. That must have changed then.

 

[Bouowmx]

I do not think Intel 300 series chipsets will be compatible with Ice Lake, because it is a new architecture. Socket H chipsets looking like (rows indicating cross-compatible products):

• Lynnfield: 5 series
• Sandy Bridge, Ivy Bridge: 6, 7 series
• Haswell, Broadwell*: 8*, 9 series * Broadwell and 8 series are incompatible
• Skylake, Kaby Lake, Coffee Lake, Cannon Lake: 100, 200, 300 series
• Ice Lake: ?

See that different rows are not compatible with each other, and between each row is a Tock, or now, Architecture.