Coffeelake thread, benchmarks, reviews, input, everything.

[TahoeDust]

Nice find, thanks for the link! Perhaps the most useful bit, a piece of the conclusion, for those who may not have the time to read through the review:

"If you have a bit more money to invest, the next [after 2666] good option is 3200 MHz memory, with CL14 or CL15. 3200 MHz CL16 is roughly equal to 3000 MHz CL14 in speed, so consider that option too. Coffee Lake also sees good gains from improved timings, which often makes it more sensible to buy lower latency memory than to go for the highest clock speed you can find."

Interesting. I wonder how 3600MHz CL16 compares to 3200MHz CL14?

 

[StinkyPinky]

I am confused why the broader community suddenly decided to have conversation about Multi-core enhancement, now.., when it could have had it 5 years ago. Manufactured drama.

I was thinking the same thing. My z97 haswell board had this feature.

 

[PeterScott]

I was thinking the same thing. My z97 haswell board had this feature.

The problem this time was Asus turned on MCE by default, which lead to discrepancies in performance between reviews with different MBs.

 

[StinkyPinky]

The problem this time was Asus turned on MCE by default, which lead to discrepancies in performance between reviews with different MBs.

Quite frankly, that is a poor job on the reviewers part. They should be checking bios settings like that.

 

[TahoeDust]

Quite frankly, that is a poor job on the reviewers part. They should be checking bios settings like that.

I had the same thought. It's pretty irresponsible not to check how many cores are running and at what frequency before performing benchmarks you plan on publishing.

 

[TheF34RChannel]

Interesting. I wonder how 3600MHz CL16 compares to 3200MHz CL14?

It should be on par if not slightly better? I wonder about the same thing about my CL16 3333MHz. But yeah I feel better after the TPU article, 1% pff

 

[VirtualLarry]

The results are pretty much what you would expect. Below the top supported memory frequency (2666), performance plunges. It plunged way more than 3% in several benchmarks (such as 2133 was 10+% slower than 2666 in WinRAR Compress and CFD). Above that top supported 2666 frequency, the performance gains are measurable but quite small. Latency matters far more than high memory frequency in that region (above 2666).

I could tell the difference, just web browsing, between running my DDR4 at 2800 versus 2133, in my i3-8100 quad-core 3.6Ghz rig, with a PCI-E M.2 SSD. It was a bit sluggish, it just felt like, until I tweaked the RAM's XMP settings properly. (They were not automatically applied on my ASRock Z370 ITX/ac board.)

 

[dbrons]

It is Kabylake with 2 more cores that comes with almost no added power cost, that is awesome. When will this be available for purchase at recommended prices though?

Well I have a little duifferent take on this. I'm kinda glad I have to wait for my 8700k. I guess it will be a month or two but I'm using this time to figure out what other parts I want and just to refamilliarize myself with the process.

It's been almost 7 years since I built my 2600k system. I used to always wait a few months to read reviews and let some of the bugs come out. I had originally planned on building Kaby Lake but the reports I read made it seem like maybe, yet again, a new system not worth upgrading for.

So I've been waiting for over a year and I went ahead and bought a couple parts - a Corsair rmi650 PS An 8tb WD red hdd, and 8tb hgst 7200rpm hdd. when they were on sale. I pre-ordered an 8700k.
The PS and hard drives are pretty straight forward but lots of the rest of my build is taking some thought.. For instance, I am figuring on the define r5 as I love my r3 but I'm waiting on the chance an r6 may be released soon. Same thing with my SSD, Samsung is supposedly coming out with a 970 pro before years end. so I'll wait at least till my cpu gets here.

Oh and I mentioned that I was planning on the Asus Prime board and someoe called it a cheapie which I still disagree with, but I'll tell you, I took a look at the Maximus X Hero and really like the components and, don't laugh, the back plate Always seems weird to me to spend a lot of money on a system and have that flimsy metal back plate that I usually cut myself on, So the Hero it is.

The only things I haven't reallly decided on short of availability of new parts, is the RAM. and the aio cooler I would like to do some overclocking - not too severe, but hey, It should be a fun system to see what it can do. And the cooler, I think I'll buy after I have my case so I can try to measure and figure out where I will be putting things. (leaning towards a 240mm radiator up front) The RAM I don't know that much about I'm reading up here and elswhere - obviously my first time buying ddr4. I'm figuring 16 gb for my system that I use compiling a very large video and music collection. I use the onboard graphics I see 3200 speed(?)

Not sure if anyone is interested in my plans but if you've got any suggestions I'll listen. I'm not concerned about getting the cheapest parts as I'm hoping to have a pretty serious sytem that will process a 50gb video file in a lot less time than it takes me now,
Dave

 

[TheF34RChannel]

Well I have a little duifferent take on this. I'm kinda glad I have to wait for my 8700k. I guess it will be a month or two but I'm using this time to figure out what other parts I want and just to refamilliarize myself with the process.

It's been almost 7 years since I built my 2600k system. I used to always wait a few months to read reviews and let some of the bugs come out. I had originally planned on building Kaby Lake but the reports I read made it seem like maybe, yet again, a new system not worth upgrading for.

So I've been waiting for over a year and I went ahead and bought a couple parts - a Corsair rmi650 PS An 8tb WD red hdd, and 8tb hgst 7200rpm hdd. when they were on sale. I pre-ordered an 8700k.
The PS and hard drives are pretty straight forward but lots of the rest of my build is taking some thought.. For instance, I am figuring on the define r5 as I love my r3 but I'm waiting on the chance an r6 may be released soon. Same thing with my SSD, Samsung is supposedly coming out with a 970 pro before years end. so I'll wait at least till my cpu gets here.

Oh and I mentioned that I was planning on the Asus Prime board and someoe called it a cheapie which I still disagree with, but I'll tell you, I took a look at the Maximus X Hero and really like the components and, don't laugh, the back plate Always seems weird to me to spend a lot of money on a system and have that flimsy metal back plate that I usually cut myself on, So the Hero it is.

The only things I haven't reallly decided on short of availability of new parts, is the RAM. and the aio cooler I would like to do some overclocking - not too severe, but hey, It should be a fun system to see what it can do. And the cooler, I think I'll buy after I have my case so I can try to measure and figure out where I will be putting things. (leaning towards a 240mm radiator up front) The RAM I don't know that much about I'm reading up here and elswhere - obviously my first time buying ddr4. I'm figuring 16 gb for my system that I use compiling a very large video and music collection. I use the onboard graphics I see 3200 speed(?)

Not sure if anyone is interested in my plans but if you've got any suggestions I'll listen. I'm not concerned about getting the cheapest parts as I'm hoping to have a pretty serious sytem that will process a 50gb video file in a lot less time than it takes me now,
Dave

Gosh you keep taking what I said out of context; stop trying to provoke me, it won't work.

 

[dbrons]

I guess you are misunderstanding me. When I looked at the Hero board I realized you were right and I decided to take your advice.
Dave

 

[epsilon84]

I could tell the difference, just web browsing, between running my DDR4 at 2800 versus 2133, in my i3-8100 quad-core 3.6Ghz rig, with a PCI-E M.2 SSD. It was a bit sluggish, it just felt like, until I tweaked the RAM's XMP settings properly. (They were not automatically applied on my ASRock Z370 ITX/ac board.)

Wow, I didn't know web browsing is so memory bandwith dependant?!

I don't have an i3 8100 but my laptops i5 7300HQ is pretty close (max turbo 3.5GHz) which is paired with 16GB DDR4-2400 and I don't feel any sluggishness, I really wonder if there are other factors at play here rather apart from the CPU or RAM speed?

Did you run any benches between the different RAM speeds, also?

 

[PeterScott]

Wow, I didn't know web browsing is so memory bandwith dependant?!

I'm more inclined to believe it's placebo dependent.

 

[Dufus]

I was thinking the same thing. My z97 haswell board had this feature.

Haswell is a special case and to use MCE on locked chips required the use of ucode version less than 8 to set it.

Once people had bought their boards they might have found sometime later it was taken away with a BIOS update. No more firmware fixes from the manufacturers if those guys wanted to continue using MCE.

example
https://rog.asus.com/forum/showthread.php?36156-Turbo-of-4770-only-3700MHz-after-bios-update

 

[VirtualLarry]

I'm more inclined to believe it's placebo dependent.

Well, that could well be, and when I was comparing, my monitor was set to 30Hz, rather than 60Hz, because of course of Intel's iGPU limitations on the HDMI 1.4 port. (Limitations that NVidia were able to bypass, on GPUs as old as Kepler with HDMI1.4, and can run at 4K60 at reduced Chroma sub-sampling.)

But I though that my RAM was already running at 2800, as the UEFI main screen showed 2800 16-16-16-36 1.20V. But apparently, that wasn't active by default. It was bugging me enough (the slight lag during browsing and Windows), that I went back into the UEFI with a fine-toothed comb, and found out that it wasn't actually enabled.

So, I really don't think it's strictly placebo. Also, I've read some reviews and commentary around the web, starting with Skylake, that 2133 RAM "holds back" the CPU (Sky / Kaby / Coffee Lake) a bunch, whereas faster (2400 and up), does not. For whatever reason, I don't know exactly.

Somewhere in my recent posts, I compared the benchmarks between 2133 and 2800, there was a difference, however slight.

 

[Bouowmx]

Web browser responsiveness: forget memory bandwidth. Why not think something more obvious? Speed Shift is disabled and processor transitions slowly from low to high frequency. Software solution is use ThrottleStop to enable Speed Shift and set Energy-Performance Preference parameter very low, like "2", so that processor would shift to high frequency even under the lightest loads, such as moving mouse.

 

[Dufus]

@VirtualLarry did you check if BCLK Aware Adaptive Voltage was enabled and if enabled then does disabling it help with achieving a bclk OC and top ratio?

Just a thought.

 

[USER8000]

Quite frankly, that is a poor job on the reviewers part. They should be checking bios settings like that.

Did you even bother to read what Jay2cents said in the comments section of the Adoredtv video??

FOR THE RECORD.... after we were called out for testing with default board settings (by the audience by the way, not you) ASUS reached out to me to explain that MCE is OFF by default and Sync all Cores should NOT have been enabled by default, I explained to them that this is indeed FALSE. ASUS themselves aren't even clear on what the default settings were and it was only after clearing CMOS again and showing them what the optimized defaults are did they agree that they need to reel in their BIOS team and get to the bottom of this... I also said in my video that we as reviewers need to be better at this, so although you calling me out in this video is accurate, I had already updated my content showing where the discrepancy was. Please dont be one of those channels who builds their rep on the "Im more righteous than you..." we have enough of those channels already. As for the "Intel hiding something" of course they are... they are widening their stack on purpose... It sucks, but its business and it doesnt take much investigative research to figure that out.
Show less
REPLY748

So Asus was still keeping it active if you switched it off.

 

[DaveSimmons]

Did you even bother to read what Jay2cents said in the comments section of the Adoredtv video??

So Asus was still keeping it active if you switched it off.

That's another reason for me to stick with ASRock, along with my 5-6 year-old mobo for the i5-2500 being problem free. I really don't want my motherboard to overclock without telling me just so they can win at benchmarking.

 

[TahoeDust]

So Asus was still keeping it active if you switched it off.

That is not what I understood from reading his comments...I have re-read it three times. Asus was keeping it active when setting it to default. He was still able to turn it off.

 

[DrMrLordX]

Interesting. I wonder how 3600MHz CL16 compares to 3200MHz CL14?

Well it depends on the task. For anything that taxes all 6 cores of the 8700k, I would think you would see some impressive performance improvement all the way up to around DDR4-3466 <--> DDR4-3733, somewhere in that department. You aren't going to see that in a whole lot of "real world" applications, but you might see it in some HPC-ish stuff like y-cruncher that is affected by memory performance. You won't see it on SuperPi.

Anyway from DDR4-3733 on up, you are going to see the same performance gain from increasing clockspeed by 266 MHz (equivalent) or decreasing CAS/CL by 1 (assuming you also get to decrease other timings/subtimings in step). So DDR4-3733 CAS/CL 17 is going to be about the same as DDR4-4000 CAS/CL 18.

For anything that does not fully utilize all the cores, you will see the same basic situation in Skylake/Kabylake: at any speed DDR4-2400 and higher, +266 MHz (equiv) = -1 CAS/CL. So DDR4-3200 CAS/CL14 is going to be about the same as DDR4-3733 CAS/CL 16 in those scenarios.

DDR4-3600 CAS/CL16 should be a little slower.

 

[deathBOB]

Did you even bother to read what Jay2cents said in the comments section of the Adoredtv video??



So Asus was still keeping it active if you switched it off.

How did the dude not spot that in CPUZ or whatever application that shows core clocks? Wouldn't you at least take a look at that while benching a new set of hardware?

 

[PingSpike]

Did you both read the full 11 pages, or just look at the few gaming graphs?

The results are pretty much what you would expect. Below the top supported memory frequency (2666), performance plunges. It plunged way more than 3% in several benchmarks (such as 2133 was 10+% slower than 2666 in WinRAR Compress and CFD). Above that top supported 2666 frequency, the performance gains are measurable but quite small. Latency matters far more than high memory frequency in that region (above 2666).

I was going to ask you if you read the pages but then I read your post so here is my response:

We're in agreement then. It barely matters. I swear I've seen other benchmarks that suggested in games it made a difference in minimum frame rates but that isn't tested here. If anything, this review suggested to me that memory frequency matters far less than I initially thought it did.

 

[dullard]

We're in agreement then. It barely matters. I swear I've seen other benchmarks that suggested in games it made a difference in minimum frame rates but that isn't tested here. If anything, this review suggested to me that memory frequency matters far less than I initially thought it did.

Memory frequency matters a lot, until it doesn't. After about 2666 with the Coffee Lake processor, you have reached the point of diminishing returns on frequency. It is better to invest in lower latency at that point than to invest in higher frequency. But, below 2666 (like in Virtual Larry's 2133 web browsing example), memory frequency matters a whole lot.

I do not have minimum frame rate data.

 

[piesquared]

The paper launch was a bad move on Intel's part. But, you can't honestly come in here and say up to 50% more processing capability (6 cores vs 4) with just 4.4% more power used on average (95 W vs 91 W) is not a spectacular engineering achievement. Especially, considering that it was based on a fairly power efficient mobile-based core to begin with.

Yes i can, because it's true. The difference is in the process like i said. Coffeelake is just 2 cores added on to the 4 existing Kabylake cores, i don't consider that a spectacular engineering achievement at all. It's and engineering effort yes, but certainly not spectacular achievement. If it was a 50% IPC improvement, that would be spectacular. Adding 50% more cores doesn't translate into 50% more processing capability.
The TDP doesn't align with the power draw of the chip, that is special condition number that intel seems to have tagged on for marketing purposes. Is it 50% more power for 50% more cores? No it isn't, because the process improved.
This is a pivotal moment in the industry. In some respects, intel has now lost that process lead. A physics wall is fast approaching and any remaining lead is going to diminish to nearly nothing. This is where the rubber is going to meet the road and who has the best architecture for scalability. I think intel has unfortunately relied too long on it's process for it's dominance in the industry and forward thinking on design has taken a back seat.

To the poster that claimed Coffeelake wasn't a paper launch, it is the exact definition of one, not sure how you can say otherwise. This is a trickle of product to the market in ultra low volume.

 

[piesquared]

Thats a very unfair comment. Coffeelake is a spectacular engineering achievement. Intel's 14++ process offers the highest transistor performance and CPU chip frequency in the semiconductor industry. AMD would love for GF to deliver a similar process node but thats not going to happen till GF 7LP arrives in 2019. AMD cannot match Coffeelake in ST and gaming perf even if they achieve clock parity as Zen has >= 10% lower IPC than Coffeelake. In memory latency sensitive workloads like games the per clock difference is surely above 10%. I think 8700k will hold up very well even against Pinnacle Ridge for MT perf. There are architectural/design choices that AMD has made (CCX and IF) to be able to scale Zen core from 4 core notebooks to 32 core servers(with multi die MCM). Intel are going to adopt a multi die with EMIB CPU design sometime in 2020 timeframe when they might have to address the problems seen with Zen like design. But until then their single die ring bus based approach works extremely well for client computers. Lets give credit where its due.

Yes, so what it boils down to is a 10% IPC lead, which subsequent versions of Zen should eliminate, and a process that clocks higher. So at clock parity, the 8700K would have a 10% lead in stricly single threaded applications (of which there are very few) and fall well short in performance for multi threaded applications compared to the uncut 8 core Ryzen chip at the same price point. As i said in the other post, yes it is an engineering effort but not spectacular. The process allows it to clock higher.