Kaby Lake Overclocking Revealed: Simultaneously impressive AND disappointing

[Prysin]

3 hours ago, patrickjp93 said:

At the rate of consumer code optimization, Sandy Bridge will have its own second coming when AVX starts getting used in games. Just see my SIMD Game Engine code blog here on LTT.

 

Whatever CPUs start getting big bandwidth upgrades at the same latencies as now will also have a large impact beyond that.

nah, AVX would be saving Sandy if its mem controller wasnt as poor as it is. Same for Bulldozer/Piledriver. If their memory controllers werent weak they too would fare better with AVX. If Sandy and FX had DDR4, they would be much more relevant even today (just look at bristol Ridge. Sandy IPC with FX clocks and DDR4, it is just barely matching ULV skylake i5s.... that is impressive for something that old, but that is also the limit)

 

Thing is, as your own blog suggests, memory bandwidth is too low with DDR3 to ever be able to EFFICIENTLY use AVX, at 256bit, it would be fast, i bet even at 128bit it would be faster then a lot of planar code solutions out there. But even at 128bit, you would risk hitting the upper limits of DDR3 bandwidth  and stability (not to mention the cost to get higher then 2400MHz DDR3 RAM these days, and a mobo with clean enough memory power delivery to maintain such clocks)

 

40 minutes ago, Lays said:

 

I ran p95 at 4.8 1.375v for an hour, no crash, bumped up to 5g 1.43, ran for 30 mins and didn't crash, then ran realbench for like 1hr 30 min I think, then GSAT on Linux for 1 hour.   I've been running 1.425 or 1.43 in bios ever since for like 4 months and haven't BSOD once, so I'm just going to assume I'm good lol   All i really do on my PC is game/netflix/ watch movies and listen to music lol

kinda what i do with my FX... although i was most impressed when i managed 2 hours P95 Large FFTs at 4.77 GHz (Bclk 201, multiplier 23.5, 1.475v. Memory at 2400Mhz XMP dual channel with +0.250v offset to the NB to help stabilize the memory <- yes for FX, bumping NB voltage helps memory OC. I know its not supposed to work that way, but it does) on all 8 cores, and only after 2 hours did two of them fail. Mind you, i was cooling it with a 92mm Noctua NH-D9L air cooler with 2x NF-F9 PWM fans...

 

 

[patrickjp93]

2 minutes ago, Prysin said:

nah, AVX would be saving Sandy if its mem controller wasnt as poor as it is. Same for Bulldozer/Piledriver. If their memory controllers werent weak they too would fare better with AVX. If Sandy and FX had DDR4, they would be much more relevant even today (just look at bristol Ridge. Sandy IPC with FX clocks and DDR4, it is just barely matching ULV skylake i5s.... that is impressive for something that old, but that is also the limit)

 

Thing is, as your own blog suggests, memory bandwidth is too low with DDR3 to ever be able to EFFICIENTLY use AVX, at 256bit, it would be fast, i bet even at 128bit it would be faster then a lot of planar code solutions out there. But even at 128bit, you would risk hitting the upper limits of DDR3 bandwidth  and stability (not to mention the cost to get higher then 2400MHz DDR3 RAM these days, and a mobo with clean enough memory power delivery to maintain such clocks)

 

kinda what i do with my FX... although i was most impressed when i managed 2 hours P95 Large FFTs at 4.77 GHz (Bclk 201, multiplier 23.5, 1.475v. Memory at 2400Mhz XMP dual channel with +0.250v offset to the NB to help stabilize the memory <- yes for FX, bumping NB voltage helps memory OC. I know its not supposed to work that way, but it does) on all 8 cores, and only after 2 hours did two of them fail. Mind you, i was cooling it with a 92mm Noctua NH-D9L air cooler with 2x NF-F9 PWM fans...

 

 

The bandwidth is too low to use AVX the way I did in isolation. I will be posting another page soon. I'm doing some testing now.

[Lays]

30 minutes ago, MageTank said:

Wow. I STILL have to do this on my ASrock Z170 Fatality ITX/AC. You make me regret going ITX x.x. Half of my profiles are basically primer profiles, lol. 

Here's what I run on that profile, it boots everytime

 

 

 

 

[Prysin]

5 minutes ago, patrickjp93 said:

The bandwidth is too low to use AVX the way I did in isolation. I will be posting another page soon. I'm doing some testing now.

DDR4 4266 should handle 128bit, that much we do know. However DDR3 wouldnt have a chance. Even at 3000, the latency of the memory itself would get so high that unless you were a damn good memory OCer, you should see diminishing return just from latency added to the fetch/store loops

[patrickjp93]

2 minutes ago, Lays said:

Here's what I run on that profile, it boots everytime

 

 

Mega SNIP

 

Dude, put those in a spoiler please.

[patrickjp93]

2 minutes ago, Prysin said:

DDR4 4266 should handle 128bit, that much we do know. However DDR3 wouldnt have a chance. Even at 3000, the latency of the memory itself would get so high that unless you were a damn good memory OCer, you should see diminishing return just from latency added to the fetch/store loops

The story is not over. We will get back to the point we can breathe again.

[Lays]

1 minute ago, patrickjp93 said:

Dude, put those in a spoiler please.

forgot those dam things even existed, haven't used or needed to use one in like 5000 years lol

[Prysin]

2 minutes ago, patrickjp93 said:

Dude, put those in a spoiler please.

oh, so when you brag about shit its ok, but when others does it they have to be considerate about it?

[patrickjp93]

Just now, Prysin said:

oh, so when you brag about shit its ok, but when others does it they have to be considerate about it?

I put my long-winded stuff in blog pages on my own area of the site. And it takes a lot of explaining to get ideas through to some people. I worked as a TA for 2.5 years in university. I know how learning works.

[Prysin]

5 minutes ago, patrickjp93 said:

The story is not over. We will get back to the point we can breathe again.

you planning to look into 128bit AVX? or is it not possible to go below 256 with conventional code (aka, needing a custom machine code compiler to make it happen)

Also, question if you know:
Does AVX allow for what we in GPGPU call "Half precision", i know AVX is based on 8bit increments, so technically its sub half-precision as is, but then again, GPU "single precision" is 32bits... mainly due to core count i suppose

[patrickjp93]

4 minutes ago, Prysin said:

you planning to look into 128bit AVX? or is it not possible to go below 256 with conventional code (aka, needing a custom machine code compiler to make it happen)

Also, question if you know:
Does AVX allow for what we in GPGPU call "Half precision", i know AVX is based on 8bit increments, so technically its sub half-precision as is, but then again, GPU "single precision" is 32bits... mainly due to core count i suppose

You mean SSE1/2/3/4.1/4.2 lol? You can. Literally change the 256 in all of the intrinsics I call to 128, and then make the pointer stride 4 instead of 8.

 

You might need AVX2, but one of those two ISAs does have half precision.

[Cryptonite]

12 hours ago, Dabombinable said:

Its could be that the FIVR in Haswell degraded above 1.35V. Though its possibly just Intel's 22nm process, since Ivybridge had a safe maximum of 1.2V

wait what? Ivybridge max 1.2?? some chips run at 1.2 on stock clocks (badly binned but still) it's 1.35 for Ivy (provided temps allow it). some will say more however. My bro is running at 1.22 volts atm and his pc isn't even overclocked. (H series mobo max turbo boost to 3.8 GHz i5 3570K)

[JoshM813]

http://www.tweaktown.com/news/54528/core-i7-7700k-hits-6-7ghz-ln2-i5-7600k-5-1ghz-air/index.html

 

6.7GHz on the 7700k with LN2 cooling and 5.1GHz on the 7600k on air.

 

Voltage on 7600k seems a bit scary though. 

 

 

 

...well crap just realized this was posted in the news section of the forums haha

 

 

[Maihym]

Damn, I'll take the 7700k, any idea how much it's going for? Nevermind I read the article and they say it's going to be around 370.

[iRileyx]

they should of done the i7 with air/watercooling and shown the overclock for that. not like everybody uses LN2 cooling so very few people will care for that. (I'm aware they were doing it for the sake of it, but it still woulda been nice to see it with a "everyday joe" kinda set up")

[WkdPaul]

1 hour ago, JoshM813 said:

...well crap just realized this was posted in the news section of the forums haha

no problem, merged

[lukart]

I wonder if they used a different mobo, I mean OC-Formula is pretty much the very best you can have for overclocking...

[MEC-777]

LN2 OCing is useless to the mass majority. Tell me what it can do on air, safe and stable, and show me the performance improvements from Skylake. That's what really matters, IMO. 

 

I am neither impressed nor disappointed by LN2 numbers. They're just meaningless. 

[MageTank]

9 hours ago, Lays said:

Here's what I run on that profile, it boots everytime

 

 

  Reveal hidden contents

 

 

 








 

 

 

 

 

Man, those OCF boards give you far more voltage options than these average "gamer" boards. DRAM boot voltage looks extremely appealing. 

 

9 hours ago, Dackzy said:

unless you run SLI/CF, m.2 NVMe, wifi card, sound card and other things

Sadly, you wouldn't want to run all of that on a PLX board. PLX, when using the same devices, acts as a multiplexer. When using different devices, it acts as a scheduler. The amount of added latency you get when running storage devices on top of GPU's is horrendous. This is why people try to avoid running mixed-mode multiplexing. The PCIE lanes wired to the CPU (not the DMI ones) are completely wired to the PLX bridge. Remaining DMI lanes are still wired to southbridge.

 

This PEX8747 datasheet might better explain it, as I am having a difficult time describing how it works. http://www.avagotech.com/docs/12351854

 

I also know that, depending on how the board itself is wired, you can get additional performance degradation when trying to use PCIE M.2 devices AND PLX, if the M.2 slot is wired to the CPU instead of DMI (Sata Express port). 

[Dackzy]

5 minutes ago, MageTank said:

SNIP

makes sense.

[vorticalbox]

20 hours ago, Ryan_Vickers said:

So for haswell the commonly accepted safe voltage was 1.35, and then for skylake it strangely went up... had that happened again here?  My gut would tell me that it should be the same as skylake but I haven't heard anything yet.

amd chips normally use 1.3 stock with most chips using ~1.5v for 5GHz.

[mr.squishy]

Being an ASRock Z170m OC Formula owner, this pleases me, but I thought Kaby Lake was going to require a different chipset than Skylake?

[MageTank]

1 minute ago, mr.squishy said:

Being an ASRock Z170m OC Formula, this pleases me, but I thought Kaby Lake was going to require a different chipset than Skylake?

Require? No. However, Z270 offers additional features not present on Z170, such as native thunderbolt 3, 4 additional PCIE lanes (24 vs the 20 on Z170) and Optane support just to name a few. 

 

A BIOS update is all that is required in order to support Kaby CPU's on current Z170 boards.

[Blackie Sheen]

Even the result on air is not even something you would do on a daily basis. 5,1 GHz is very impressive and all, but running on 127,51 is just not something you would do if you want your pc to be truly stable. I want to see how high a mutiplier-only based overclock would fare against previous gens. Not this.

[MageTank]

31 minutes ago, Blackie Sheen said:

Even the result on air is not even something you would do on a daily basis. 5,1 GHz is very impressive and all, but running on 127,51 is just not something you would do if you want your pc to be truly stable. I want to see how high a mutiplier-only based overclock would fare against previous gens. Not this.

BCLK in and of itself doesn't make something unstable. 127BCLK is perfectly fine. If you have memory that is at a stable strap at that BCLK, then it's no different from a multiplier overclock. 

 

What makes the 5.1 result impressive is the fact that it is on air, an unspecified air cooler at that. Potentially, a custom loop could get that voltage down quite a bit. We already know Skylake is safe at 1.4v if thermals are taken care of, so Kaby should be the same in that regard.