Makentox :
Cache frequency has very little impact on CPU performance, but can have a noticeable effect at higher memory frequencies by improving latency results. This why it took less time in prime. However fps gain in games is 0. Ive never heard someone setting uncore above core. It usually brings instability when i set it to = core and sadly 0 perfomance gain for my everyday tasks. Also do u get your magic boost in all games or which? What ur ram mhz/timing? Your system spec?
I have been running my 6600K for 18 months at 4.6 GHz after manually setting VCore at 1.380 V in the BIOS. With this, I set the Uncore Ratio (cache) to "49" (or 4.9 GHz) - yes, much higher than the Core ratio! It is mighty stable under anything Prime95 can throw at it and for as long as I want...
Here are details of my current, and ultra stable/safe, BIOS settings for i5 6600K on Gigabyte GA-Z170XP-SLI with 16 GB of RAM (4 x 4):
[1] - CPU Core Ratio:
46
[2] - FCLK Frequency For Early Power:
1 GHz
[3] - Uncore Ratio:
49
[4] - CPU Flex Override:
Disabled
[5] - Intel Turbo Boost Technology:
Disabled
[6] - CPU Enhanced Halt (C1E):
Disabled
[7] - C6/C7 State Support:
Disabled
[8] - C8 State Support:
Disabled
[9] - CPU Thermal Monitor:
Enabled
[10]-CPU EIST Function:
Enabled
[11]-Voltage Optimization:
Enabled
[12]-Residency State Registration (RSR):
Disabled
[13]-Hardware Prefetcher:
Enabled
[14]-Adjacent Cache Line Prefetch:
Enabled
[15]-Extreme Memory Profile (X.M.P.):
Profile 1
[16]-System Memory Multiplier:
32
[17]-Memory Enhancement Settings:
Relax OC
[18]-Channel Interleaving:
Enabled
[19]-Rank Interleaving:
Enabled
[20]-CAS Latency:
15
[21]-tRCD:
17
[22]-tRP:
17
[23]-tRAS:
28
[24]-Command Rate (tCMD):
1
[25]-CPU VCore Loadline Calibration (LLC):
High
[26]-CPU VCore:
1.380V
[27]-CPU VCCIO:
Auto
[28]-CPU System Agent Voltage:
Auto
[29]-PCH Core:
Auto
[30]-DRAM Voltage (CH A/B):
1.360 V
[31]-Internal Graphics:
Disabled
Now then, let's address your comments:
<<Cache frequency has very little impact on CPU performance>> true for values below the Core freq. since the Uncore remains just a bottleneck for the CPU. But once you get true stability with a suitable VCore (1.380 V in my case) at Core/Uncore parity, e.g. 46/46, then increasing the Uncore freq. will start having a great impact on CPU performance (I get 14.5 % !!!) and without affecting stability. Stop increasing Uncore once you detect that you have reached the peak (even go back a little if required). Again, you detect that by measuring the time the PC takes to perform the exact same run (test), like the Prime95 "In-Place large FFTs test", in the exact same conditions. And repeat the measurements to be absolutely sure...
<<a noticeable effect at higher memory frequencies by improving latency results>> no, NOT "noticeable", just fractions of a single percentage point, particularly once you already have memory at 3200 MHz. Beyond that it becomes nil.
<<This why it took less time in prime.>> I also used the Prime95 "In-Place large FFTs test" to measure how long the system takes to complete the 128 K test on the four workers (around 8 min. 38 sec. with Uncore 49). This test uses very little memory (they say "some RAM tested"). UnCore 50 took 8 min. 39 sec., and UnCore 51 was taking 8 min. 59 sec. indicating that the peak had already been reached.
<<Ive never heard someone setting uncore above core.>> well, now you have. Try it yourself rather than following the herd mentality. You can only get the immense performance gains that I am getting.
<<It usually brings instability>> yes, and I had a BSOD with Uncore 47, Core 46 and VCore 1.370. BUT once I found the stable VCore of 1.380 V, my system has become more stable than it has ever been. I now throw anything at it and it never fails.
<<sadly 0 perfomance gain for my everyday tasks>> that's why it's time for you to change and go the other way, way over the Core value. I use my 14.5 % performance gain in my everyday tasks with great benefits, because this is "real" computational performance, not some rendering/fps gain in some game. By the way, I get in all the games because I play them with my two GTX 970 cards mounted in SLI which requires a fair bit of additional CPU work. Once again, this is not some specific game rendering gimmick (polygon count...) , this is real processing performance at the PC level by no longer having the Uncore as a bottleneck to its CPU but really assisting the CPU.
In the BIOS of my Gigabyte GA-Z170XP-SLI motherboard, when I click on "Uncore" it indicates something like:
"In overclocking configurations, it is better to set the Uncore ratio at or above the Core ratio". I really can't take a picture now because I would have to reboot and go into the BIOS of course. I will try that the next time I reboot to show you. The Gigabyte engineers are not completely stupid IMO and surely would be careful when printing such a widely distributed statement to prevent being taken for ignorant fools. Admittedly I too was surprised when reading it at first because it goes against the herd mentality. But then I thought "what if"? What if these guys were not complete idiots. And so I tried from parity (46/46) and kept increasing the Uncore ratio while measuring at every new value the time my PC took to accomplish the exact same test run (and more than once each time). From Uncore 46, then 47, then 48 I had already shaved off over 6 minutes on a total of 48 minutes. At Uncore 49 I was 7 minutes faster than my baseline which is the same test performed in exactly the same conditions (except that the ambient temperature was 0.3 degree Centigrade lower) at an Uncore ratio of 38 (46/38). I had the Uncore at 38 with the Core at 46 to allow me to run at VCore 1.345 V fully stable. I was also running at 46/38 under VCore 1.335 V but then only partially stable (30 min. of Prime95 Blend test before one worker started to fail).
Besides, when now running the Prime95 Blend test for a long period of time, my max Temp periodically reaches 86 C for a few seconds (at over 26 C ambient) and the power tops at 104.50 W. My CPU fan goes up to 1745 RPM Max. for a very short while. Most of the time, my Temp. stays under 70 C, even now at 27.2 C ambient. Let us remember that the Temp values are standard at 22 C ambient. So at 86 C with 26 C ambient, I am still way under 85 C with 22 C ambient which is fine. My CPU fan (SYS 1) is designed to go up to 2000 RPM, so it still has enough left over (I have seen it reach 1989 RPMs when testing on 4.7 GHz - too high).
This is a true 14.5 % performance gain on real (practical) "PC runtime", not some fictitious fractions of percentages over a colourful graph from TweakTown. However, it is not for me to take all the credit. I really thank the Gigabyte engineers and/or designers for having placed that little note as a help for overclockers in the BIOS... and will try to publish a picture of it from my screen. I looked in the printed/online doc. but can't see any mention of it (not surprised, many other things from the BIOS are missing from the doc).
So please, look into this and try it thoroughly, it is really worth the effort... would it be just to make sure.
OK, I just got the camera out, rebooted, and took the following pictures:
(Note also my 46 and 49 ratios)
And here is a new Prime95 Blend test result from CPUID HWMonitor of 41 minutes that I ran under a slightly higher ambient temp. (up to and including the completion of the 800K of the Blend test for the 4 workers):
Note: On the last pic I indicate with my finger the word "
higher" in case it were missed (this is the most important word in this text after all). This one replaces a previous pic showing my middle (longest) finger and parts of my hand, but pointing up that way could have been misconstrued ;-).
So here it is, you have the exact wording. Nowhere does it even suggest that it should be anything LOWER than the Core ratio. Yet almost everybody has it around 400 MHz lower (?). Why? ... fashionable he... well, all MISGUIDED guys. Besides, when it mentions
"Note: In overclocked configurations it is recommended to set the Uncore ratio equal to or higher than CPU clock ratio.", it doesn't indicate by how much it is recommended to set it
HIGHER. I went all the way from parity 46/46 (that's the "equal" part) to 46/49 (that's the "higher" part) and stopped there before performance starts dropping. Et voila!
As I wrote above, are these Gigabyte guys completely bonkers then? Hmmm... didn't think so. But maybe it's just me.