The Overclocking Thread

[Phuncz]

It's been a while for me and where it was relatively "simple" in the Pentium 4 and Athlon XP era to overclock, a lot of overclocking possibilities have become available the last few years. CPUs are more involved with on-board voltage regulation, but also refocused overclocking variables and motherboard manufacturers take it more seriously.

Vccin, LLC, PLL, cache voltage, Vdroop, System Agent voltage, BLCK, CPU ring/uncore, etc.
These are all very vague terms for the uneducated overclockers and I also don't know exactly what they do (just a basic understanding). It might even be stated that overclocking doesn't fit with SFF, I've stated that more than once when a 450W, K-series CPU and high-end GPU were stuffed inside an Ncase M1.

This thread isn't meant as my personal OC thread, I hope others will join with their amazing or disappointing endeavours. I'll be kicking off this thread with my own project in the post below.

 

[Phuncz]

My scavenged OC system:

• Intel Pentium G3258 (3.2GHz, 1.01V)
• AsRock Z87E-ITX (v2.50)
• Noctua NH-L12 with single Scythe 120mm PWM fan
• Crucial Ballistix VLP 2x 4GB PC3-12800 (9-9-9-24)
• WD Green 1TB SATA HDD
• Sigmatek Nebula C with front and side panels removed, 1x 120mm Scythe PWM fan
• Fans set at "performance mode"
• Windows 7 x64 - basic install (including all drivers
• AIDA64 for checking basic stability/temps/volts

Some basic settings I've made to start:
CPU Ratio - All Core (so both cores get the same ratio)
BCLK - 100.0
Spread Spectrum - Disabled
Intel SpeedStep - Disabled
RAM - XMP Profile 1 (DDR3-1600 9-9-9-24 1.35V)
CPU Vcove Voltage Mode - Override (fixed voltage)

First step: Probing.

Temps are average CPU after 3 minutes of AIDA stability test. I'll be going from the default 1.01 to 1.30V, over 1.40V isn't recommended and I'd like to keep some options for the end stretch.

Fixed CPU voltage at 1.01V with 1.70V input voltage

1. 32x100 at 1.01V --- ~38°C (default clocks)
   
2. 36x100 at 1.01V --- ~43°C (3.6 GHz)
3. 38x100 at 1.01V --- ~44°C (3.8 GHz)
   
4. 40x100 at 1.01V --- ~45°C (4.0 GHz - 25% overclock)
5. 42x100 at 1.01V --- No boot/POST

Fixed CPU voltage at 1.10V with 1.90V input voltage

1. 42x100 at 1.10V --- ~49°C (4.2 GHz)
2. 44x100 at 1.10V --- No boot/POST

Fixed CPU voltage at 1.20V with 1.90V input voltage

1. 44x100 at 1.20V --- ~54°C (4.4 GHz)
   
2. 45x100 at 1.20V --- ~57°C (4.5 GHz - 41% overclock)
3. 46x100 at 1.20V --- Unstable

Fixed CPU voltage at 1.25V with 1.90V input voltage

1. 46x100 at 1.25V --- ~64°C (4.6 GHz)
2. 47x100 at 1.25V --- Unstable

Fixed CPU voltage at 1.30V with 1.90V input voltage

1. 47x100 at 1.30V --- ~68°C (4.7 GHz - 47% overclock)
   
2. 48x100 at 1.30V --- No boot/POST

Noteworthy: idle temps went from 28°C to 30°C going from 1.01V to 1.30V, didn't expect that with SpeedStep disabled ! C-states is still active so it must be responsible for this.

I'll update this post until I reached the limit with 1.30V for the CPU and without touching the other voltages or settings. EDIT: done, next !

 

[Phuncz]

From the previous post we have established that 4.7 GHz is workable. Now lets get it stable. At this point I'm very happy, I'm almost at a 50% overclock while still playing it safe, maybe we can even reach 4.8 GHz stable with 1.3V using specific settings. So let's get to it:

Second step: Stabilizing.

47x100 at 1.30V was stable "at first sight" but it might as well tip over when we unleash something like Prime95 or Intel's Burnin Tool.
Since a CPU is more than just a Vcore, multiplier and base clock, lets look at the other tasty settings we have at our disposal:

• CPU Cache Ratio (if the cache is the issue)
• Filter PLL Frequency
• Internal PLL Overvoltage
• PCIE PLL Selection (should be set to SB PLL when OCing)
• CPU Cache Voltage (referenced as Uncore)
• System Agent Voltage
• CPU analog IO Voltage Offset
• CPU digital IO Voltage Offset
• FIVR Efficiency Mode
  
• CPU Load-Line Calibration

I'm going to start setting the options that seem to be important for OC stability. But first let me test how stable the current 4.7 GHz is without them, using Prime95, monitoring using AIDA64. Then I'll gradually do settings to achieve (hopefully) stability:

1. All settings at default settings --- Crash within a few seconds
   
2. Setting PLL Freq Filter, PCIe PLL Selec and FIVR Efficiency to their "recommended for OC" settings --- Crash after ~10 seconds
   
3. Also decreasing CPU Cache Ratio to 32x (default clock) --- 75-80°C - Passed: 640K, 8K, BSOD
4. Also enabling Internal PLL Overvoltage ---75-80°C - Passed: 640K, 8K, 720K,12K, 800K, 16K, 960K, 24K, 1120K, 32K, 1200K, 48K, 1344K, 64K, 1536K, 80K, 1680K, 96K <-- at this point I'm an hour in and I consider it stable enough to move on.

So I've determined potential issues with stability seem to be solved with the PLL Overvoltage setting, but I still want the CPU Cache Ratio to be the same as the clock speed. So I'm going to try to achieve that:

1. Increase CPU Cache Ratio to 40x --- 75-80°C - Passed: ~1hour, moving on
2. Increase CPU Cache Ratio to 44x --- 75-80°C - Passed: 640K, 8K, 720K,12K, 800K, 16K, 960K, 24K, 1120K, 32K, BSOD

Clearly the cache seems to be the problem. Lets try upping the voltage on that one.

1. Increase CPU Cache Voltage to 1.20V, Ratio 44x -- 75-80°C - Passed: ~20 minutes, freeze
   
2. Increase CPU Cache Voltage to 1.30V, Ratio 44x -- 75-80°C - Passed: ~10 minutes, reboot

That isn't helping at all, so it must be something else that's over it's limit. Lets try setting the Cache Voltage back to Auto and upping System Agent Voltage, CPU analog IO Voltage Offset and CPU digital IO Voltage Offset by 0.20V each:

1. Increased Agent, analog and digital Voltage with 0.20V, Ratio 44x --- 75-80°C - Passed: ~1hour, moving on
2. Increased Agent, analog and digital Voltage with 0.20V, Ratio 47x --- Crash after ~10 seconds
3. Also applied CPU Cache Voltage to 1.30V, Ratio 47x --- Crash after ~5 seconds

So the Cache is being a problem at 4.7 GHz, so let's see if it's stable at 4.5 GHz with Cache Voltage set to Auto:

1. Increased Agent, analog and digital Voltage with 0.20V, Ratio 45x --- 75-80°C - Passed: ~1hour, moving on

At this point I'm satisfied with the cache at 4.5 GHz. Maybe I can try getting 4.8 GHz on the cores to work ?

Noteworthy 2: the CPU cooler is barely above room-temperature but temperatures seem to react fast enough to be sure it's doing its job. If temperature would be the problem, I would have to delid the CPU as it seems the heat at load stays trapped at the silicon.

 

[Phuncz]

Now that 4.7 GHz is stable and I've determined the limit out of the cache, it is time to nudge it further. This will be tiny steps, but with a 50% overclock at arm's length, this is hopefully going to be a fruitful endeavor.

Let's recap what we've set so far to reach the 4.7 GHz:

• CPU Ratio - All Core
• BCLK - 100.0
• Spread Spectrum - Disabled
• Intel SpeedStep - Disabled
• RAM - XMP Profile 1 (DDR3-1600 9-9-9-24 1.35V)
• CPU Vcove Voltage Mode - Override
• CPU Voltage Fixed - 1.30V
• FIVR Voltage Fixed - 1.90V (input voltage)
  
• CPU Cache Ratio - 45x
  
• Filter PLL Frequency - High BCLK Mode
  
• Internal PLL Overvoltage - Enabled
  
• PCIE PLL Selection - SB PLL
  
• System Agent Voltage Offset - 0.200V
  
• CPU analog IO Voltage Offset - 0.200V
  
• CPU digital IO Voltage Offset - 0.200V
  
• FIVR Efficiency Mode - Disabled

This still leaves some options. I have yet to touch CPU Load-Line Calibration because I need to look into it to find out how this works on this particular board. I can also adjust the BCLK to improve overclocks but it involves a lot of other aspects like the RAM which in my build is not meant for overclocking.

For now I'm going to try to focus on achieving the highest Core clock I can, beginning at 4.8 GHz.
Remember in the earlier test: 48x100 at 1.30V --- No boot/POST

Currently it's purring at that setting 15+ minutes in AIDA 64's System Stability Test, before I unleash the relentless Prime95 onto it. But as you may have noticed, this means that 4.8 GHz is possible to achieve, even though it seemed not to be at all possible when just adjusting the Core Voltage.

Third step: Nudging

1. 48x100 at 1.30V and Cache Ratio x45 --- 75-80°C - few minutes of Prime95, BSOD
2. 48x100 at 1.325V and Cache Ratio x45 --- 77-82°C - Passed: 640K, BSOD
3. 48x100 at 1.325V and Cache Ratio x44 --- 77-82°C - Passed: 640K, BSOD

At this point I'm not seeing much stability with 4.8 GHz so it might as well be that we've reached the limit with air-cooling or possibly this board. We shouldn't forget 82°C is close to the CPU's limit so even more voltage is not going to get us great results either way. Lets not forget at 1.30V we have increased the voltage with 30% already.

After reviewing some settings, I've noticed I had analog and digital voltage set, although this should only matter for RAM overclocking. I've also slightly increased the System Agent Voltage offset to 0.250V as a gut feeling.

I've seen others also topping out at 4.7 GHz, with roughly the same settings. This might mean we're hitting a wall somewhere, which I'm too inexperienced to be making guesses at. So instead of reaching for the stars, lets find a more comfortable setting by slowly reducing the voltage and see how low I can go to attain the maximum clock rate.

1. 47x100 at 1.300V and Cache Ratio x45 --- 75-80°C - ~1 hour of Prime95, moving on
2. 47x100 at 1.275V and Cache Ratio x45 ---

 

[Vittra]

A couple of notes:

-One of the most confusing aspects about overclocking is actually what stability software to use. You will see various discussions (sometimes outright arguments) regarding what piece of software should be used, or even what specific version of a software to use (most often regarding Prime95). These discussions also vary by the platforms specific generation. I believe for an overclock to be truly stable, it should pass most of the common tests (Prime95 / AIDA 64 / Rog Realbench / IntelBurnTest), even with the latest punishing AVX instructions... so the question really becomes "how much instability am I willing to accept?". Usually the margin of this question can be quite high if gaming is the focus.

-VCCIN/Input Voltage has been said to play a pivotal role in stabilizing higher OCs on Haswell and Haswell-E based chips. You may want to spend some more time tweaking this particular value on your G3258 to see if a 4.8ghz OC can be stabilized.

-If you do decide to revisit that 4.8ghz target, drop your cache speeds back down to default as it could affect your overall core speed. Cache speed within 400-500mhz of the core speed is actually considered to be sufficient, although cache speeds seem to provide very little benefit overall.

I did quite a bit of overclocking with the 5930K but wasn't really pleased with the results, hence why that rig was ultimately sold. Regardless of setting tweaks, it refused to go beyond 4.2GHZ, and my target was 4.4. My 4770K was mostly run at stock, and prior to that my 2600K ran at 4.7ghz for the time I had it.

I haven't touched overclocking with my 6700K. Yet. I actually specifically bought this CPU because it boosts to 4.2GHZ without my intervention. This time around, I will probably use the OC capability Asus baked into the UEFI directly (largely based on Asus Suite) or install the Asus Suite in Windows to see what the algorithms can achieve, and then work based on those values.

 

[Phuncz]

Thanks for the tips Vittra, today I've looked at tips on OCN about Haswell overclocking and I made the same conclusions: more focus on Input Voltage, drop the cache ratio and see if 4.8 GHz is attainable. At the moment it's running through AIDA64 at 4.8 GHz with the cache at 32x and Input Voltage at 2.1V stably at over 8 minutes now.

My main concern about stability is Prime95 as that seems to bring out potential problems. AIDA64 seems good to see if it might actually run "basically stable", Prime95 shows me if I should worry about heavy usage scenario stability.

EDIT:

So I may be on to something, I'm 40+ minutes in on AIDA64, the stable 4.8 GHz OC seems attainable. This CPU seems to rarely go beyond 4.7 GHz so it looks like I won the chip lottery ! Although those few more percents of clock aren't really amazing, it still hits the 50% overclock mark if it proves to be stable.

What I've done is set the Input Voltage to 2.1V, the Core Voltage to 1.325V and Cache Ratio to x32. I have yet to touch LLC but it may help with Prime95 stability. After an hour of AIDA64, I'm going to try Prime95 with Load Line Calibration Level 1, which should keep the voltages in check.

1. 48x100 at 1.325V and Input Voltage at 2.100V --- 77-82°C - Passed: 640K, BSOD
2. 48x100 at 1.325V and Input Voltage at 2.150V --- 77-82°C - Passed: 640K, BSOD
3. 48x100 at 1.337V and Input Voltage at 2.150V --- After ~1 minute BSOD
4. FIVR faults disabled, Cache Voltage set to 1.3V --- 80-87°C - Passed: 640K, 8K, BSOD
5. 48x100 at 1.375V and Input Voltage at 2.150V --- 87-93°C - Passed: 1+ hour, moving on

 

[Phuncz]

So I've determined that 4.8 GHz is attainable on my air-cooled scrapheap, although just barely. But I don't like that part and all that for a smidge less of performance. I could potentially delid the CPU to lower the temps by 7-10°C but I'm not liking those voltages either way.

These are the achieved overclocks at the different voltage stages:

40x100 at 1.010V --- Input Voltage at 1.700V --- ~45°C (4.0 GHz - 25% overclock)
42x100 at 1.100V --- Input Voltage at 1.900V --- ~49°C (4.2 GHz - 31% overclock)
45x100 at 1.200V --- Input Voltage at 1.900V --- ~57°C (4.5 GHz - 41% overclock)
47x100 at 1.300V --- Input Voltage at 1.900V --- ~68°C (4.7 GHz - 47% overclock)
48x100 at 1.375V --- Input Voltage at 2.150V --- ~90°C (4.8 GHz - 50% overclock)

What I'm going to try next is to gain as much as I can from 1.2V Core and see how much I can lower the input voltage. 4.5 to 4.6 GHz is going to be the focus now.

Some lessons I took away from this:
- an AsRock Z87E-ITX can be an excellent overclocking tool
- the G3258 is a lot of dual-core performance and OC'ing toy for the low price
- this shouldn't be done in a packed air-cooled mITX case with low noise in mind
- overclocking Haswell feels much more confident than the Core 2 platforms and earlier

 

[Phuncz]

I'm picking up where I left off, namely getting as much from 1.20V core voltage as I can.
Input Voltage at 1.900V:

• Core 45x100 at 1.200V --- Cache x45 at 1.200V --- 62-65°C - 1 min Prime95 - BSOD
• Core 45x100 at 1.200V --- Cache x40 at 1.200V --- 66-69°C - 33 min Prime95 - BSOD
• Core 45x100 at 1.210V --- Cache x40 at 1.210V --- 66-69°C - 1:45 hour Prime95 - moving on

Input Voltage at 1.800V:

• Core 45x100 at 1.210V --- Cache x40 at 1.210V --- 63-66°C - 1:07 min Prime95 - moving on

Input Voltage at 1.750V:

• Core 45x100 at 1.210V --- Cache x40 at 1.210V --- 63-66°C - 2 hours Prime95 - moving on

Input Voltage at 1.700V:

• Core 45x100 at 1.210V --- Cache x40 at 1.210V --- 63-66°C - 2 hours Prime95 - moving on

So 4.5 GHz at 1.21V core with 1.7V input. I'm comfortable with this result, it's a good overclock and the voltages are OK. Next is to implement the voltages I need in a more dynamic scenario.

 

[Phuncz]

So today I'm doing the final step, incorporating the new settings into a more dynamic whole.
I've changed the CPU Voltage, Cache Voltage and Input Voltage from their fixed state (useful for finding an overclock) to an adaptive state, meaning as there is need for more voltage, this will be supplied. How ?

While I was trying to find an overclock, I made sure the voltage and core frequency were stable, so 1.21V and 4.5 GHz no matter the load. By re-enabling SpeedStep, the CPU will be able to lower it's clock speed to a minimum of 800 MHz (0.8 GHz) in a few steps. This means that while idle or using very limited CPU power, the processor will only work at a limited speed. This in turn allows the voltage to go down for these lower stages. I've set the voltage for 1.010V and added an +0.200V offset, meaning it will default to 1.000V and if it needs to step up it's game it will increase the volt up to 1.210V, the voltage I succesfully tested 4.5 GHz at.

But dynamic voltages have always been less precise than fixed voltages for me, so when I choose the above, I ended up with 1.020V in idle and 1.270V at load. This could be the Load-Line Calibration at work here, I have to test it.
So I chose 1.000V and +0.200V offset, which results in 0.924V in idle and 1.213V at load. I'm not so much worried about the idle voltage since 1.010V was enough for 4.0 GHz clocks.

 

[Phuncz]

Time for another run at some overclocking, today I'm finally going to squeeze the juice out of my Core i5-4670K.

My "little engine that could" OC system:

• Intel Core i5-4670K (3.4 GHz, 1.5 V --- Boost: 3.8 GHz, 1.15 V)
• Asus Maximus VII Impact (BIOS 3003)
  
• Noctua NH-L12 with two fan setup
  
• Crucial Ballistix Sport VLP 2x 8GB PC3-12800 (9-9-9-24)
• Samsung 960 EVO 256GB
  
• In Win Chopin with the sidepanel removed
  
• Fans set at "performance mode"
• Windows 10 x64 - basic install (including all drivers)
• AIDA64 for checking basic stability/temps/volts

Some basic settings I've made to start:

• CPU Ratio - All Core (so all cores get the same ratio)
• BCLK - 100.0
• Intel SpeedStep - Disabled
• Intel TurboBoost - Disabled
• Asus MultiCore Enhancement - Disabled
  
• RAM - XMP Profile 1 (DDR3-1600 9-9-9-24 1.35V)
• Cache min ratio: 35
• Cache max ratio: 35

First step: Probing.

Temps are average CPU after 15 minutes of AIDA stability test. I'll be trying to maximize the clock from the default 1.15 V up to 1.45 V, with the expectation to be running 1.3 V or lower long-term. The first part was done using only the 92mm fan on the heatsink.

Fixed CPU voltage at 1.15 V with 1.85 V input voltage (default)

1. 38 x 100 at 1.15 V --- ~70°C (default clocks)
   
2. 40 x 100 at 1.15 V --- ~75°C (4.0 GHz)
3. 42 x 100 at 1.15 V --- ~78°C (66°C package temp)
   
4. 44 x 100 at 1.15 V --- ~80°C (68°C package temp)
5. 46 x 100 at 1.15 V --- FAIL after 10 minutes

Fixed CPU voltage at 1.20 V with 1.85 V input voltage

1. 46 x 100 at 1.20 V --- ~90°C (77°C package temp)
   
2. 48 x 100 at 1.20 V --- FAIL at login

Fixed CPU voltage at 1.25 V with 1.85 V input voltage, adding 120mm fan to heatsink for dual-fan configuration

1. 48 x 100 at 1.25 V --- ~85°C (72°C package temp)
   
2. 50 x 100 at 1.25 V --- FAIL during boot

Fixed CPU voltage at 1.25 V with 1.90 V input voltage

1. 50 x 100 at 1.25 V --- FAIL during boot

Fixed CPU voltage at 1.30 V with 1.90 V input voltage

1. 50 x 100 at 1.30 V --- FAIL during boot

Here I decided to increase the CPU voltage to 1.35 V input voltage to 2.00 V, but also set the cache voltage to 1.25 V, System Agent voltage to 1.25 V and disable SVID Support.

1. 50 x 100 at 1.35 V --- FAIL at login
2. 50 x 100 at 1.375 V --- FAIL at login
3. 50 x 100 at 1.40 V --- FAIL at login

Fixed CPU voltage at 1.40 V with 2.10 V input voltage

1. 50 x 100 at 1.40 V --- FAIL at start of AIDA stress test
   
2. 50 x 100 at 1.43 V --- FAIL at start of AIDA stress test
3. 50 x 100 at 1.45 V --- FAIL at start of AIDA stress test

It's clear I'm hitting a ceiling here, either the CPU goes into thermal shutdown or I'm htting some other limit. Nonetheless I was able to boot into Windows, take a CPUID screenshot and do a submission to HWBOT.org. But clearly I should start finetuning with the CPU multiplier at 48, since I'm not planning to run continuously at 1.40 V or higher anyway and am not planning to use a beefier cooler.

 

[msystems]

What's the input voltage about? Don't think I've seen that before

 

[Phuncz]

Anand from Anandtech can answer this better than I ever can:

The other big part of the Haswell power story is what Intel is calling FIVR: Haswell’s Fully Integrated Voltage Regulator. Through a combination of on-die and on-package circuitry (mostly inductors on-package), Haswell assumes responsibility of distributing voltages to individual blocks and controllers (e.g. PCIe controller, memory controller, processor graphics, etc...). With FIVR, it’s easy to implement tons of voltage rails - which is why Intel doubled the number of internal voltage rails. With more independent voltage rails, there’s more fine grained control over the power delivered to various blocks of Haswell.

Thanks to a relatively high input voltage (on the order of 1.8V), it’s possible to generate quite a bit of current on-package and efficiently distribute power to all areas of the chip. Voltage ramps are 5 - 10x quicker with FIVR than with a traditional on-board voltage regulator implementation.

In order to ensure broad compatibility with memory types, there’s a second input voltage for DRAM as well.

FIVR also comes with a reduction in board area and component cost. I don’t suppose this is going to be a huge deal for desktops (admittedly the space and cost savings are basically non-existent), but it’ll mean a lot for mobile.

So increasing core voltage without also increasing input voltage isn't always going to be enough. On my G3258 this is 1.7 V by default, on my i5-4670K it is 1.85 V. I've set both to 2.1V when reaching for maximum clocks.

 

[msystems]

Oh thanks for that info, it was removed for Skylake apparently but it might be back in next generation Intel. Not sure if that is good or bad for overclocking.