At 1.32 V CPU core, ASRock’s Turbo 4.60 GHz overclocking profile provided more than enough voltage to prevent crashing on this Core i7-4770K.
Unfortunately, that high voltage also caused it to throttle within seconds of starting an eight-thread run of AVX-optimized Prime95. Experimentation showed that a minimum of 1.275 V was needed to make this core run at 46 x 100 MHz, but that this voltage still caused thermal bottlenecking within minutes of my stability tests.

Some overclockers will point out that my stress test is unrealistically tough, that thermal throttling prevents damage, and that this protection mechanism takes minutes to start under Prime95. So, it'd be unlikely to affect our benchmarks. But when I look for a stable overclock, I avoid throttle conditions altogether. I instead dropped the CPU multiplier to 45x and began looking for the lowest voltage that would keep the CPU stable at 4.5 GHz.

A 1.24 V VCore worked great, so I also dropped the ASRock 4.60 GHz profile’s 1.30 V “CPU Cache” voltage to 1.24 V.

I increased the memory subsystem from its DDR3-1866 profile to DDR3-2133, and also configured it to 1.60 V.

To achieve stability at DDR3-2133, the three primary latencies all needed to be increased by one cycle beyond XMP settings. I usually find that tRAS can also be tightened, but that wasn’t true for this combination of memory and CPU.

As with the CPU, I don’t like the idea of graphics cards that throttle down when you need full power the most. I first increased the GPU power limit to its maximum slider setting, then began bumping up GPU clock in 25 MHz increments. After finding +175 MHz unstable and +150 MHz stable, I picked a frequency that would result in a nice round number for the GPU clock. Adding 157 MHz yielded a 1020 MHz GPU base clock and a 1059 MHz typical GPU Boost frequency.
- Our High-End Build Evolves
- Graphics, CPU, And Memory
- Motherboard, Case, And Power
- CPU And Motherboard Cooling
- An Alphabet Soup Of Storage: SSD, HDD, And ODD
- Hardware Installation
- Overclocking
- Test Hardware And Benchmark Settings
- Results: 3DMark And PCMark
- Results: SiSoftware Sandra
- Results: Battlefield 4 And Far Cry 3
- Results: Grid 2 And Arma 3
- Results: Audio And Video Encoding
- Results: Adobe Creative Suite
- Results: Productivity
- Results: File Compression
- Power, Heat, And Efficiency
- A Gaming Build That Works Hard
1.) Start the system, wait for all processes to load, take a measurement (Active, but idle)
2.) Load the CPU using eight thread of AVX-optimized Prime95, take a reading (CPU Load).
3.) Load GPUs with 3DMark 11 Test 1 in loop, take max reading as it heats up (GPU Load).
4.) Load both applications (CPU+GPU Load).
The "math problem" is that any program used to fully load the GPU also partly loads the CPU. So when test 4 is Prime95+3DMark, Prime95 can only use whatever CPU resources are left with 3DMark running.
So the most accurate system power reading is with "CPU+GPU Load" applied. The system measurement for "CPU Load" still includes the power of an idle GPU. And the system power measurement for "GPU Load" still includes the amount of CPU energy it takes to run the GPU's test application.
Power supplies of greater capacity and similar reliability at this price tend to be lower-efficiency units. And we like efficiency too.
1.) Start the system, wait for all processes to load, take a measurement (Active, but idle)
2.) Load the CPU using eight thread of AVX-optimized Prime95, take a reading (CPU Load).
3.) Load GPUs with 3DMark 11 Test 1 in loop, take max reading as it heats up (GPU Load).
4.) Load both applications (CPU+GPU Load).
The "math problem" is that any program used to fully load the GPU also partly loads the CPU. So when test 4 is Prime95+3DMark, Prime95 can only use whatever CPU resources are left with 3DMark running.
So the most accurate system power reading is with "CPU+GPU Load" applied. The system measurement for "CPU Load" still includes the power of an idle GPU. And the system power measurement for "GPU Load" still includes the amount of CPU energy it takes to run the GPU's test application.
The "math problem" is that any program used to fully load the GPU also partly loads the CPU. So when test 3 is Prime95+3DMark, Prime95 can only use whatever CPU resources are left with 3DMark running.
So the most accurate system power reading is with "CPU+GPU Load" applied. The system measurement for "CPU Load" still includes the reading of an idle GPU. And the system power measurement for "GPU Load" still includes the amount of CPU power it takes to run the GPU.
Very much appreciated and satisfying answer.
Thanks Crashman
http://pcpartpicker.com/p/3fuGw
Wondering how much of a difference would non-reference cards make. Obviously, CPU cooler and RAM could be different, BR drive optional, storage drive as well.
Shouldn't that be DDR3-1866?
my fix is get a 700gb ssd, 780ti no sli problems, and a i5 4670, this is a much better gaming pc, and can go quiet build.