TiN
04-11-2011, 05:06 AM
Hi, dear extreme overclockers.
Most of you already benchmarked and overclocked lots of different videocards, from low-end to super-high-end multi-GPU solutions. But not all understand the fact about power required for successful system benching during LN2 overclock sessions. In general running system with limited power available would not do anything very bad, you will just suffer from OCP limits, throttling and other issues. But in rare cases some really bad damage may happen, if power supply will not handle big overload and blow whole rig, even killing HDD's and mouse with keyboard (yes, this is possible in practice :D). Also having margin in power supply rating to system consumption helps for better stability and no throttiling of any kind.
We all know about TDP, TDC and similar specs of various hardware, including CPU's and GPU's. But that values relate only to stock clocks and stock voltages, with stock air temperatures.
In this thread we will discuss all about power for CPU's and GPU's. But principle is same for any hardware, including HDDs, memory, motherboard chipsets etc, FYI.
There are few facts:
1. Power draw increases very slowly with raising clocks (almost by linear law).
2. But contra to (1) consumption fly like a rocket when you raise voltage, even by not much values, like +100mV will get you 20-30% of more power draw on GPU, with _SAME_ videocard frequencies.
GPU POWER section
How to figure out how many power does videocard take?
We need to measure both voltage on the videocard input power connection, and current flowing from MB/PSU to videocard. While measuring current on 6/8pin connection is not a hard task, it will not be true power consumption of card, because there is also +12V rail going from PCI-Express slot. So we need to use interposer/adapter with sensing curcuitry between motherboard and videocard.
For example for next GTX 580 card with extra VRM power testing I use next rig:
http://xdevs.com/evga/power/power_meter.jpg
It's important to have both current and voltage measurent on same time for accurate measurement, because loading is not constant in time but changes a lot during different image rendered by videocard each frame. So it's not like stready line of 500W, but like wavy, 50W->190W->500W->300W->600W->70W->90W->200W and so on, but very fast, these transitions in power occur hundreds times per each second of benchmark run.
This is why using multimeters is useless to measure power, you will just have some average value, having max power needed unknown.
So I used high-speed oscilloscope (Tek TDS5034B) with voltage probe and current probe (rated up to 150A) to constantly collect power data. Scope also allows you to see actual power graph, and comfortable to make screenies which are easy to understand.
I used current probe to sense sum of currents from motherboard slot to videocard (+12V contacts, and power 8+8 connectors on videocard).
This was +12V power was taken from 24pin main connector/aux connector, and from +12V5, +12V6 rails from power supply
http://xdevs.com/evga/power/curr_s3.jpg
You can see on photo adapter board between videocard (GTX590 on photo), current probe (big stuff on bottom left, with red sense wires +12V) and voltage probe for +12V voltage measurement.
http://xdevs.com/evga/power/curr_s2.jpg
I've used kingpincooling.com Tek9 6.66 VGA pot for ability to use any temperature from -180C to +80C during power testing.
All tests were done on sandybridge platform with CPU speed fixed to 5G.
http://xdevs.com/evga/power/curr_s1.jpg
Close up on current sensing probes. 3 wires: +12V5, 12V6 and 12V from motherboard.
Lets check different videocards power rating:
EVGA GeForce GTX 580 (reference PCB), stock reference card clocks and 1.2G overclock with LN2, Tek9 6.66, 1.42-1.5Vgpu, 1.82-1.85Vmem from Untouchables, +20°C ambient air temp, -100°C pot temp range
3D11, stock reference clocks, just raised voltages.
Just idle in windows, lowest power state.
60-70W already
http://xdevs.com/evga/power/580/772_2k/idle.png
GT1 - 500-540W taken by card.
http://xdevs.com/evga/power/580/772_2k/gt1.png
GT2 - 520-570W
http://xdevs.com/evga/power/580/772_2k/gt2.png
GT3 - 520-600W
http://xdevs.com/evga/power/580/772_2k/gt3.png
GT4 - 540-600W
http://xdevs.com/evga/power/580/772_2k/gt4.png
Temperatures kept at -80°C during this test.
Now lets overclock card to 1200-1250 GPU clock, and 1250MHz memory.
Same voltages, just overclock and repeat benchmark run.
Just idle in windows, lowest power state.
65-75W already
http://xdevs.com/evga/power/580/1250_25k/idle.png
GT1 - 650-700W taken by card.
http://xdevs.com/evga/power/580/1250_25k/gt1.png
GT2 - 670-740W
http://xdevs.com/evga/power/580/1250_25k/gt2.png
GT3 - 670-820W
http://xdevs.com/evga/power/580/1250_25k/gt3.png
GT4 - 660-850W (!). Peak current from +12V is about 68-70 AMPs.
http://xdevs.com/evga/power/580/1250_25k/gt4.png
For those who forget, that is SINGLE videocard, with 1.45-1.48Vgpu, and 1.85Vmemory.
Now little interesting.
Physics test, at begin.
http://xdevs.com/evga/power/580/1250_25k/phy_start.png
Card in 3D mode, but works only little fraction of time. 620-650W spikes.
And after first second, card goes low-power mode.
http://xdevs.com/evga/power/580/1250_25k/phy_mid.png
Now low power, 160-180W.
And it stays low-power during whole combined test load time, so no pouring needed here.
Then first second of combined test card still low-power mode, that's why we have slowdown startup, and then finally it switch to 3D mode, with full clocks.
Combined test after switch to 3D
http://xdevs.com/evga/power/580/1250_25k/comb.png
600-670W, much less than GT tests.
Result score during this power test:
http://xdevs.com/evga/power/580/3d11.png
Now some other benchmarks measurements:
3Dmark Vantage Performance
GT1 Vantage
http://xdevs.com/evga/power/580/1250_25k/gt1_vant.png
600-720W, less than 3D11 GT tests, but I personally think that Vantage loads more on memory, and 3D11 loads more on GPU.
GT2 Vantage
http://xdevs.com/evga/power/580/1250_25k/gt2_vant.png
640-680W (actually peak values less than GT1)
Result
http://xdevs.com/evga/power/580/vantage.png
3Dmark03 Nature
http://xdevs.com/evga/power/580/1250_25k/nature03.png
530-600W. Not as much for this oldie
Aquamark default run
http://xdevs.com/evga/power/580/1250_25k/aqua.png
440-500W, even less :)
For comparsion: Reference NVIDIA GeForce GTX 580 PCB without any mods, just stock card on stock cooling:
Same 3D11, P-run
GT1
http://xdevs.com/evga/power/580/772_2k/gt1.png
GT2
http://xdevs.com/evga/power/580/772_2k/gt2.png
20-24A mean that stock reference GTX 580 eat about 240-280W from power supply during 3D11 test.
EVGA GeForce GTX 590, stock reference card clocks, stock cooling, stock voltages, +20°C ambient air temp
3Dmark11, Xtreme preset (power rating actually same as Performance preset) , GT1 run
http://xdevs.com/evga/power/590/gt1.png
Power draw here is in range 330-350W.
3Dmark11 , GT2 run
http://xdevs.com/evga/power/590/gt2.png
Power draw here is in range 340-360W.
3Dmark11 , GT3 run
http://xdevs.com/evga/power/590/gt3.png
Power draw here is in range 340-360W.
3Dmark11 , GT4 run
http://xdevs.com/evga/power/590/gt4.png
Power draw here is in range 340-380W.
TEMPERATURE related power deviation.
TBW later.
CPU POWER section
TBW later.
Running multi-PSU rigs
Why this thread importance is underestimated? Here is the answer :)
All this power ratings, delivery and distribution of +12V rails knowledge does not matter much when you run single CPU, single VGA system. In that simple cases it's enough just to have some beff PSU, like 1-1.2kW and you are fine and safe.
But when we talk about 2-way SLI systems with both GPUs under LN2 and big overvoltage, paired with CPU LN2, wrong rails distibution will crush whole benching session, or at least limit overclocking.
How to solve that problem?
TBW later.
Stay tuned. :pconfused:
Most of you already benchmarked and overclocked lots of different videocards, from low-end to super-high-end multi-GPU solutions. But not all understand the fact about power required for successful system benching during LN2 overclock sessions. In general running system with limited power available would not do anything very bad, you will just suffer from OCP limits, throttling and other issues. But in rare cases some really bad damage may happen, if power supply will not handle big overload and blow whole rig, even killing HDD's and mouse with keyboard (yes, this is possible in practice :D). Also having margin in power supply rating to system consumption helps for better stability and no throttiling of any kind.
We all know about TDP, TDC and similar specs of various hardware, including CPU's and GPU's. But that values relate only to stock clocks and stock voltages, with stock air temperatures.
In this thread we will discuss all about power for CPU's and GPU's. But principle is same for any hardware, including HDDs, memory, motherboard chipsets etc, FYI.
There are few facts:
1. Power draw increases very slowly with raising clocks (almost by linear law).
2. But contra to (1) consumption fly like a rocket when you raise voltage, even by not much values, like +100mV will get you 20-30% of more power draw on GPU, with _SAME_ videocard frequencies.
GPU POWER section
How to figure out how many power does videocard take?
We need to measure both voltage on the videocard input power connection, and current flowing from MB/PSU to videocard. While measuring current on 6/8pin connection is not a hard task, it will not be true power consumption of card, because there is also +12V rail going from PCI-Express slot. So we need to use interposer/adapter with sensing curcuitry between motherboard and videocard.
For example for next GTX 580 card with extra VRM power testing I use next rig:
http://xdevs.com/evga/power/power_meter.jpg
It's important to have both current and voltage measurent on same time for accurate measurement, because loading is not constant in time but changes a lot during different image rendered by videocard each frame. So it's not like stready line of 500W, but like wavy, 50W->190W->500W->300W->600W->70W->90W->200W and so on, but very fast, these transitions in power occur hundreds times per each second of benchmark run.
This is why using multimeters is useless to measure power, you will just have some average value, having max power needed unknown.
So I used high-speed oscilloscope (Tek TDS5034B) with voltage probe and current probe (rated up to 150A) to constantly collect power data. Scope also allows you to see actual power graph, and comfortable to make screenies which are easy to understand.
I used current probe to sense sum of currents from motherboard slot to videocard (+12V contacts, and power 8+8 connectors on videocard).
This was +12V power was taken from 24pin main connector/aux connector, and from +12V5, +12V6 rails from power supply
http://xdevs.com/evga/power/curr_s3.jpg
You can see on photo adapter board between videocard (GTX590 on photo), current probe (big stuff on bottom left, with red sense wires +12V) and voltage probe for +12V voltage measurement.
http://xdevs.com/evga/power/curr_s2.jpg
I've used kingpincooling.com Tek9 6.66 VGA pot for ability to use any temperature from -180C to +80C during power testing.
All tests were done on sandybridge platform with CPU speed fixed to 5G.
http://xdevs.com/evga/power/curr_s1.jpg
Close up on current sensing probes. 3 wires: +12V5, 12V6 and 12V from motherboard.
Lets check different videocards power rating:
EVGA GeForce GTX 580 (reference PCB), stock reference card clocks and 1.2G overclock with LN2, Tek9 6.66, 1.42-1.5Vgpu, 1.82-1.85Vmem from Untouchables, +20°C ambient air temp, -100°C pot temp range
3D11, stock reference clocks, just raised voltages.
Just idle in windows, lowest power state.
60-70W already
http://xdevs.com/evga/power/580/772_2k/idle.png
GT1 - 500-540W taken by card.
http://xdevs.com/evga/power/580/772_2k/gt1.png
GT2 - 520-570W
http://xdevs.com/evga/power/580/772_2k/gt2.png
GT3 - 520-600W
http://xdevs.com/evga/power/580/772_2k/gt3.png
GT4 - 540-600W
http://xdevs.com/evga/power/580/772_2k/gt4.png
Temperatures kept at -80°C during this test.
Now lets overclock card to 1200-1250 GPU clock, and 1250MHz memory.
Same voltages, just overclock and repeat benchmark run.
Just idle in windows, lowest power state.
65-75W already
http://xdevs.com/evga/power/580/1250_25k/idle.png
GT1 - 650-700W taken by card.
http://xdevs.com/evga/power/580/1250_25k/gt1.png
GT2 - 670-740W
http://xdevs.com/evga/power/580/1250_25k/gt2.png
GT3 - 670-820W
http://xdevs.com/evga/power/580/1250_25k/gt3.png
GT4 - 660-850W (!). Peak current from +12V is about 68-70 AMPs.
http://xdevs.com/evga/power/580/1250_25k/gt4.png
For those who forget, that is SINGLE videocard, with 1.45-1.48Vgpu, and 1.85Vmemory.
Now little interesting.
Physics test, at begin.
http://xdevs.com/evga/power/580/1250_25k/phy_start.png
Card in 3D mode, but works only little fraction of time. 620-650W spikes.
And after first second, card goes low-power mode.
http://xdevs.com/evga/power/580/1250_25k/phy_mid.png
Now low power, 160-180W.
And it stays low-power during whole combined test load time, so no pouring needed here.
Then first second of combined test card still low-power mode, that's why we have slowdown startup, and then finally it switch to 3D mode, with full clocks.
Combined test after switch to 3D
http://xdevs.com/evga/power/580/1250_25k/comb.png
600-670W, much less than GT tests.
Result score during this power test:
http://xdevs.com/evga/power/580/3d11.png
Now some other benchmarks measurements:
3Dmark Vantage Performance
GT1 Vantage
http://xdevs.com/evga/power/580/1250_25k/gt1_vant.png
600-720W, less than 3D11 GT tests, but I personally think that Vantage loads more on memory, and 3D11 loads more on GPU.
GT2 Vantage
http://xdevs.com/evga/power/580/1250_25k/gt2_vant.png
640-680W (actually peak values less than GT1)
Result
http://xdevs.com/evga/power/580/vantage.png
3Dmark03 Nature
http://xdevs.com/evga/power/580/1250_25k/nature03.png
530-600W. Not as much for this oldie
Aquamark default run
http://xdevs.com/evga/power/580/1250_25k/aqua.png
440-500W, even less :)
For comparsion: Reference NVIDIA GeForce GTX 580 PCB without any mods, just stock card on stock cooling:
Same 3D11, P-run
GT1
http://xdevs.com/evga/power/580/772_2k/gt1.png
GT2
http://xdevs.com/evga/power/580/772_2k/gt2.png
20-24A mean that stock reference GTX 580 eat about 240-280W from power supply during 3D11 test.
EVGA GeForce GTX 590, stock reference card clocks, stock cooling, stock voltages, +20°C ambient air temp
3Dmark11, Xtreme preset (power rating actually same as Performance preset) , GT1 run
http://xdevs.com/evga/power/590/gt1.png
Power draw here is in range 330-350W.
3Dmark11 , GT2 run
http://xdevs.com/evga/power/590/gt2.png
Power draw here is in range 340-360W.
3Dmark11 , GT3 run
http://xdevs.com/evga/power/590/gt3.png
Power draw here is in range 340-360W.
3Dmark11 , GT4 run
http://xdevs.com/evga/power/590/gt4.png
Power draw here is in range 340-380W.
TEMPERATURE related power deviation.
TBW later.
CPU POWER section
TBW later.
Running multi-PSU rigs
Why this thread importance is underestimated? Here is the answer :)
All this power ratings, delivery and distribution of +12V rails knowledge does not matter much when you run single CPU, single VGA system. In that simple cases it's enough just to have some beff PSU, like 1-1.2kW and you are fine and safe.
But when we talk about 2-way SLI systems with both GPUs under LN2 and big overvoltage, paired with CPU LN2, wrong rails distibution will crush whole benching session, or at least limit overclocking.
How to solve that problem?
TBW later.
Stay tuned. :pconfused: