SPCR

One year passed and nobody's talking about owning and tweaking them...

My G3220 has been running at 0.800V at 3GHz without any CPU-caused errors since day one, workload including overnight x264 and overnight gaming, surely that is worth something.

0.8 because that is the minimum my motherboard allows. Has anybody gone 3GHz and higher at lower voltages?

wwenze wrote:surely that is worth something

Not necessarily. Have you measured the effect?

HFat wrote:

wwenze wrote:surely that is worth something

Not necessarily. Have you measured the effect?

In terms of? At this stage in technological development it is mostly enthusiasm that creates market for over price and minimal power saving.

Whatever you like. Do you actually save a significant amount of power for instance?
What actually happens with the CPU voltage isn't as simple or transparent as it used to be...

And for what it's worth, I disagree with your assessment of the market.

There's not a lot of data out there on Haswell undervolting and motherboard mfgrs don't bother to state whether or not their UEFI will allow for negative core voltage offsets. The few data points I've seen show it's fairly easy to get -0.1V at stock speeds (Roughly .98V vs 1.08V) and YMMV vary lower than that. Hardwareluxx got to 0.9V on a i7-4770K sample unit.

The 0.1V drop is ~18% power savings. So, if your particular CPU uses 70W or so running Handbrake, undervolting to 0.1V drops it to ~58W. Not a lot on the power bill, but it does provide significant thermal benefit making silent CPU cooling an easier chore.

It's not that simple anymore (for one thing, the autonomous voltage management may disregard your settings). Measurements are required to assess any actual power savings.
And no task will ever cause that CPU to burn 70W. Likely not even half unless you torture-test the GPU. Obviously there's more potential for gains with quad-cores.

(edit: I recalled wrong. SPCR didn't undervolt or underclock a dual-core in that test.)

HFat wrote:It's not that simple anymore (for one thing, the autonomous voltage management may disregard your settings). Measurements are required to assess any actual power savings.
And no task will ever cause that CPU to burn 70W. Likely not even half unless you torture-test the GPU. Obviously there's more potential for gains with quad-cores.

(edit: I recalled wrong. SPCR didn't undervolt or underclock a dual-core in that test.)

I refer to quad core Haswell, like the one referred to in my post. 84W TDP can see 70W with high load application. Second, Voltage Offset Mode in UEFI specifically refers to offsetting the core voltage at load and works in hand with dynamically changing load voltage.

My asrock has voltage override alongside voltage offset. If using override according to cpuz (and bios) it stays at that 0.8V regardless of frequency.

I would reckon while VR is integrated much of the control would still remain external, otherwise overclockers would either have a high chance of destroying the CPU via auto voltage increments or the platform would be terrible for over clocking with auto voltage limits.

And going 0.785V using a combination of 0.8V override plus -0.015 offset causes freeze up after sone duration of stress. 0.775 bsods at logon screen iirc. So I was already near the limits.

Disregard voltage control, at near threshold voltages the common electrical relationships aren't reliable anymore either. Not that they have been since leakage came into the equation circa 2002.

Dug out my power meter, measured values are from wall

System is Antec TP-550 (Bronze) and GTX 460 (hence the high idle power)

Prime 95 Large FFT
1.0V (stock) @ 3GHz: ~78W
0.8V @ 3GHz: ~68W

Idle
0.8V @ 800MHz (same value whether undervolting or not): ~52.5W

Worth it? Yea my stock heatsink seems to think so.

wwenze wrote:Prime 95 Large FFT
1.0V (stock) @ 3GHz: ~78W
0.8V @ 3GHz: ~68W

Idle
0.8V @ 800MHz (same value whether undervolting or not): ~52.5W

Worth it? Yea my stock heatsink seems to think so.

Only worth it if you crunch a lot of numbers and basically useless when idling. That's generally been people's experience with undervolting modern Intel CPU but it's the first time I see Haswell measurements. If you wanted to do a lot of number crunching, would you use that CPU though? That's part of why I think it's a waste of time on dual-cores.
What I have yet to see is serious stability testing. My own stability testing with older CPUs when undervolting mattered enough that I cared strongly suggested that the testing procedures suggested on consumer forums are inadequate.

It makes sense the CPU would automatically set the voltage low when idling anyway but did you try to use no other voltage control than underclocking and the offset? If offsetting doesn't make a significant difference while crunching numbers at 800 Mhz, it would suggest the CPU may ignore the offset not only in turbo mode or when overclocking but also when it underclocks itself autonomously.

exactly. the only real savings is when you run the cpu full on 24-7. -and if you are doing that kind of work, you probably require a truly stable and reliable system. I don't see any real world benefits in undervolting anymore.

Useful or not, I can say that the Asus Z97I supports undervolting, without a lower limit as far as I can tell. There is auto, manual, adaptive and offset mode.

It is more complicated since both the CPU and stuff on the motherboards adjust voltages automatically. Hard to predict how these things will affect each other...
Even so - with all power saving features enabled my "Haswell Refresh" CPU (4690K) does run 3-4°C cooler if I set a negative offset of 80 mV.

After stress-testing all cores for at least 30 minutes:
* Stock settings, all cores at 3.9 GHz (turbo mode) - max 67°C.
* Limit cores to 3.5 GHz (makes fans much quieter!) - max 61°C.
* Limit cores to 3.5 GHz + 80mV negative offset - max 57°C.
Limiting cores to 3.5 GHz + 100mV negative offset causes kernel panic within minutes.

Seems to me like undervolting is still working just like with Core 2 Duo...?

Thanks for the data.

I'm coming from AMD Phenom II territory, where undervolting during idle made a big difference. I've moved to Haswell G3258, and it's been a whole 'nother world for me to learn.

Correct me if I'm wrong, but with the C3/C6/C7 C-states, there is no point anymore in enabling dynamic frequency, C1E, or EIST? I just set it to be pegged at my desired voltage and frequency, and let C6/C7 do its job?

And I still haven't figured out what the difference is between Core C-state and Package C-state.

If anyone has a link to a guide explaining all of this, that would be awesome. My interest over the years has been getting the best of both worlds between overclocking and undervolting by dynamically undervolting/underclocking the chip at idle. So I can get better performance from the overclock at load, and energy savings at idle from the undervolt/underclock. Seems all the rules have changed though from the AMD land I'm coming from.

davidh44 wrote:If anyone has a link to a guide explaining all of this, that would be awesome.

I don't think there's ever been a lot of info on undervolting, and it's even less with Haswell. If you find such a guide, please share it here.

davidh44 wrote:My interest over the years has been getting the best of both worlds between overclocking and undervolting by dynamically undervolting/underclocking the chip at idle. So I can get better performance from the overclock at load, and energy savings at idle from the undervolt/underclock.

Trying to save energy in idle doesn't make much sense to me as the CPU handles that very well with the technologies you mention. For me it's much more interesting to shave off a few degrees at peak performance, where cooling matters most.

Also you can now (?) undervolt using offset mode, which I believe reduces voltage across all states. See my last post.

davidh44 wrote:And I still haven't figured out what the difference is between Core C-state and Package C-state.

If anyone has a link to a guide explaining all of this, that would be awesome.

Not exactly a guide but this Wikipedia article might be useful, especially if you follow the links/references at the bottom of the page.