AVX512 hardware recommendations?

[ewmayer]

Originally Posted by sdbardwick

My power adapter from same model NUC bought new in 2019 (hi-res for easy reading)

Yep, that's the one - ginormo-brick, relative to the size of the NUC proper.

Tried a little experiment - nuc1 draws a steady 18.4W under full 4-thread/avx2-math load ... kill Mlucas, wattage at wall drops to 5.5W. Thus assuming the wattmeter is providing accurate numbers, plugged in but NUC off the older smaller power brick draws under 1W, system-on adds ~5W, Mlucas-crunching adds another 13W.

OK, repeat experiment with nuc2 - wait, I unplugged the wattmeter from that so I could use it on nuc1. So power down both, transfer wattmeter to nuc2, run the experiment in the opposite direction: Hmm, this time on plug-into-wattmeter the brick draws ~0.6W, similar to nuc1's. Switch nuc2 on, spikes to ~18W ... maybe I mentally mixed up that sequence in my post yesterday? Anyway, this time before firing up Mlucas I watched the wattmeter for a few minutes ... it's all over the place for the first minute as the system boots up and inits various things, but then settles down to between 8-9W, a few watts more than nuc1 in the same state. Fire up 4-threaded Mlucas_avx512, we go to the now-familiar 33-34W, i.e. Mlucas adds 25W.

That 25W figure seems odd, though - for this CPU, per Tom's Hardware, bolds mine:

The Crimson Canyon NUCs are built around the 8th generation Intel Core i3-8121U processor. The chip is based on the Cannon Lake microarchitecture and produced with Intel's 10 nm manufacturing process. The Core i3-8121U is a dual-core part with Hyper-threading that runs at 2.2 GHz and has a turbo clock that boosts up to 3.2 GHz. The processor also has 4 MB L3 cache and a low, 15 W TDP (thermal design power) rating.

I will still try hooking both NUCs up to the old power supply when the splitter arrives - worst case it will save no watts, but will allow me to have both NUCs on/around my monitor (currently nuc2 is elsewhere due to lack of plugs in my under-desk power strip) with just 1 power cable. Nuc1 - a shortie like Mike bought, due its SSD memory stick - will remain affixed to back of monitor, but nuc2 will fit nicely between left edge of the monitor and the desktop fan I use to blast air over my multi-GPU desktop supercruncher. That fan has its top 2" sticking up over top of the upper glass monitor stand above the supercruncher, putting nuc2 there will allow it to benefit from that currently-wasted airflow. And if/when the older 65W power brick fails, I have the newer beefier (90W rating) one as a backup.

[Xyzzy]

You can adjust the TDP in the BIOS.

There are tons of options. You can turn off boost, hyperthreading and even cores. You can adjust voltages and multipliers. It is kinda overwhelming!

Our NUC has a 28W CPU. This is part of the reason we chose it over a newer model.

https://www.reddit.com/r/intelnuc/co...0_performance/
https://www.reddit.com/r/intelnuc/co...nuc8_vs_nuc10/

[mackerel]

Originally Posted by sdbardwick

For example, at low loads some newer PSUs draw only near the +- peaks of the AC line; (less current for same power) and rely on capacitors to maintain output when not drawing from the AC line. This intermittent draw can fool power meters into misrepresenting the actual power consumption.

Sounds like you're describing a full wave bridge rectifier arrangement. While I'm not familiar with worldwide power efficiency standards, I think they've pretty much been consigned for the history books outside of very low power scenarios.

The drive for higher power factor in loads requires more even loading throughout the power cycle, which should reduce potential differences between apparent power and true power.

Originally Posted by ewmayer

That 25W figure seems odd, though - for this CPU, per Tom's Hardware, bolds mine:

The power you measure at the wall will include any efficiency losses in the power supply and any other power conversion before it reaches the CPU. Things other than the CPU will also be running under load e.g. ram.

[kriesel]

Pplug = (CPU_TDP + ram + storage + fan + gluelogic + gpu + NIC + audio + USB)/(onboard voltage converters effective efficiencies)/(efficiency of copper lines internal)/(brick and external cords efficiency)

Pbilling = Pplug/(UPS_efficiency)/(house_wiring_efficiency)*(utility_meter_scale_factor)

House wiring transmission efficiency is close to unity on a lightly loaded circuit but could become a factor if there are a few RadeonVIIs drawing on the same circuit as the NUC. Running above 75% of the circuit rating may involve a few volts and few % power loss between utility meter and wall outlet.

[ewmayer]

The coax-power-plug splitter cable arrived yesterday ... before unplugging nuc2 and moving it onto my desktop monitor stand I checked wall-watts for both nucs ... nuc1 a stady 18W, nuc2 28W, which is 5W less than the 33W I reported previously, no idea whence the difference, seems to be some random luck-of-the-draw thing on powerup. After hooking both nucs to 2 power-out plugs of the available 4 on the splitter, powering up and firing up my usual 4-HT-threaded Mlucas jobs on each, power is a steady 46W, just the sum of the 2 separate-power wall wattages. But it was very nice to be able to put nuc2's big power brick and attendant cabling back in the box, and the splitter cable is invisible tucked away behind the monitor stand, as shown in the pic below.

Having still-unused power-outs on the splitter and a now-unused 19V brick makes it tempting to look for a cheapie used NUC online, perhaps one whose power brick failed ... something along these lines. But that particular one would need another $15 for a wifi stick, driving the total to over $100 once CA sales tax is factored in. Plus I resolve to stay disciplined - ha, ha, wait for it - in the sense that another further hardware-$ outlay this year will go to a 5th and final Radeon VII GPU which would join my very first one in my Haswell-CPU ATX-case system and finish the build-out of my existing setup. But will keep eyes out for sub-$100-total used NUCs using SSD memory, i.e. of the shorter form-factor like the aforelinked one. I'm down to just 3 of the 12 Andoid broke-o-phones which were my main GIMPS-crunching hardware outlay last year, I doubt more than one of those will last through end of this year, the batteries slowly deteriorate and apparently the system software does not allow the damn things to run sans working battery, simply off a microUSB power/recharge cable.

[ewmayer]

Update on the AVX-512 NUC: I noticed from the 'sensors' data that nuc2 runs 10-15C cooler than my older avx2-using nuc1. So 'can I overclock this sucker?' came to mind. Reboot <f2> shows promising stuff in the Intel Visual BIOS, but AFAICT the bios won't let me edit any of the cpu or ram performance-related settings. Mike suggests there is a hidden 'unlock' somewhere, I asked him if he could find it next time he reboots his NUC, please let me know. But there was one possibly-performance-related setting I *was* able to edit, that for max power, which was set at a default = 15w, I bumped up to 25w and upped max-burst-power from 20 to 30w. Those got save on bios-exit-and-reboot, and the result on my ongoing Mlucas run @5.5M FFT was a 15% speedup. Watts-at-wall rose from 28 to 36, and 'sensors' indicates core temp 10C higher, up from 60c to 70C, roughly what my old NUC runs at. Nice.