GPU/CPU folding approach theoretical max system power draw?

[ocmusicjunkie]

Hey guys, I am trying to figure out the ballpark on just how much power is actually being used by my folding machines without having to go get a meter and test them all just for the rough estimate. Basically I know what a PSU calculator will tell me each rig would need at 100% load across the whole system. The part I don't know for certain is if folding at 100% on all CPU and GPU cores would be pushing on that figure, or if that would still leave some parts of a given system at low/idle consumption enough to matter. These are the three machines primarily being discussed along with their PSU output ratings, which are all at least 80+ bronze high quality units selected to be very near the calculated draw for each rig. Am I close enough to assume each would fold at around those wattage figures?

#1:
FX-8350 @ 4.3ghz (cpu:8 slot)
GT-620 x 2 (two GPU slots)
PSU 620w

#2:
X4 965 @ 4.1ghz (cpu:4 slot)
HD 6870 @1000mhz core (core 17 GPU slot)
PSU 550w

#3
X4 965 @ 3.8ghz (cpu:4 slot)
HD 6670 @825 core (core 17 GPU slot)
PSU 430w

With the cost of last month's electric bill generated by breaking 1m points for the month, it might be time to look at a 4p or some other more economical way to keep that PPD rate where it is with those systems...

[P5-133XL]

The problem with your power estimates is that you are busy OC'ing stuff and no calculator will estimate anything OC'ed they all base it on factory clocks. OC'ing often drastically increases the power draw especially if you end up increasing voltage as well as clock speed. You really need to incorporate a significant buffer into the PS especially if you intend to keep things running long-term.

A kill-a-watt meter is really quite inexpensive compared to paying to oversize/under-size you PS's. I recommend that route as opposed to your calculator method.

You will note that power supply's, in general, run most efficiently at around 50% of their max and the PS standards only use from 20%-80% capacity to judge their efficiency. Over 80% and under 20% you can end up in a quite inefficient range. Also as PS age, their maximum capacity decreases. If you are into keeping your PS long-term you will want to keep an additional buffer for that reason too. You really do not want to find out that after a year your PS no longer can supply the power it used to and now you are over-loading it.

I would recommend at least a 30% buffer beyond what a kill-a-watt measures. That should give you at little room for efficiency and aging.

[k1wi]

Getting a Kill-a-watt is really a sensible option for anyone who 1) cares about power efficiency, 2) has multiple systems 3) wants to overclock.

A particular note I'd make about using kill-a-watt measurements is that it's measuring plug usage, not component usage. Therefore to get the true measure of PSU utilisation you need to take off the 10-15% odd percent that the PSU is loosing at or around peak efficiency. I.e. a draw of 400W at the kill-a-watt on a 800W PSU is not 50% utilisation, but probably more like ~43-45%.

Further, if you're serious about PSUs then there are a ton of things to consider.

I'd personally use that 50% net usage quoted as a low-side value. You'll go from 50% through 75% at the absolute peak performance - depending on the PSU and it's own particular curve. Further, going above 75% towards 100% is really only going to cost you about a 2% drop in efficiency. What matters more at that point is PSU noise, which does start to increase relatively rapidly. There is an up front capital cost, albeit relatively small, in overspecing a PSU that may easily offset that 2% reduction in efficiency.

That said, I can see the argument in aiming for the lower % to start with to take into account component aging, I've just never really found my PSUs to reduce in efficiency as much as is commonly advised (no where near the 10% degredation p/a often quoted).

With a system that is folding non-stop (or whenever the system is turned on) you have more leeway to go down towards that 50% percentile, because it'll more or less always be running flat out. But if there are times when parts of your system are going to be running idle then there are some advantages in moving up closer towards the 70% value, because you will gain efficiency at the lower end when the system is idle. With systems being increasingly good at running at idle this becomes more of an issue.

Further you'll need to consider where you personally expect things to go in the future - how far you 'look out' depends on how long you expect to have the PSU. If you're anticipating the purchase of additional GPUs etc for your system in the future then obviously add that into the mix. If you're intending to make upgrades to the system, for example replacing the GPU, and intend to keep the PSU then unless you're going up a class in hardware (i.e. from a 460 to a 680), and you expect components to become less power hungry, then you might want to expect a long term reduction in power consumption...

The other thing to note is you have 3 different PSUs with different outputs and I'd assume ages, buying a new PSU is going to involve additional sink costs that may not offset any improvements in efficiency. On the other hand, with the kill-a-watt you might want to look at moving the PSUs to get the best configuration.

In any event, go get a kill-a-watt, there are other uses for it too.

[ocmusicjunkie]

Thanks for the thorough sharing of your thoughts- I read through all of it and will keep some if it in mind to be certain when looking at planning for future folding rigs. I should clarify, I did actually use a PSU calculator that accounts for the CPU overclock (although not GPU), and gave myself ~20% room to spare in terms of the unit's wattage rating compared to the minimum amount spec'ed per the calculator. To me, that's "very close" to the minimum requirement as I know there are all sorts of issues with the potential of running a unit at nearly full draw around the clock, and under the sort of case temps that are created by folding to boot. The 620w and 430w units are Seasonic and I trust to be more than capable for a good while. The 550w is a OCZ ZT that I took a chance on after reading some surprising reviews about it's noise/ripple and overall build quality- so far, so good.

I guess I might have not been very clear on what it is I'm trying to figure out- although I know that a kill-a-watt meter is probably the easiest way to be sure, I'm just curious if it's probable that those three units would actually be consuming close to the 1500w total rating of their power supplies when figuring draw from the outlet. I guess that I hadn't really given much thought to just how much the folding has seemed to change my power bill last month (nearly doubled to over $500)... the only plan available for my home here in CA is the ridiculous tiered system where at some point, adding 15% more consumption could add 50% to the dollar amount owed. It isn't a question of having to bail on the project because of power cost... rather, it's reconsidering just how much more logical the one-time investment in a single 4p system capable of running bigadv units with what might be 30-50% of the power requirements might become if this is going to be a normal monthly cost now...

[kscott]

I think you should sell all 3 and get a Intel system with 2 or 3 GPUs and a good 80+ Gold PSU
I'm running one machine with 4 clients and got almost 2mil for last month and my monthly bill is low

[Nathan_P]

Just to give you some idea of what power drawn can be, based on rigs I have owned or where I have seen power draw reported by more than one person

x6 1090T + GTX 460, GTS 250 and GTX 275 folding 24/7 approx. 500w for 37-45k PPD
x6 1090T + 1 GTX 460 - folding 24/7 350w for 25-30k PPD
Dual Xeon L5640 folding 24/7 217w for 45k PPD
Dual Xeon x5670 folding 24/7 (bigadv) 297w for up to 160k PPD
Dual E5 Xeon @2.3Fhz folding 24/7 (Bigadv) 248w for up to 248k PPD

4p 8 core AMD system not OC 400-500w
4p 12 core AMD system Upto 1100 w depending on overclock but upto 600-700k PPD

[mdk777]

rather, it's reconsidering just how much more logical the one-time investment in a single 4p system capable of running bigadv units with what might be 30-50% of the power requirements might become if this is going to be a normal monthly cost now...

I have to agree with kscott.

For the last 4 years, investing in a 4p and reducing your electric bill was the quickest and surest way to 200k to 400k ppd.

However, if you read the posts in the blog...GPU with opencl is going to quickly become the future path.

slap one 7970 onto any one of those rigs...running one client only...in the very near future(less than a year hopefully) you will be looking at close to 80,000 ppd(my prediction..nothing is written in stone). from something like 300 watts.

EDIT: pulled out the old kill-a-watt. I have not checked recently only running the GPU core alone.

My rig
X4 965 @ 3.8ghz
AMD 7970 OC 1050
750 watt gold rated PSU

Folding GPU only and web browsing = 235 watt.

check the beta information for current ppd.

Can you still do better from a ppd/watt with a 4p ? assuming they continue to exist in the current form and supply...sure.

However, you can't buy and set up a 4p for $450.

Anyway, this is all educated guessing until the new core17 launches with openmm5.1. (weeks, several months...hopefully less than a year.)

PS, you can get a PLAT 850 PSU on sale at fresh poultry site for $110(easily run 3 high end GPU {the most that easily fits on the average MB anyway}) . No reason to use anything other than PLAT for 24/7/365 folding now that prices have come down so much...if you do consider a new rig....saving $110 over an existing PSU would take some years of power savings...but can be justified depending on cooling and local power costs.

Yes, electricity is the most important factor in long term folding...but low initial investment costs should make gpu folding more attractive soon with advanced core and QRB.

[Jesse_V]

However, if you read the posts in the blog...GPU with opencl is going to quickly become the future path.
...
Anyway, this is all educated guessing until the new core17 launches with openmm5.1. (weeks, several months...hopefully less than a year.)

The blog posts and other information does indicate that OpenCL will be the future for F@h on GPUs. I know the scientists are quite excited about OpenMM 5.1, which according to this blog post introduces optimizations that deliver a 2x speedup. The official OpenMM site only offers 5.0, but since the blog post said "We have been aggressively working on OpenMM" I would speculate that we'll see 5.1 in FahCore 17 fairly "soon", perhaps measurable in weeks, but only time will tell. It will come out when it comes out...

[kscott]

However, if you read the posts in the blog...GPU with opencl is going to quickly become the future path.
...
Anyway, this is all educated guessing until the new core17 launches with openmm5.1. (weeks, several months...hopefully less than a year.)

The blog posts and other information does indicate that OpenCL will be the future for F@h on GPUs. I know the scientists are quite excited about OpenMM 5.1, which according to this blog post introduces optimizations that deliver a 2x speedup. The official OpenMM site only offers 5.0, but since the blog post said "We have been aggressively working on OpenMM" I would speculate that we'll see 5.1 in FahCore 17 fairly "soon", perhaps measurable in weeks, but only time will tell. It will come out when it comes out...

That would be nice especially if Nvida comes out with drivers to support it
Right now I can get 38k on 7662 with my 570

[k1wi]

However, if you read the posts in the blog...GPU with opencl is going to quickly become the future path.
...
Anyway, this is all educated guessing until the new core17 launches with openmm5.1. (weeks, several months...hopefully less than a year.)

The blog posts and other information does indicate that OpenCL will be the future for F@h on GPUs. I know the scientists are quite excited about OpenMM 5.1, which according to this blog post introduces optimizations that deliver a 2x speedup. The official OpenMM site only offers 5.0, but since the blog post said "We have been aggressively working on OpenMM" I would speculate that we'll see 5.1 in FahCore 17 fairly "soon", perhaps measurable in weeks, but only time will tell. It will come out when it comes out...

That would be nice especially if Nvida comes out with drivers to support it
Right now I can get 38k on 7662 with my 570

Drivers to support OpenCL? It already does - just not as well as the AMD drivers do.

Support the 2x increase in performance? Seeing as that's an openMM improvement, rather than a change to OpenCL, then the improvement should be relatively NVidia/AMD agnostic. Although, given the complex relationship between architecture, drivers and machine language one may make more of an improvement than the other. But as far as support goes, at the moment there is no NVidia lack of support.

[Joe_H]

Drivers to support OpenCL? It already does - just not as well as the AMD drivers do.

Support the 2x increase in performance? Seeing as that's an openMM improvement, rather than a change to OpenCL, then the improvement should be relatively NVidia/AMD agnostic. Although, given the complex relationship between architecture, drivers and machine language one may make more of an improvement than the other. But as far as support goes, at the moment there is no NVidia lack of support.

The question remains though. Since AMD only supports OpenCL 1.1, will the improvements in OpenMM port or do they depend on features in OpenCL 1.2? What there is may not be "no nVidia lack of support", but a lack of sufficient support.

[k1wi]

Drivers to support OpenCL? It already does - just not as well as the AMD drivers do.

Support the 2x increase in performance? Seeing as that's an openMM improvement, rather than a change to OpenCL, then the improvement should be relatively NVidia/AMD agnostic. Although, given the complex relationship between architecture, drivers and machine language one may make more of an improvement than the other. But as far as support goes, at the moment there is no NVidia lack of support.

The question remains though. Since AMD only supports OpenCL 1.1, will the improvements in OpenMM port or do they depend on features in OpenCL 1.2? What there is may not be "no nVidia lack of support", but a lack of sufficient support.

AMD and Intel currently has driver support for OpenCL 1.2 on their applicable hardware, NVidia only supports OpenCL 1.1 at present.

My understanding is that the speed up is due to OpenMM changes, not OpenCL changes. However, because this is the general section of the forums, I cannot go into greater detail than that.