Power Consumption Characteristics of Crypto Mining Rigs

[Fisherman199]

I have occasion to attempt to alleviate a lot of confusion on behalf of my utility (and me). Despite a well-trained search engine I can't seem to find a white paper from another utility/IEC/IEEE that tackles this question.

Does anyone here have a good source?

 

[anorlunda]

What about server farms? I see no special reason why a crypto mining load looks much different than a server farm load.

 

[Fisherman199]

What about server farms? I see no special reason why a crypto mining load looks much different than a server farm load.

Exactly. I don't either but this is one of the areas of confusion I hope to clear up. There's a lot of talk internally about how "different" the loads are but I can't think of a reason off hand why that would be the case. (speaking electrically)

I've read sporadic reports that triplen harmonics from bit-mining loads can be be more significant than in other "server farms," but, again, I've not found empirical data to back this idea.

 

[russ_watters]

I don't see what question you are trying to answer.

I can speculate someone might calculate a coin per kWh mining efficiency, but I don't know if that's what you are after.

I also don't know who cares. You mention the "utility". You mean the electric utility? Why do they care? Power factor?

 

[Fisherman199]

I don't see what question you are trying to answer.

I can speculate someone might calculate a coin per kWh mining efficiency, but I don't know if that's what you are after.

I also don't know who cares. You mention the "utility". You mean the electric utility? Why do they care? Power factor?

I'm referring to general consumption characteristics like harmonic injection, response to transients, etc.

The question is: Can bitcoin mining load be treated as analogous to other data center loads in characteristic? I thought I'd made it clear, so my apologies.

 

[anorlunda]

It is just computers. They may have powerful graphics GPU cards, but still just computers that use standard computer DC power supplies. I see no reason why their load characteristics should be different.

I could buy a used server farm, then set it to work mining bitcoins. It's load characteristics don't change.

I may be wrong, but I never heard of custom computer hardware for bitcoin mining.

 

[phinds]

I may be wrong, but I never heard of custom computer hardware for bitcoin mining.

I have. I've read that Nvidia and others have created special chips to use in servers specifically for bitcoin mining. That may make the power characteristics of those servers different than the average general purpose server.

https://www.nvidia.com/en-us/cmp/

 

[berkeman]

I have occasion to attempt to alleviate a lot of confusion on behalf of my utility (and me).

Is this in your home? Most server farms are located in industrially-zoned areas, with different power metering characteristics and requirements...

 

[anorlunda]

Is this in your home?

The OP is asking about how it from the power company point of view. In addition to power and VARS, each kind of load has different transient characteristics. The OP mentioned harmonic injection.

For example, some utility engineers are concerned that LED, incandescent or fluorescent lighting each has different load characteristics. Motor loads have inrush currents during startup. Utility engineers model those differences in their grid simulations.

I have. I've read that Nvidia and others have created special chips to use in servers specifically for bitcoin mining. That may make the power characteristics of those servers different than the average general purpose server.

I doubt if the chips in a computer make a difference to the power grid. The power supply (PSU) is the load seen by the grid. IMO no DC components downstream of the PSU are likely to be significant.

When I said I didn't know of special purpose hardware for bitcoin mining, I was thinking about non-digital-computer hardware that might connect directly to the AC mains. I'm pretty sure no such exotic stuff exists, but I'm surprised every other day by something new.

 

[Tom.G]

It seems that a server farm load would vary with (the local?) time, same as the Internet traffic does.

Where-as a Bitcoin farm would be a much more constant load.

Any arguments pro or con?

 

[Fisherman199]

It is just computers. They may have powerful graphics GPU cards, but still just computers that use standard computer DC power supplies. I see no reason why their load characteristics should be different.

I could buy a used server farm, then set it to work mining bitcoins. It's load characteristics don't change.

I may be wrong, but I never heard of custom computer hardware for bitcoin mining.

There is custom hardware built/designed specifically for mining. It's the only way to get a good hash rate. Mining space is saturated (though less so now that it's banned in China). It's all very impressive but the machine are useless for anything besides mining bitcoin.

 

[Fisherman199]

Is this in your home? Most server farms are located in industrially-zoned areas, with different power metering characteristics and requirements...

No, this is on utility scale.

 

[Fisherman199]

I have. I've read that Nvidia and others have created special chips to use in servers specifically for bitcoin mining. That may make the power characteristics of those servers different than the average general purpose server.

https://www.nvidia.com/en-us/cmp/

Yes, but the machines installed that I'm concerned about have hash rates in the TH/s.

 

[Fisherman199]

The OP is asking about how it from the power company point of view. In addition to power and VARS, each kind of load has different transient characteristics. The OP mentioned harmonic injection.

For example, some utility engineers are concerned that LED, incandescent or fluorescent lighting each has different load characteristics. Motor loads have inrush currents during startup. Utility engineers model those differences in their grid simulations.
I doubt if the chips in a computer make a difference to the power grid. The power supply (PSU) is the load seen by the grid. IMO no DC components downstream of the PSU are likely to be significant.

When I said I didn't know of special purpose hardware for bitcoin mining, I was thinking about non-digital-computer hardware that might connect directly to the AC mains. I'm pretty sure no such exotic stuff exists, but I'm surprised every other day by something new.

I don't think there is either but this seems to be a source of debate. An engineer yesterday was talking about the chips of the mining rigs as if they were significant to overall transient/harmonic injection. I'm constantly surprised by how much I don't know so I didn't say anything during the meeting. I stole myself away in my office to run some equations and I can't find anything to suggest I need to care about anything happening below the power supply level. I get the mining rigs are a more constant load and this presents its own set of problems but the idea they're fundamentally different than other computational loads escapes me.

Another thing that really worries me is the more I speak with other util engineers the more I find out many of the filters that data centers will install aren't typically installed for bit-miners.

At any rate, my question is pretty much answered. Thanks folks.

 

[Fisherman199]

It seems that a server farm load would vary with (the local?) time, same as the Internet traffic does.

Where-as a Bitcoin farm would be a much more constant load.

Any arguments pro or con?

They are constant. This represents a range of issues but none that can't be addressed from the inclusion of relaying or reclosing.

 

[Rive]

The power supply (PSU) is the load seen by the grid.

Maybe worth noting that SMPS loads has a negative resistance characteristic. Usually not a problem, but crypto HW always running on a constant load with close to the nominal power of PSUs involved, and often congregates on grids with low cost of electricity (what may mean poor grid properties).
With some of the reported/suspected load percentages of such grids it may be a problem.

 

[anorlunda]

Maybe worth noting that SMPS loads has a negative resistance characteristic.

Yes, the transformers in SMPS do present a particular characteristic load the to grid; especially considering harmonics. However, the SMPS load does not change its harmonic impedances depending on what software the computer is running. Thank goodness for that. Just imagine if the government could spy on what software you are running just by looking at your power records. That would be even worse than the Tempest fiasco.

Yes, bitcoin mining might tend to be more time invariant than other software applications.

But note that the OP says that he has his answer.

 

[Fisherman199]

Jut to give a little back to the community, here's a quick run-down of what happened:

A bitcoin miner is fed off a single 1200 amp breaker. Regulators have been changed to 12% regulation to handle more amperage (usually set to 5%). and the breaker TOC settings are at 720 (per phase) with a lockout timer of just 1 minute. Bank transformer can handle ~1,100 amps (loaded very near 90% of that now)

Over the weekend the bit-miner added another trailer of mining rigs and cut them in late Saturday night. After 20 minutes the breaker operated (opened and reclosed). This happened 8 more times. Each reclose was 20 minutes apart, almost exactly (this is a characteristic in itself). Looking at the phase current graphs from SCADA captures it took 20 minutes after each event for the load to break over the TOC pickup. Obviously, since the lockout timer is only 1 minute (fine under most circumstances) the breaker never locked out. Each time and event occurred the regulators (which are already running very hot) coil, paper, and oil are subjected to inrush, ultimately diminishing their life if this phenomena persists. This also affects the station transformer which is a big no-no.

Part of the OG question stemmed from the curious phenomena of the slow ramp-up in load after a quick reclose (out only 5 seconds) for so long a period of time. Cold load is a known phenomena but not one that occurs without power being gone for many minutes/hours and it's not something that continues to ramp up for 20 minutes (at least, not in my experience or the experience of many of the white-papers I've read). The slow ramp-up in load is not something I, or anyone else I've spoken with, is familiar with when comparing against data center load.

From here, since this aspect of expressed load ramp-up was something unfamiliar, questions about harmonic injections, responses to system transients, and other questions came forward. Thus, my appearance here.

 

[Rive]

Just a thought: the load consists only of mining rigs, or does those have some serious cooling too?

 

[anorlunda]

The slow ramp-up in load is not something I, or anyone else I've spoken with, is familiar with when comparing against data center load.

The slow ramp-up could be something explicitly programmed into the data center's administration. It might even be required in some locations that couldn't stand a step change. It could also be possible that the owners transferred the software, including ramp-up, from some other location without modification.

So, once again, the problem may be data center related, not specifically bitcoin mining data center related.

 

[rbelli1]

A data center will have battery/generator back up since unexpected shutdowns and continued down time is very expensive. Cost/benefit on providing that continuity is is very clear.

For a mining rig you experience money loss when not mining but you don't have thousands of customers and their lawyers calling up and demanding refunds. I can see how battery and generator backup might be much harder to justify on that kind of work.

The data center is going to switch over a large amount of already running hardware when the power comes back. If there is no backup on the mining rigs then you will be booting all of them when power returns. What causes them to boot slowly or at different rates is not clear.

BoB