seasonal or long term trends

[chalsall]

Originally Posted by kriesel

Dude, quit while you're behind. ;)

Dude! It's nice being able to insult, and be insulted, while still sharing smiles.

Sincerely and seriously.

I'm not sure I'm behind. But more than happy to be corrected!

[retina]

So ignoring the "green" topic as introduced by chalsall ...

In the cold season (if you have one) place your computer(s) in your living space and use the waste heat to increase your comfort levels for "free". In the warm season (if you have one) place the computer in another room, or another building, or outside with proper environment protection, and away from your A/C intake.

Forget about the "cloud" IMO. It is troublesome to manage (your time isn't free, right?); and, unless you have really really expensive power prices, it will cost you more.

[Xyzzy]

Originally Posted by kriesel

Anything to contribute on-topic, toward the question about seasonal or long term variation in cloud computing cost for GIMIPS?

In ten years you will be able to do at lot more work for a lot less expended energy.


So wait ten years?

[GP2]

Originally Posted by kriesel

Hi,
I'm wondering if those who use cloud computing frequently have any data about seasonal or long term trends in availability or cost per unit throughput for cloud computing.

Spot prices on AWS used to have some seasonal variation. Particularly during the end-of-the-year holiday season, prices would spike. But last year it didn't happen.

Google on the other hand has fixed pricing for their preemptible virtual machines.

Availability isn't an issue, all the more so for the small one-core or two-core instance types.

Historically, cost per Mersenne exponent has declined over time. As Intel introduces new architectures (Haswell, Skylake, ...), new instance types become available. These typically have the same per-hour pricing as the older architectures but offer better performance.

I don't know if that trend will continue. What will Intel do after AVX-512? Can x86-64 CPUs offer any further major improvements to make LL testing and PRP testing faster? I suspect that GPUs are the way forward. But GPUs on the cloud are way too expensive because they are being used for machine learning by companies with very deep pockets.

[GP2]

Originally Posted by Xyzzy

In ten years you will be able to do at lot more work for a lot less expended energy.


So wait ten years?

Ten years from now you will wait ten more years.

[GP2]

Originally Posted by retina

Forget about the "cloud" IMO. It is troublesome to manage (your time isn't free, right?); and, unless you have really really expensive power prices, it will cost you more.

The whole point of the cloud is that it's less work to manage than your own on-premises hardware.

[M344587487]

Originally Posted by GP2

...
I don't know if that trend will continue. What will Intel do after AVX-512? Can x86-64 CPUs offer any further major improvements to make LL testing and PRP testing faster? I suspect that GPUs are the way forward. But GPUs on the cloud are way too expensive because they are being used for machine learning by companies with very deep pockets.

Memory bandwidth is the killer and the bottleneck worsens as core counts explode. They can rain cache everywhere and take steps to move memory closer to the cores which I think they will do as more and more workloads become memory-bound. 128MB of fast L3 cache accessible by all cores on a package would be ideal for our niche, that seems unlikely. It'll be really interesting to see the topology of zen2/Epyc2, that massive I/O die must be for something other than memory controllers and I'm hoping it's a unified cache of some variety. Down the line I think HBM2 or its successor will be used as L4 cache if not replace DDR outright, there's only so many DDR channels you can add before it physically becomes a challenge to add more.

[chalsall]

Originally Posted by GP2

The whole point of the cloud is that it's less work to manage than your own on-premises hardware.

To resonate with your point...

I sometimes have some jobs which needs a very large compute engine. But very rarely, and only for a brief period of time while I run the job. Tends to cost me only a few dollars per hour for some serious kit. Spin it up; use it; shut it down.

It would make little economic sense for me to buy the kit I needed, import it, deploy it, host it, manage it, and then use it rarely.

There is a reason both Netflix and Apple (two quick examples) are both dependent on AWS for their service offerings.

FWIW.

[kriesel]

Originally Posted by GP2

Ten years from now you will wait ten more years.

That and its implied recursion does not work well, given that human life expectancy is finite. Some of us are near, at or past retirement age.