The real benefit of the AWS solution is that you could rent a large number of AWS nodes and run the calculations to completion right now if you wanted to. As in, literally tonight. And you wouldn’t have to source the hardware, assemble machines, install software, power it all, cool it, figure out all of the power complexities and so on.
> There is no runtime available right now. Please change the compute type or try again later.
recently, that's not always true.
It's always funny to me when you run into a wall of "no more instances" in AWS. It doesn't happen often unless you're at a massive scale, but it does happen surprisingly often with the GPU types
It's pretty sad how Google/Amazon can't at least temporarily move their own workloads to free up resources for their public clouds while they install new hardware.
That's likely not the reason. It's more likely caused by inherent spike-y ness of the traffic, given that most GPU instances are probably used for batch numerical computations (like training neural networks). The pricing incentivizes using a lot of nodes for a shorter amount of time, much like in the SHA example.
If their own public services slowed down in high demand times, it wouldn't look good. The sales channels could try to spin it as “giving you priority over us” but who would genuinely believe them.
They could reduce resources available to low priority internal stuff during spikes of other activity, but that would not have a large effect considering the combined size of activity on the public cloud resources.
People also overestimate how much effort it is to manage hardware, especially for raw compute (ie 99% SLA fine) operations.
The cooling and power is handled by your datacentre. You don't need to deal with that.
I've built a 200 GPU mining operation in one workweek, pay me $170k?
And that is just raw uptime, as an end-user you also have to live with you jobs waiting for the scheduler to fit them into the queue. A day or two waiting time is pretty common on larger jobs.
I dig the cloud for production, but for R&D I like my people to have fixed costs and close to zero marginal costs.
But you are right that the marginal costs will be mostly made up of opportunity costs.
Unless you have only one researcher, it doesn't matter for an individual researcher whether their marginal costs are opportunity costs of their co-workers or AWS bills.
(And if you don't have any mechanism for your individual researchers to feel the opportunity costs of their co-workers; that is the same thing as saying that one researcher can hog up all the compute time.
Social pressure is also a way to make them feel opportunity costs.)
Legally, even money you have in your account now might not stay there, if there was fraud involved. But the mined coins will be gone (and the compute time, too).
I imagine it's the cause of several support tickets along the lines of "I got hacked and racked up a $50,000 bill because of bitcoin mining"
I didn't do the math, just my feeling.