| >The only exception is the cost of the hardware, but over the long term this is a relatively small factor compared to the cost of electricity and bandwidth. My electricity rate is is 15.7 cents per kw-hour. During typical usage (MS Office, web browser, programming), my Intel 6-core desktop (without LCD monitor on) draws about 150 watts. For back-of-the-napkin estimates, let's round the kwh cost up to 16 cents and the wattage to 300 watts (to cover scenario of some of cpu cores being 100% pegged). The electricity cost of 24x7 for one year would be ~$413. What remains would be bandwidth costs -- if any. I like many others have Verizon FiOS and even if others have Comcast or ATT, there's no obvious residential bandwidth costs I can think of to calculate. Maybe... if the homeowner wants to upgrade the speed from 75GB@$99/month to 150@$199/month because the he wants to download the datasets faster. That extra $100 wouldn't have been spent for plain web browsing. So conceivably, that would be $1200 per year. What we don't know is how big the datasets are that must be downloaded. I assume the upload size would be minimal because the compute tasks appear to be variations on "y_output = computecombinationsmontecarlobruteforce(x_input)." The y_output answer would usually be order-of-magnitude(s) smaller than x_input. Assuming there are no extra bandwidth costs, it would be hard for a company to buy computer hardware for less than a homeowner's $413/year electricity cost. Perhaps suchflex's particular business model is financially wrong. However in general terms, it does seem possible to find a monetizing sweet spot of computing tasks that takes advantage of the idle and wasted resources of existing home computers. However, if the homeowner has to buy extra hardware that was only dedicated to suchflex, that's probably where the economics won't make as much sense. |
That's the issue though, this wouldn't be similar to your typical usage. Instead, if they're using your GPU to train neural networks, it'll be running close to or at full capacity.
I realize that you rounded the costs up, but lets just look at the costs of a GPU often used for machine learning - Nvidia GTX 980 TI. According to Nvidia, it draws 250W under load which according to your figures would result in a yearly cost of roughly $344. That's just for the reference card, a typical card that a consumer would purchase would draw even more. You can buy a 980 TI for a little more than $400. That doesn't even begin to look at hardware actually designed for commercial and research applications.
I think that it's possible to find a way of monetizing computer resources, however, I think it has more to do with arbitraging differences in electricity costs. Suchflex's model certainly wouldn't work where I live (electricity costs in NYC are roughly 20 cents per kw-hour) but parts of the US are under 10 cents. I could see a company attempting to profit from these differences by setting up hardware in a cheap state and negotiating a favorable electricity rate. Heavy computation could then be done on these networks for significantly less than it could in New York or California.
In summary, the value of a consumer's unused computer has more to do with their electricity rate than their hardware.