[evervevdww221]

$2 per hour per machine is not a made-up number, just check how much Amazon charges for a GPU vm.

Again, we were a small company, we had to use other cloud providers. I don't know why amazon or alike are so expensive. We tried colocation too, not any cheaper to be honest.

Also, when buying a computer or a console, you don't pay by usage. you pay one time to own the hardware and the remaining 5 year usage is free, plus power bills no one really cares. 5% utilization times 5 year is still some decent hours, the cost is not necessarily more than a cloud vm.

For a cloud vm, you pay by usage. The more you use, the more you pay, the overall payment is not capped. Eventually the cost will surplus that of the hardware. And I'm saying "Eventually" is actually a month.

[mantap]

Amazon is a bad fit for this use case. Amazon's value proposition is: "you are making money, pay us a big chunk of it and we will help you scale up with less effort and fewer staff.

It doesn't seem surprising to me that making a cloud gaming service would necessitate assembling your own servers. Games just have different hardware requirements to everything else, it's well known that "pro" graphics cards are not meant for gaming.

[aeturnum]

Not Op, but they specifically said the using a colocation (i.e. setting up your own servers) was not helpful in keeping down costs.

[mantap]

I know but I find it hard to believe that Colo provides no benefit. I'm guessing a large part of their AWS bill would be bandwidth since this kind of a pathological use case for AWS (streaming individual video with no possibility to use CDN).

[Tsarbomb]

> And I'm saying "Eventually" is actually a month.

I was the technical lead on a project where I warned the product owner that implementing a certain feature the way they wanted in AWS would have astronomical costs. I was told "let me worry about the money".

A month later we were told to scale down our ECS and EMR usage as the bill was astronomical.

[akvadrako]

I assumed you were not using Amazon because it's obviously a bad choice for this use case. $2/hour is $87k over 5 years — not surprising you can't make money.

One can rent a rack for about $3k a year which comfortably fits 20 servers. Obviously, colocation is much cheaper.

[HeWhoLurksLate]

You've also gotta get servers to put in there, and in a 42u rack with, say 9 4u servers, a 2u managed switch, and now you have 4 units left to handle power distribution, power backups, any kind of external remote management (say an out-of-band KVM, or a second network for IPMI traffic, or anything else).

ALSO, all those servers? Say the fancy, high-speed switch you need to route all that traffic is $2000.

You wind up needing spend $5000 on auxiliary equipment and installation. This includes IE power cables, network cables, Velcro, and that darn cable you forgot you needed.

Now we're up to $10,000 including the colo costs, and we haven't even gotten to servers yet.

We need 9 4u-tall servers. The reason for the 4u height is because that's the size you want to be the most space-efficient with your full-height GPU's.

You'll want good base servers to slap your graphics cards into. I'm a little out of the loop on the latest and greatest in the server world, but we'll assume it's around $7,500 for a fairly moderate AMD EPYC system (EPYC because A] they wind up being cheaper than their Intel counterparts and B] have many more PCIE lanes).

9 $7,500 servers is $67,500 without any GPU's.

You don't want nVidia's consumer GPU's, because they're hostile to virtualization, so it's either AMD's consumer or enterprise stuff Or nVidia's enterprise stuff.

For GPU's you want to use, you'll be paying at least ~$600 per, and there's around 8-10 slots per server. $600 * 8 * 9 = $43,200.

I'm sure I've missed stuff- haven't included data storage, for one- but we're already at $120,700 and you'll probably want some new GPU's in a couple of years, and the total cost of servers over their lifespan winds up being around double what the initial cost was.

-Summer Glau

[akvadrako]

I wasn’t suggesting doing it this way. I said to use the equivalent of having a system at home: no virtualization, consumer GPUs. That way it’s an apples to apples comparison - you just get higher utilization and cheaper power. The cost of the switch and rack are minor when divided over 20 systems.

[hrgiger]

AMD consumer GPUs doesnt support SR-IOV probably something like this https://www.amd.com/en/graphics/workstation-virtual-graphics