minerpart - this is one reason why I (and several others) think you need to step back. you lose any shred of credibility when you try to argue with an expert.
I showed his statement which made the 1-2 PH plan look void to be 100% incorrect, with hard evidence. Thanks for the support.
Hard evidence? Come on now, that's just silly. I didn't say it was void, just that getting it done in Q2 when the chips are slated to arrive in Q2 is next to impossible.
Let's play with some numbers, shall we? Please fill in the blanks. Let's assume for a minute that this farm will be going into a large datacenter already, and that Ken will just be able to show up with the rack mount equipment ready to be bolted into place.
How long do you think that installation will take?
Before he can do that though, all those machines need to be configured. We're talking about on the order of 1000 4U rack cases here assuming 1TH/s per unit. How long will that take?
Before that can happen though, he needs to build those 1000 4U cases up and test them. How long will that take?
I would posit that from the time cases are ready to be built to the time the whole system is live hashing will not be less than a month. Feel free to provide your own numbers.
Before that happens, he needs boards which will need to be tested. Let's just spitball that he's got a very high density system and has 32 chips per board for ~61GH/s, with 16 boards per 4U chassis. That's probably optimistic, but we'll roll with it. That would be 1TH/s per 4U at about 2kW which works nicely. To build out 1PH/s, you would need 16,000 boards. Even with automated programming and test jigs, how long will that take to do?
Of course, before that you need to assemble these 16,000 boards. Large assembly houses can process a huge volume, but setup and inspection does not happen instantaneously. How long will assembly take?
Before assembly can happen though, you need all the components. These can be done in parallel, but not before the board/system is tested and finalized.
1. Boards
2. ASICs
3. Everything else
You can say that the boards could be ready beforehand, but ordering 16,000 PCBs before the design is tested is foolhardy. You're not going to be able to place your volume order until after you have tested the complete system. What's the lead time on an order of 16,000 PCBs?
Likewise the ASICs. To get 1PH/s you're talking about over half a million chips. If you're getting around 4k chips per wafer, that's 128 wafers. I think that many wafer starts with a completely untested design is a bit unwise anyway, but let's say you do it. You would almost certainly wait on metalizing most of those wafers until you have preliminary functional tests so you don't have to scrap hundreds of thousands worth of chips if there is an issue. How long will it take to finish the wafers, saw and package them once testing is done?
For everything else, what do you think your lead times are there? You can order all the components in advance assuming the board design will just work, so let's say 0. Otherwise, expect 6-8 weeks if you're ordering something like 64,000 power inductors for the DC/DC convertor.
So, for the above items from a finalized design to having 16,000 hashing boards done (plus whatever ancillary boards are needed), what are your time estimates on that?
Before that you need to have a finalized design. This is going to be far and away the most variable part of the deployment. Best case, from the time you first get sample chips to when you sign off on the design might be a week if you're reckless and extremely lucky. That's the absolute best case. More likely you're looking at several weeks including debugging software and a couple board spins. Remember this isn't even like KnC where they were shipping early units a week later despite issues with boards blowing up due to power supply issues that got fixed on a later spin. If you're compressing the schedule as much as we're trying here, you have to torture test that design or face writing off millions of dollars if issues pop up that you didn't catch in your couple days of testing.
So, when you add that all together, what kind of time from chips off the line to 1PH/s hashing do you see? Note there's no real allowance for anything going wrong here, if you want to you should probably add in a factor to account for not everything going perfectly.
Keep in mind this is all entirely contrived. There's a reason why large deployments like this are staged, the entire thing becomes unmanageable if it gets too large. Doing it based around an entirely untested design is insanity. I would actually be more worried if Ken said he was immediately ordering 1PH/s (and millions of dollars) worth of hardware in the first batch before field deploying anything than if he said it will take a few months from the time the first hardware is online to when the mine is fully deployed.
Wow, a 1 million square foot DC is an absolutely huge facility. I might be out of date, but that would have been one of the largest in the world just a couple years ago. That's going to be a multi-billion dollar renovation. I'm sure a facility like that could definitely squeeze in an extra couple MW without batting an eye.