Diceware cracking math for @thorsheim

[atom]

Basis dictionary from diceware is 7776 words and there are 5 words takes from it.

Makes a total combination keyspace of 7776^5 = 28430288029929701376

Since the diceware RNG is hopefully a perfect one we can assume to crack a passphrase on an average at 50% of the keyspace.

So we do 28430288029929701376/2 = 14215144014964850688.

A single (3 year old) hd6990 card runs with 10935 MH/s against a single NTLM hash.

So for a single card, we do 14215144014964850688/10935000000 = 1299967445 seconds.

Then, 1299967445 / 60 = 21666124 minutes.

Then, 21666124 / 60 = 361102 hours.

Then, 361102 / 24 = 15045 days.

Now, if we have 150 cards:

15045 / 150 = 100 days

Some funny theory to continue:

If we'd sacrifice some speed, let's say from 10935 MH/s down to 5042 MH/s per card, we'd able to crack 500,000 of those hashes at once!

But in this case we have to scan the entire keyspace, which is 28430288029929701376.

Makes a runtime of 400 days in total.

But, since we're cracking 500,000 in parallel, we can do:

500,000 / 400 = 1,250 cracked passphrases per day

[magnum]

Would you not end up with more like 400 days total for 500k hashes full keyspace since you dropped speed to about half? Or did I miss something.

To complete the picture it would be cool to produce ballpark figures for the cost of that cluster and for the electric bill

[epixoip]

At retail prices, you could build a proper VCL cluster with nineteen 4x 7990 compute nodes, a cluster controller, and a 24-port InfiniBand switch with required cabling for about 350,000 USD. It would require approximately 38 kW of power.

You could obviously do a milkcrate build for much less, but it would be hard to manage that many milkcrates

[magnum]

I believe 38kW of power means (very very roughly) a need for another 38kW of cooling, right? Maybe this gets too OT but it's an interesting thought.

[epixoip]

sure, but this isn't really that much for any standard datacenter to absorb. target 10kW per rack, spread the systems out across four racks (five systems or 20U per rack), shouldn't be a problem for even the smallest of datacenters.

[undeath]

you don't have 150 GPUs but 150 cards = 300 GPUs

[atom]

Would you not end up with more like 400 days total for 500k hashes full keyspace since you dropped speed to about half? Or did I miss something.

you don't have 150 GPUs but 150 cards = 300 GPUs

Fixed both, thanks!