New AWS G3 Instance Lineup

Amazon has released their new G3 instances. Utilizing a Tesla M60 GPU platform (http://www.nvidia.com/object/tesla-m60.html) and scaling up to 4 GPU's. Along with this you also get 64 CPU's and almost half a TB of RAM. I figured it was time to crunch some numbers on Hashcat performance of the current AWS cloud options.

For all the new G3 M60 GPU benchmarks, please visit (https://gist.github.com/binary1985/881c2df6310659413102eaf5e349e999)

Current state of GPU Hash Cracking

I have been following GPU performance trends since it first became popular with bitcoin. A recent trend in the cloud has me optimistic that "some" day high intensity GPU work loads will make fiscal sense to run cloud based. The last time I checked the current state of AWS GPU Hashcat performance I was disappointed in the findings however since then 2 new lines of GPU instances have come into play. Lets take a look at the current top tier options, using on demand pricing structures.

G2.x16(4 GPU)
NTLM Performance ~ 16GH
SHA Performance ~ 3GH
Price Per Hour = $2.60

G3.x16(4 GPU)
NTLM Performance ~ 73 GH
SHA Performance ~ 16 GH
Price Per Hour = $4.56

P2.x16(16 GPU)
NTLM Performance ~ 136
SHA Performance ~ 32
Price Per Hour = $14.40

So if we take the Cost / NTLM Performance:
G2.x16 = $0.1625  per GH/hr
G3.x16 = $0.0624 per GH/hr
P2.x16 = $0.1058 per GH/hr

So its nice to see that the new G3.x16 instances are moving in the correct direction, however its not close to matching modern desktop GPU based crackers. Using a Nvidia 1080ti the cost per GH is $13.18, which means you would need to run only 250 hours to balance out the purchase. This excludes costs associated with power, cooling, and other hardware.

$699/53.1 GH ~ $13.18
$13.18/$0.0624 = 211hr

If I had to estimate, I would balance our current cracker at being used 80% of the time by at least one consultant. As such we use our system about 7000 hours a year, using this model we can backwards compute what an acceptable pricing model would be for cloud infrastructure rental to outweigh a purchase.

$13.18/x(unknown cost)=7000hr
x(cost)=$0.00188 per GH/hr

As expected, homegrown cracking solutions are the way to go from a cost analysis point of view. While my formulas left out power and maintence of hardware, we still have a huge rift between the economics of purchasing a GPU rig, versus using cloud options. It might be that you don't perform password cracking often, if you do less than 211 hours of cracking a year it may be cheaper for you to use cloud resources. You should be able to use my formulas to determine what is right for you.