Six hours ago I was trying to figure out CPU vs. GPU performance with memory-hard algorithms and ran into a wall when hashcat turned out not to support argon2 at all and scrypt modes only up until 16 MiB of memory.
Looking a bit deeper, someone had opened an issue about that 28 days ago and 19 days ago a patch was committed to master that should support a cost factor of up to 1M (which would use 1 GiB of memory at the recommended block size[1]). Given the slow release cycle and my having a talk on the topic next week, I downloaded the hashcat source, compiled it, struggled a bit more with the driver (switched OSes because compiling for Windows seemed like a ring of hell I wasn't yet ready for, so needed a new driver installation) and... now they just release the binary! Could have saved that effort, but oh well. Now, instead of erroring out at 16 MiB, it errors out at 110 MiB, saying not enough allocatable memory (it says on startup that the GPU has 6949M allocatable) and at 114 MiB it gives an error that, when you google or duck it, actually gives zero hits.
The speed boost of scrypt in this release is about 5 times in my benchmarks, but note that the old release was run on Windows and the new release on Linux, so driver differences probably affect that as well.
So that's my highlight of this release: better scrypt support. People that were using memory factors above 16 but below 110 MiB, beware! (Not that 64 < memory_usage < 128 MiB are valid and usable in many (most?) implementations, but hashcat runs with it.)
Of course, the real strength is in the password, but you know your users...
Looking a bit deeper, someone had opened an issue about that 28 days ago and 19 days ago a patch was committed to master that should support a cost factor of up to 1M (which would use 1 GiB of memory at the recommended block size[1]). Given the slow release cycle and my having a talk on the topic next week, I downloaded the hashcat source, compiled it, struggled a bit more with the driver (switched OSes because compiling for Windows seemed like a ring of hell I wasn't yet ready for, so needed a new driver installation) and... now they just release the binary! Could have saved that effort, but oh well. Now, instead of erroring out at 16 MiB, it errors out at 110 MiB, saying not enough allocatable memory (it says on startup that the GPU has 6949M allocatable) and at 114 MiB it gives an error that, when you google or duck it, actually gives zero hits.
The speed boost of scrypt in this release is about 5 times in my benchmarks, but note that the old release was run on Windows and the new release on Linux, so driver differences probably affect that as well.
So that's my highlight of this release: better scrypt support. People that were using memory factors above 16 but below 110 MiB, beware! (Not that 64 < memory_usage < 128 MiB are valid and usable in many (most?) implementations, but hashcat runs with it.)
Of course, the real strength is in the password, but you know your users...
[1] 128 * r * N, where r is the block size (recommended to leave it at 8) and N the cpu/mem cost factor, via https://stackoverflow.com/a/30308723/1201863