Well, the specific part that was wrong was "GPU always performs better than CPU, even on slow hashes". Slow hashes are especially faster on GPU than CPU (until you start involving memory, that is) because the hash algorithm is slow enough that even a few candidates will keep the GPU loaded, so you don't have to give them much work to gain acceleration.
In reality, it's actually fast hashes & anything involving memory that you have to be concerned about. For example, all of the following are faster on CPU than GPU:
- Attacks without amplifier against fast hash (e.g. straight wordlist or pipe against MD5, MD4/NTLM, SHA1, SHA2, SHA3, etc. )
- Attacks with high hit rate (hundreds+ cracks per second) against medium-to-large sized lists (results in too frequent hash table lookups & collisions + large host-device transfers, in addition to being IO bound from writing to potfile, and god help you if you're also using --remove)
- Very large hash lists (steep drop in bitmap/hash table efficiency + taxing the card's memory controller)
- bcrypt (at one point high-end CPUs and GPUs were about equal for single hash, but CPU still won multi-hash and even a handful of bcrypt hashes can make an ASIC hang.)
- scrypt when 128*r*N >~ 1MiB (I'm actually not entirely sure where the precise tipping point is, but N=16384,r=8,p=1 will definitely make an ASIC hang with even a single hash.)
I know there are more examples, but fuck it I'm drunk, so you take what you get AND YOU LIKE IT
Also keep in mind the architecture changes in v2.10 will allow you to use both CPU & GPU simultaneously, so yeah, CPU kind of matters now as it can provide a decent boost in performance -- especially when you have multiple systems clustered, you can get like another GPU's worth of performance out of your CPUs. For example, http://pastebin.com/MY2BrmEA -- that's 6 GH/s NTLM on CPU alone, which is faster than some of the GPUs you guys are still using!
In reality, it's actually fast hashes & anything involving memory that you have to be concerned about. For example, all of the following are faster on CPU than GPU:
- Attacks without amplifier against fast hash (e.g. straight wordlist or pipe against MD5, MD4/NTLM, SHA1, SHA2, SHA3, etc. )
- Attacks with high hit rate (hundreds+ cracks per second) against medium-to-large sized lists (results in too frequent hash table lookups & collisions + large host-device transfers, in addition to being IO bound from writing to potfile, and god help you if you're also using --remove)
- Very large hash lists (steep drop in bitmap/hash table efficiency + taxing the card's memory controller)
- bcrypt (at one point high-end CPUs and GPUs were about equal for single hash, but CPU still won multi-hash and even a handful of bcrypt hashes can make an ASIC hang.)
- scrypt when 128*r*N >~ 1MiB (I'm actually not entirely sure where the precise tipping point is, but N=16384,r=8,p=1 will definitely make an ASIC hang with even a single hash.)
I know there are more examples, but fuck it I'm drunk, so you take what you get AND YOU LIKE IT
Also keep in mind the architecture changes in v2.10 will allow you to use both CPU & GPU simultaneously, so yeah, CPU kind of matters now as it can provide a decent boost in performance -- especially when you have multiple systems clustered, you can get like another GPU's worth of performance out of your CPUs. For example, http://pastebin.com/MY2BrmEA -- that's 6 GH/s NTLM on CPU alone, which is faster than some of the GPUs you guys are still using!