Cryptohaze.com

http://stacksmashing.net/2010/11/15/cra ... instances/



Just wait until the multiGPU support comes out with Fermi optimizations...

So, slight fail...

Their Fermi cards are 448 cores at 1.15ghz - even slower than a 470...

Very poor numbers for the Tesla EC2:

[root@ip-10-17-130-104 CUDA-Multiforcer-Linux-0.72]# ./CUDA-Multiforcer -d 0 -l -h NTLM -f hash.txt –min=8 –max=8 -c charsets/charset-upper-lower-numeric-symbol-95.chr -o outhash.txt –threads 768 –blocks 60 -m 1024
Version 0.72, length 0-14
CUDA Device Information:
Device 1: “Tesla M2050″
Number of cores: 112
Clock rate: 1.15 GHz
Performance Number: 16058
Hash type: NTLM
Hashes loaded (22 hashes)
Launching kernel for password length 8
Done: 0.00% Step rate: 1110.7M/s Search rate: 24435.0M/sec

NOTE: This is with the new V0.72 NOT the V0.70 in report

Code: Select all

yum install automake autoconf gcc44-c++
wget  http://developer.download.nvidia.com/compute/cuda/3_2/sdk/gpucomputingsdk_3.2.12_linux.run
sh gpucomputingsdk_3.2.12_linux.run
rpm -Uvh --nodeps --provides ftp://ftp.pbone.net/mirror/archive.fedoraproject.org/fedora/linux/releases/10/Fedora/x86_64/os/Packages/libstdc++-4.3.2-7.x86_64.rpm
wget http://cryptohaze.com/releases/CUDA-Multiforcer-Linux-0.72.tar.bz2
tar xvf CUDA-Multiforcer-Linux-0.72.tar.bz2
cd CUDA-Multiforcer-Linux-0.72
./CUDA-Multiforcer -d 0 -l -h NTLM -f hash.txt –min=8 –max=8 -c charsets/charset-upper-lower-numeric-symbol-95.chr -o outhash.txt –threads 768 –blocks 60 -m 1024


Cost Breakdown (above variables):
$2.20 /hour
1569 hours until completion
--------
$3,551.80

NOTE: You can cut that price in half when Multi-GPU support is added to Multiforcer

Note the --threads 768, 64-bit gets bitchy with "--threads 1024" so I had to throttle back. 32-bit will take --threads 1024, but is still a bit slower in tests

Plus, at some point, some sse2 magic thrown in. The ec2 boxes have decent CPUs too.

@hackjar, i think the GTX470/80 is faster than the tesla M2050 and costs like 10% of it.

So, a few thoughts:

- The Fermis they are using *are* slower than a 470. Same shader count, slower clock. Theoretically, lower power consumption, allowing them to be passively cooled (the Fermi Teslas I've seen have a *giant* heatpipe covering the entire card, designed for low profile server cooling).

- They will still be usable for smaller password audits or wordlists.

- They have the potential to be *very* useful for rainbow tables for those who lack the bandwidth or GPUs to effectively use mine.