![]() There are some great resources (including those linked above) that can show how fast passwords would conceivably fall with this level of performance. For a nominal increase in cost (approximately 10% more) simply getting bigger GPUs is worth the nominal investment. We checked the summary numbers against other completed runs.įor those who want the summary, this system is about 25% faster than an 8x NVIDIA GTX 1080 machine. Here is a raw output of one of the runs we had. ![]() 8x NVIDIA GTX 1080 Ti Hashcat Benchmark Results We would certainly call this a reproducible anomaly but it is likely a driver/ software version issue. This same system can run deep learning training models for weeks without issues. On three separate occasions, we ran the benchmarks and saw the machine eventually crash and need to be restarted. We fit hashcat benchmarking times between running other workloads. As one may imagine, a system that uses more than $1/ hour in northern California electricity is expensive to leave workloads unscheduled on. Second, we do not have a full set of benchmark results. We tried this on three different days and each time we saw between 3.1kW and 3.2kW. The figure of just shy of 3.2kW was achieved using hashcat benchmarking. When we fired up hashcat we saw major power spikes: DeepLearning10 Hashcat Power Peak We can see machines run in that range for days on end while training models. This system will train most deep learning models and other GPU compute workloads we run in the 2.4kW to 2.6kW range on a fairly constant basis. Uncharacteristic power consumption and stability.įirst, power consumption was much higher. The number one result was that some of the algorithms hashcat tests significantly stress the GPUs. Our test procedure was to utilize the latest 381 series NVIDIA drivers haschcat 3.40 and a simple benchmark command and see what would result.īefore we get to the performance numbers, we did have a few observations. We did absolutely zero special setup on this machine. Swordfish Movie (2001 Era) Encryption using a similar system and 8x AMD FirePro S9150 cards. We actually did a post and video a long time ago about Today’s (mid-2016) Hardware v. One of the interesting features of these large multi-GPU systems is that there are a number of intriguing workloads you can run for $15,000 or so beyond just deep learning. Supermicro 4028GR TR Red V Black NVIDIA GTX 1080 Ti 8x GPU If you try this, get the FE cards not the 2/ 3 fan variants. The black SLI covers are Gigabyte and Zotac versions on the CPU1 PCIe root. ![]() ![]() The red SLI covers are MSI GTX 1080 Ti Founders Edition cards on CPU2’s PCIe root. The GTX 1080 Ti is the go-to value card for deep learning at the moment. At the heart of the system are 8x GTX 1080 Ti’s. That sounded fun, and while we are waiting for results of the 10x GTX 1080 Ti “DeepLearning11” machine, we decided to use our DeepLearning10 build to with 10x NVIDIA GTX 1080 Ti’s.įor the quick background, DeepLearning10 is a build that we made specifically for deep learning work, e.g. In that post, a password cracking tool ( hashcat) was cited with 8x NVIDIA GTX 1080 8GB cards and some impressive numbers put forward. ![]() This post was inspired by Jeff Atwood’s work seeing how secure passwords are using “low cost” commercially available systems. One area that is particularly fascinating with today’s machines is password cracking. ![]()
0 Comments
Leave a Reply. |