Machine Learning build

[Dieselboy]

Built the 2nd system for machine learning. One of the things that's amazed me about this build is the Samsung 960 SSD drive (nvme). The thing is about as big as a stick of ram, may be shorter and a few mm thicker. around 2000mb/s read and write and 100,000 iops. done some copy tests and I'm getting 1GB/s in Windows 10 between disks.

The difficult thing was getting the three GPUs squeezed in. Running load tests on it now doing SETI and Einstein at the same time. The system flies through the work units. For example, my own GTX560ti GPU that I bought back around 2010 has 380 cuda cores and takes the gpu around 45 mins after a bit of tweaking to compute 1 work unit. This system is doing three gpu units simultaneously and is finishing them off in under four minutes. Each gpu has little over 3800 cuda cores (near 12,000 in total for the one system) On top of that there's 40 cpu threads.

It does zap a bit of power though. I see about a 2A increase in UPS load when the system ramps up to 100% on all cores.

The thing is a beast and not even seen minesweeper 

[deanwebb]

Interesting note about the 2A increase... is that within stated limits on the power draw or is that in excess of what is indicated on the power info for the device?

[Dieselboy]

It's all custom built by myself. I have a 1500W power supply as it has dual Xeon CPU, 8 sticks of RAM (128GB), 2 storage drives, a whole bunch of fans (8 in total including the CPU fans) and 3 x NVidia titan Xp GPUs which each of them take up two power feeds from the PSU.

2A @ 230v = about 460W so well within the PSU limit. Before I spec'd it, I put all the components in a PSU calculator and I worked out that the 1500W would be the safest.

[deanwebb]

Good to know... power calculations can be very critical.