Description

The world's most popular Machine Learning Computer. This is the official Lambda GPU DevBox Build. You can purchase direct from Lambda here:

https://lambdal.com/nvidia-gpu-workstation-devbox

Log in to rate comments or to post a comment.

Comments

  • 35 months ago
  • 1 point

Those temps are gonna be amazing...

  • 35 months ago
  • 2 points

We run a 50 GPU cluster and temps are fine in a climate controlled environment.

  • 17 months ago
  • 1 point

Blower style cooler, so temps should be fine

  • 33 months ago
  • 1 point

While the motherboard supports all those GPUs ("quad x16/x16/x16/x16"), the CPU does not ("Max # of PCI Express Lanes: 40"). As I understand it, that means this machine runs one GPU at 16x and the others at 8x. Is that correct?

  • 30 months ago
  • 3 points

Incorrect. This motherboard has PLX switches which allow the cpu to switch between all 4 GPUs and provide 16x speed to any one of the GPUs. That said, you are correct that the maximum throughput would still be 32x speed from root to leaf. Essentially what the PLX switches do is that during the device enumeration step all PCIe devices are recognized as 16x devices and the PLX switches handle the routing so that the CPU can provide up to 16x speed for each device at any point in time. Without PLX switches you would have a hard wired setup with something like 16x/8x/8x/8x but in this case it enumerates @ 16x for all GPUs.

  • 21 months ago
  • 1 point

The build has incompatible warning as the following. Should I be concerned?

Some Intel X99 chipset motherboards may need a BIOS update prior to using Broadwell-E CPUs. Upgrading the BIOS may require a different CPU that is supported by older BIOS revisions.

  • 19 months ago
  • 1 point

I see how you keep the thermals down: you're using blower style cards. Interesting. Looks like the 6850k has only 40 pcie lanes which means all the gpus would have to be running in x8 mode. This would leave 4 lanes for the m2 and 4 lanes for the south bridge.

Do you have any benchmarks on a dataset like image net to quantifying the performance hit? Also given that you have a 1600 watt PSU have you increased the voltage limit on the gpus to 300 watts and once again benchmarked on image net?

Also it would appear that changing out the mobo to a zenith extreme and a threadripper 1950x wouldn't really increase the cost of the system, basically the cost of a good air cooler. Such an upgrade would have enough pcie lanes for two of the cards to run at x16. Have you guys thought about doing that?

Also thanks for posting the parts list! It's reassuring to see something that many folks think is not possible being built and run successfully.

  • 35 months ago
  • 0 points

this has to be a troll

  • 35 months ago
  • 2 points

Why would this be a troll?

  • 35 months ago
  • 2 points

im talking abou that intel cooler

  • 30 months ago
  • 1 point

Having fancy cooling on your CPU is irrelevant for deep learning.

  • 20 months ago
  • 2 points

What is tbe CPU cooling like? Do you have any temps figures?

  • 28 months ago
  • 0 points

PcpartPicker says that Carbide Series® Air 540 Arctic White is not compatible with Asus - X99-E WS/USB 3.1 SSI CEB LGA2011-3 Motherboard. Did you have to do some custom work?

  • 28 months ago
  • 3 points

X99-E WS/USB 3.1 works with the AIR 540 with no custom work required.

  • 21 months ago
  • 0 points

i7-6850K supports RAM only up to DDR4-2400, the DDR4-2666 is a little bit of waste I guess. https://ark.intel.com/products/94188/Intel-Core-i7-6850K-Processor-15M-Cache-up-to-3_80-GHz