Re:Considerations when setting up deep learning hardware
by sgt101 on Monday June 13, 2016 @02:48PM
(#11896279)
I was a bit disappointed as my questions are more "what is important about a deeplearning box?" How much DRAM should I get ? What should I do about disks?
We built one recently, I got a big lump of dram (256GB), a 12 core processor 6 cards and a 2TB ssd on pci + some ssd's as a backend. We then mapped it onto HDFS, and our pipeline is from HDFS to the ssds for project work and then when the data's ready you can map it onto the pci disc for the actual run - hopefully the cpu has enough cores to run the discs and the GPUs at once.
Re:Considerations when setting up deep learning hardware
by KaiserPro on Monday June 13, 2016 @03:13PM
(#11896500)
If you're building one yourself then this might be useful:
Make sure you have enough PCI lanes. Motherboards might say they have x16 lanes for all four slots, but in practice it can drop down to as little as 4 each.
Unless something has changed recently, dual processor machines suffer a hit in speed when transferring data to and from GPU memory. However They might have fixed card affinity.
Each graphics card should be able to pull in at least 2 gigabytes a second. You therefore need to max out system ram (assuming your dataset changes that much....)
Power.
You need a PSU that is really up to it. Your wiring needs to be over specced.
Cooling.
4x 400 watts, plus 2x 150 watts for each proc, is a lot of power to dissipate. PSU efficiency really saves money
Re:Considerations when setting up deep learning hardware
by vonnik on Monday June 13, 2016 @02:55PM
(#11896337)
Anybody curious about this, especially as it pertains to production systems and the JVM, should join the Deeplearning4j gitter channel. We deal with those issues from time to time. https://gitter.im/deeplearning4j/deeplearning4j
Re:Considerations when setting up deep learning hardware
by brandmeyer on Monday June 13, 2016 @05:31PM
(#11897860)
Tiny detail when shopping for UPS: It is the kW rating that is important. Some UPS, especially at the small end, will have an artificially (marketing inflated) large kVA rating compared to the kW rating. Modern computer PFC power supplies (just about anything with an 80 Plus rating) will draw kW == kVA.
We built one recently, I got a big lump of dram (256GB), a 12 core processor 6 cards and a 2TB ssd on pci + some ssd's as a backend. We then mapped it onto HDFS, and our pipeline is from HDFS to the ssds for project work and then when the data's ready you can map it onto the pci disc for the actual run - hopefully the cpu has enough cores to run the discs and the GPUs at once.
Make sure you have enough PCI lanes. Motherboards might say they have x16 lanes for all four slots, but in practice it can drop down to as little as 4 each.
Unless something has changed recently, dual processor machines suffer a hit in speed when transferring data to and from GPU memory. However They might have fixed card affinity.
Each graphics card should be able to pull in at least 2 gigabytes a second. You therefore need to max out system ram (assuming your dataset changes that much....)
Power.
You need a PSU that is really up to it. Your wiring needs to be over specced.
Cooling.
4x 400 watts, plus 2x 150 watts for each proc, is a lot of power to dissipate. PSU efficiency really saves money
Where can I get one for next week?