Difference between revisions of "GPU Build"
Jump to navigation
Jump to search
| Line 30: | Line 30: | ||
====TL;DR==== | ====TL;DR==== | ||
Using multiple GPUs adds a lot of complexity. It has a few benefits: possible speed ups if the network can be split up (and is big enough), able to train multiple networks at once (either copies of the same network or modified networks), more memory for huge batches. Some frameworks have much better performance with multiple GPUs (pytorch, caffe 2) while others are catching up. | Using multiple GPUs adds a lot of complexity. It has a few benefits: possible speed ups if the network can be split up (and is big enough), able to train multiple networks at once (either copies of the same network or modified networks), more memory for huge batches. Some frameworks have much better performance with multiple GPUs (pytorch, caffe 2) while others are catching up. | ||
| − | |||
| − | |||
==RAM== | ==RAM== | ||
| Line 57: | Line 55: | ||
* [http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ A Full Hardware Guide to Deep Learning] | * [http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ A Full Hardware Guide to Deep Learning] | ||
* [https://www.oreilly.com/learning/build-a-super-fast-deep-learning-machine-for-under-1000 Cheap build] | * [https://www.oreilly.com/learning/build-a-super-fast-deep-learning-machine-for-under-1000 Cheap build] | ||
| − | * [https://medium.com/@SocraticDatum/getting-started-with-gpu-driven-deep-learning-part-1-building-a-machine-d24a3ed1ab1e] | + | * [https://medium.com/@SocraticDatum/getting-started-with-gpu-driven-deep-learning-part-1-building-a-machine-d24a3ed1ab1e How to build a GPU deep learning machine] |
| + | |||
| + | |||
| + | Questions to ask: | ||
| + | * Ask what libraries will be used | ||
| + | * Approx. dataset/batch size | ||
| + | * Development approach regarding multiple GPUs; splitting up large models, training multiple models | ||
| + | * Network card? | ||
| + | * DVD drive? | ||
Revision as of 16:29, 23 October 2017
| GPU Build | |
|---|---|
| Project Information | |
| Project Title | GPU Build |
| Owner | Oliver Chang, Kyran Adams |
| Start Date | |
| Deadline | |
| Primary Billing | |
| Notes | |
| Has project status | Active |
| Copyright © 2016 edegan.com. All Rights Reserved. | |
PC Partpicker build missing a lot of pieces
Motherboard/CPU/Fan
- Not a huge deal, but used for data preparation
- If using multiple GPUs, at least one core (two threads) per GPU
- Should have enough PCIe slots
- Motherboards: ASUS Z10PE-D16 [1], Dual LGA 2011 R3, DDR4 - Up to 32GB RDIMM, 16 slots
- Chips: Intel Haswell Xeon e5-2620v3, 6 core @ 2.4ghz, 6x256k level 1 cache, 15mb level 2 cache, socket LGA 2011-v3 [2]
- CPU Fans: Intel Thermal Solution Cooling Fan for E5-2600 Processors BXSTS200C [3]
GPU
- "I quickly found that it is not only very difficult to parallelize neural networks on multiple GPUs efficiently, but also that the speedup was only mediocre for dense neural networks. Small neural networks could be parallelized rather efficiently using data parallelism, but larger neural networks... received almost no speedup."
- Possible other use of multiple GPUs: training multiple different models simultaneously, "very useful for researchers, who want try multiple versions of a new algorithm at the same time."
- This source recommends GTX 1080 Tis and does cost analysis of it
- If the network doesn't fit in the memory of one GPU (11 GB),
- Might want to get two graphics cards, one for development, one (crappy card) for operating system
- Intra-model parallelism: If a model has long, independent computation paths, then you can split the model across multiple GPUs and have each compute a part of it. This requires careful understanding of the model and the computational dependencies.
- Replicated training: Start up multiple copies of the model, train them, and then synchronize their learning (the gradients applied to their weights & biases).
TL;DR
Using multiple GPUs adds a lot of complexity. It has a few benefits: possible speed ups if the network can be split up (and is big enough), able to train multiple networks at once (either copies of the same network or modified networks), more memory for huge batches. Some frameworks have much better performance with multiple GPUs (pytorch, caffe 2) while others are catching up.
RAM
- At least as much RAM as GPUs, probably twice that much (2 * Number of cards * 11 GB [GTX 1080 Ti size])
- RAM: Crucial DDR4 RDIMM [4], 2133Mhz , Registered (buffered) and ECC, comes in packs of 4 x 32GB
PSU
- Some say 1.5x-2x wattage of GPU+CPU, some say GPU+CPU+100W
- PSUs: Corsair RM Series 850 Watt ATX/EPS 80PLUS Gold-Certified Power Supply - CP-9020056-NA RM850 [5]
Storage
- M.2 Drives: Samsung 950 PRO -Series 512GB PCIe NVMe - M.2 Internal SSD 2-Inch MZ-V5P512BW [6]
- Solid State Drives: Intel Solid-State Drive 750 Series SSDPEDMW400G4R5 PCI-Express 3.0 MLC - 400GB [7] or 800GB [8]
- Regular Hard drives: WD Red 3TB NAS Hard Disk Drive [9] - 5400 RPM Class SATA 6 Gb/s 64MB Cache 3.5 Inch
Misc.
- Cases: Rosewill 1.0 mm Thickness 4U Rackmount Server Chassis, Black Metal/Steel RSV-L4000[10]
- DVDRW (Needed?): Asus 24x DVD-RW Serial-ATA Internal OEM Optical Drive DRW-24B1ST [11]
- Keyboard and Mouse: AmazonBasics Wired Keyboard and Wired Mouse Bundle Pack [12]
Other Builds/Guides
- Deep learning box for $1700 (links to https://blog.slavv.com/the-1700-great-deep-learning-box-assembly-setup-and-benchmarks-148c5ebe6415)
- A Full Hardware Guide to Deep Learning
- Cheap build
- How to build a GPU deep learning machine
Questions to ask:
- Ask what libraries will be used
- Approx. dataset/batch size
- Development approach regarding multiple GPUs; splitting up large models, training multiple models
- Network card?
- DVD drive?
