|Has title||GPU Build|
|Has owner||Oliver Chang, Kyran Adams|
|Has start date|
|Has deadline date|
|Has project status||Active|
|Has sponsor||McNair Center|
|Has project output||Content|
|Copyright © 2019 edegan.com. All Rights Reserved.|
To start with we are trying to use our existing ASUS Z10 server board, rather than switching to the Asus X99-E WS workstation class motherboard, and rather than Four TITAN X GPUs, we've got a TITAN XP and a TITAN RTX.
Note that the Asus X99-E WS is available from NewEgg for $500 now.
Single vs. Multi GPU
- GTX 1080 Ti Specs
- Since we are using Tensorflow, it doesn't scale well to multiple GPUs for a single model
- Which GPU for deep learning (04/09/2017)
- "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),
- Want to get two graphics cards, one for development, one (crappy or onboard 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).
Pros of multiple GPUs:
- Able to train multiple networks at once (either copies of the same network or modified networks). Allows for running long experiments while running new ones
- Possible speed ups if the network can be split up (and is big enough), but tensorflow is not great for this
- More memory for huge batches (not sure if necessary)
Cons of multiple GPUs:
- Adds a lot of complexity.
- NVLink can link between CPU and GPU for increase in speed, but only with the CPU IBM POWER8+.
- NVLink can link between GPU and GPU as a replacement for SLI with other CPUs, but this is not super relevant to tensorflow, even if trying to parallelize across one model.
- This source says to get the 1080 because the K80 is basically two K40s, which have less memory bandwidth than the 1080. This source agrees.
- Cases: Rosewill 1.0 mm Thickness 4U Rackmount Server Chassis, Black Metal/Steel RSV-L4000
- Consider this case: Corsair Carbide Series Air 540 High Airflow ATX Cube Case 
- DVDRW (Needed?): Asus 24x DVD-RW Serial-ATA Internal OEM Optical Drive DRW-24B1ST 
- Keyboard and Mouse: AmazonBasics Wired Keyboard and Wired Mouse Bundle Pack 
- Optical drive: HP - DVD1265I DVD/CD Writer 
- Deep learning box for $1700 (Discussion)
- A Full Hardware Guide to Deep Learning
- Cheap build
- How to build a GPU deep learning machine
- Deep Learning Computer Build useful tips, long
- Another box
- Expensive deep learning box
Double GPU Server Build
- This article says that it may be necessary to get both CPUs to get all of the PCI lanes
Double GPU Build
- Needs enough PCIe slots to support both GPUs and other units
- Motherboards: MSI - Z170A GAMING M7 ATX LGA1151 Motherboard , LGA 1151, 3x PCIe 3.0 x 16, 4 x PCIe 3.0 x 1, 6 x SATA 6GB/s, also used in this build
- At least one core (two threads) per GPU
- Chips: Intel - Core i7-6700 3.4GHz Quad-Core Processor 
- CPU Fans: Cooler Master - Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler 
- Buying this fan because it's very cheap for the reviews it got, and the stock cooler for the CPU has had mixed reviews
- 2x GTX 1080 Ti 
- Integrated graphics on CPU: Intel HD Graphics 530
- At least as much RAM as GPUs (2 * 11 GB [GTX 1080 Ti size] = 22 GB, so 32GB)
- Does not have to be fast for deep learning: "CPU-RAM-to-GPU-RAM is the true bottleneck – this step makes use of direct memory access (DMA). As quoted above, the memory bandwidth for my RAM modules are 51.2GB/s, but the DMA bandwidth is only 12GB/s!"
- Crucial - 32GB (2 x 16GB) DDR4-2133 Memory , SATA 6 GB/s interface
- If not enough, should be able to extend this by buying two more cards
- Some say PSU should be 1.5x-2x wattage of system, some say wattage+100W
- PSU: EVGA - SuperNOVA G2 1000W 80+ Gold Certified Fully-Modular ATX Power Supply 
- SSD should be fast enough, no need for M.2 
- SSD: Samsung - 850 EVO-Series 500GB 2.5" Solid State Drive 
- HDD: Seagate - Barracuda 3TB 3.5" 7200RPM Internal Hard Drive 
Other things to consider
- Water cooling? this has a good section on cooling
- Case is not rack mounted
- Setting up Ubuntu and Docker