Difference between revisions of "GPU Build"

From edegan.com
Jump to navigation Jump to search
 
(41 intermediate revisions by 2 users not shown)
Line 1: Line 1:
{{McNair Projects
+
{{Project
 +
|Has project output=Content
 +
|Has sponsor=McNair Center
 
|Has title=GPU Build
 
|Has title=GPU Build
 
|Has owner=Oliver Chang,Kyran Adams
 
|Has owner=Oliver Chang,Kyran Adams
Line 5: Line 7:
 
}}
 
}}
  
===Single GPU vs Multi GPU===
+
==Final Decision==
 +
 
 +
We decided to clone the NVIDIA [[DIGITS DevBox]]: https://developer.nvidia.com/devbox
 +
 
 +
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==
 +
 
 
*[https://www.nvidia.com/en-us/geforce/products/10series/geforce-gtx-1080-ti/ GTX 1080 Ti Specs]
 
*[https://www.nvidia.com/en-us/geforce/products/10series/geforce-gtx-1080-ti/ GTX 1080 Ti Specs]
 +
* Since we are using Tensorflow, it doesn't scale well to multiple GPUs for a single model
 
* [http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/ Which GPU for deep learning (04/09/2017)]  
 
* [http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/ 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."
 
# "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."
Line 13: Line 25:
 
# If the network doesn't fit in the memory of one GPU (11 GB),  
 
# If the network doesn't fit in the memory of one GPU (11 GB),  
 
* [https://devtalk.nvidia.com/default/topic/743814/cuda-setup-and-installation/advice-on-single-vs-multi-gpu-system/ Advice on single vs multi-GPU system]
 
* [https://devtalk.nvidia.com/default/topic/743814/cuda-setup-and-installation/advice-on-single-vs-multi-gpu-system/ Advice on single vs multi-GPU system]
# Might want to get two graphics cards, one for development, one (crappy card) for operating system  
+
# Want to get two graphics cards, one for development, one (crappy or onboard card) for operating system [https://stackoverflow.com/questions/21911560/how-can-i-set-one-nvidia-graphics-card-for-display-and-other-for-computingin-li]
*[https://stackoverflow.com/questions/37732196/tensorflow-difference-between-multi-gpus-and-distributed-tensorflow]
+
*[https://stackoverflow.com/questions/37732196/tensorflow-difference-between-multi-gpus-and-distributed-tensorflow Different uses of multiple GPUs]
# Different uses of multiple GPUs
 
 
# 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.
 
# 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).  
 
# Replicated training: Start up multiple copies of the model, train them, and then synchronize their learning (the gradients applied to their weights & biases).  
  
===Other Builds===
+
====TL;DR====
 +
 
 +
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.
 +
 
 +
=== K80, NVLink ===
 +
*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.
 +
*[https://www.quora.com/Which-GPU-is-better-for-Deep-Learning-GTX-1080-or-Tesla-K80 This source] says to get the 1080 because the K80 is basically two K40s, which have less memory bandwidth than the 1080. [https://www.reddit.com/r/deeplearning/comments/5mc7s6/performance_difference_between_nvidia_k80_and_gtx/ This source] agrees.
 +
 
 +
==Misc. Parts==
 +
*Cases: Rosewill 1.0 mm Thickness 4U Rackmount Server Chassis, Black Metal/Steel RSV-L4000[https://www.amazon.com/gp/product/B0056OUTBK/ref=oh_aui_detailpage_o04_s00?ie=UTF8&psc=1]
 +
*Consider this case: Corsair Carbide Series Air 540 High Airflow ATX Cube Case [https://www.amazon.com/dp/B00D6GINF4/ref=twister_B00JRYFVAO?_encoding=UTF8&psc=1]
 +
*DVDRW (Needed?): Asus 24x DVD-RW Serial-ATA Internal OEM Optical Drive DRW-24B1ST [http://www.amazon.com/Asus-Serial-ATA-Internal-Optical-DRW-24B1ST/dp/B0033Z2BAQ/ref=sr_1_2?s=pc&ie=UTF8&qid=1452399113&sr=1-2&keywords=dvdrw]
 +
*Keyboard and Mouse: AmazonBasics Wired Keyboard and Wired Mouse Bundle Pack [http://www.amazon.com/AmazonBasics-Wired-Keyboard-Mouse-Bundle/dp/B00B7GV802/ref=sr_1_2?s=pc&rps=1&ie=UTF8&qid=1452402108&sr=1-2&keywords=keyboard+and+mouse&refinements=p_72%3A1248879011%2Cp_85%3A2470955011]
 +
* Optical drive: HP - DVD1265I DVD/CD Writer [https://www.newegg.com/Product/Product.aspx?Item=N82E16827140098&ignorebbr=1&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
 
 +
 
 +
==Other Builds/Guides==
 +
 
 +
* [https://blog.slavv.com/the-1700-great-deep-learning-box-assembly-setup-and-benchmarks-148c5ebe6415 Deep learning box for $1700] ([https://news.ycombinator.com/item?id=14438472 Discussion])
 +
* [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://medium.com/@SocraticDatum/getting-started-with-gpu-driven-deep-learning-part-1-building-a-machine-d24a3ed1ab1e How to build a GPU deep learning machine]
 +
* [https://www.slideshare.net/PetteriTeikariPhD/deep-learning-workstation Deep Learning Computer Build] useful tips, long
 +
* [https://www.tooploox.com/blog/deep-learning-with-gpu Another box]
 +
* [http://graphific.github.io/posts/building-a-deep-learning-dream-machine/ Expensive deep learning box]
 +
 
 +
==Double GPU Server Build==
 +
[https://pcpartpicker.com/user/kyranadams/saved/gDzFdC PC Partpicker build]
 +
 
 +
*[https://www.quora.com/Can-I-double-the-PCIe-lanes-in-a-dual-CPU-motherboard This article] says that it may be necessary to get both CPUs to get all of the PCI lanes
 +
 
 +
==Double GPU Build==
 +
[https://pcpartpicker.com/user/kyranadams/saved/ykK7hM PC Partpicker build]
 +
 
 +
===Motherboard===
 +
*Needs enough PCIe slots to support both GPUs and other units
 +
*Motherboards: MSI - Z170A GAMING M7 ATX LGA1151 Motherboard [https://www.newegg.com/Product/Product.aspx?Item=9SIA85V4SC7911&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=], LGA 1151, 3x PCIe 3.0 x 16, 4 x PCIe 3.0 x 1, 6 x SATA 6GB/s, also used in [https://medium.com/@SocraticDatum/getting-started-with-gpu-driven-deep-learning-part-1-building-a-machine-d24a3ed1ab1e this build]
 +
 
 +
===CPU/Fan===
 +
*At least one core (two threads) per GPU
 +
*Chips: Intel - Core i7-6700 3.4GHz Quad-Core Processor [https://www.amazon.com/dp/B0136JONG8/?tag=pcpapi-20]
 +
*CPU Fans: Cooler Master - Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler [https://www.newegg.com/Product/Product.aspx?Item=N82E16835103099&ignorebbr=1&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
*Buying this fan because it's very cheap for the reviews it got, and the stock cooler for the CPU has had mixed reviews
 +
 
 +
===GPU===
 +
* 2x GTX 1080 Ti [https://www.newegg.com/Product/Product.aspx?Item=N82E16814487338&ignorebbr=1&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
* Integrated graphics on CPU: Intel HD Graphics 530
 +
 
 +
===RAM===
 +
*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!"[http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/]
 +
* Crucial - 32GB (2 x 16GB) DDR4-2133 Memory [https://www.newegg.com/Product/Product.aspx?Item=9SIA8PV5HF1514&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=], SATA 6 GB/s interface
 +
* If not enough, should be able to extend this by buying two more cards
 +
 
 +
===PSU===
 +
*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 [https://www.newegg.com/Product/Product.aspx?Item=N82E16817438010&ignorebbr=1&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
 
 +
===Storage===
 +
*SSD should be fast enough, no need for M.2 [http://timdettmers.com/2015/03/09/deep-learning-hardware-guide]
 +
*SSD: Samsung - 850 EVO-Series 500GB 2.5" Solid State Drive [https://www.newegg.com/Product/Product.aspx?Item=N82E16820147373&ignorebbr=1&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
*HDD: Seagate - Barracuda 3TB 3.5" 7200RPM Internal Hard Drive [https://www.newegg.com/Product/Product.aspx?Item=9SIADG25GT7889&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=]
 +
 
 +
===Other things to consider===
 +
* Water cooling? [http://timdettmers.com/2015/03/09/deep-learning-hardware-guide/ this] has a good section on cooling
 +
* Case is not rack mounted
  
* [https://news.ycombinator.com/item?id=14438472 Deep learning box for $1700] (links to https://blog.slavv.com/the-1700-great-deep-learning-box-assembly-setup-and-benchmarks-148c5ebe6415)
+
==Software tips==
 +
* Setting up Ubuntu and Docker [https://medium.com/@SocraticDatum/getting-started-with-gpu-driven-deep-learning-part-2-environment-setup-fd1947aab29]

Latest revision as of 12:39, 21 September 2020


Project
GPU Build
Project logo 02.png
Project Information
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.


Final Decision

We decided to clone the NVIDIA DIGITS DevBox: https://developer.nvidia.com/devbox

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

  1. "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."
  2. 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."
  3. This source recommends GTX 1080 Tis and does cost analysis of it
  4. If the network doesn't fit in the memory of one GPU (11 GB),
  1. Want to get two graphics cards, one for development, one (crappy or onboard card) for operating system [1]
  1. 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.
  2. 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

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.

K80, NVLink

  • 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.

Misc. Parts

  • Cases: Rosewill 1.0 mm Thickness 4U Rackmount Server Chassis, Black Metal/Steel RSV-L4000[2]
  • Consider this case: Corsair Carbide Series Air 540 High Airflow ATX Cube Case [3]
  • DVDRW (Needed?): Asus 24x DVD-RW Serial-ATA Internal OEM Optical Drive DRW-24B1ST [4]
  • Keyboard and Mouse: AmazonBasics Wired Keyboard and Wired Mouse Bundle Pack [5]
  • Optical drive: HP - DVD1265I DVD/CD Writer [6]


Other Builds/Guides

Double GPU Server Build

PC Partpicker build

  • This article says that it may be necessary to get both CPUs to get all of the PCI lanes

Double GPU Build

PC Partpicker build

Motherboard

  • Needs enough PCIe slots to support both GPUs and other units
  • Motherboards: MSI - Z170A GAMING M7 ATX LGA1151 Motherboard [7], 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

CPU/Fan

  • At least one core (two threads) per GPU
  • Chips: Intel - Core i7-6700 3.4GHz Quad-Core Processor [8]
  • CPU Fans: Cooler Master - Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler [9]
  • Buying this fan because it's very cheap for the reviews it got, and the stock cooler for the CPU has had mixed reviews

GPU

  • 2x GTX 1080 Ti [10]
  • Integrated graphics on CPU: Intel HD Graphics 530

RAM

  • 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!"[11]
  • Crucial - 32GB (2 x 16GB) DDR4-2133 Memory [12], SATA 6 GB/s interface
  • If not enough, should be able to extend this by buying two more cards

PSU

  • 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 [13]

Storage

  • SSD should be fast enough, no need for M.2 [14]
  • SSD: Samsung - 850 EVO-Series 500GB 2.5" Solid State Drive [15]
  • HDD: Seagate - Barracuda 3TB 3.5" 7200RPM Internal Hard Drive [16]

Other things to consider

  • Water cooling? this has a good section on cooling
  • Case is not rack mounted

Software tips

  • Setting up Ubuntu and Docker [17]