Nvidia:HardwareSelect

딥러닝 (Deep learning)을 위한 NVIDIA하드웨어 선택방법에 대하여 설명한다.

Hardware select

• Best GPU overall: GTX Titan X
• Cost efficient but expensive: GTX Titan X, GTX 980, GTX 980 Ti
• Cost efficient but troubled: GTX 580 3GB (lacks software support) or GTX 970 (has memory problem)
• Cheapest card with no troubles: GTX 960 4GB or GTX 680
• I work with data sets > 250GB: GTX Titan, GTX 980 Ti or GTX Titan X
• I have little money: GTX 680 3GB eBay
• I have almost no money: AWS GPU spot instance
• I do Kaggle: GTX 980 or GTX 960 4GB
• I am a researcher: 1-4x GTX Titan X
• I want to build a GPU cluster: This is really complicated, you can get some ideas here
• I started deep learning and I am serious about it: Start with one GTX 680, GTX 980, or GTX 970 and buy more of those as you feel the need for them; save money for Pascal GPUs in 2016 Q2/Q3 (they will be much faster than current GPUs)

결론은 아래와 같다.

• 돈이 충분하다면 GTX Titan X or GTX 980
• 돈이 부족하다면 GTX 960 or GTX 680
• 문제가 있지만 컨트롤할 수 있다면 GTX 970

GTX TitanX vs Tesla K40/K80

• Would you build a multi-GPU system for deep learning with GTX Titan X or Tesla K40/K80? What are the pros and cons?

Specs:

• GeForce Titan X Features: 3072 Cores, 12GB Ram, 336GB/s, 384-bit bus, Boost Clock: 1075MHz, Core Clock 1000MHz, 7GHz DDR5, 6.2 TFLOPS SP
• Tesla K40 Features: 2880 Cores, 12GB Ram, 288 GB/s, 384-bit bus, Boost Clock: 810-875MHz, Core Clock: 745MHz, 6GHz GDDR5, 4.29 Tflops SP

Pros of Geforce Titan X:

• Cheap price
• Because there is graphics support you can visualize your processed data easily
• It has latest Maxwell architecture, whereas Tesla K40 has kepler architecture (Not sure but Nvidia might not release Maxwell architecture Tesla GPUs)
• More GPU cores than Tesla K40

Pros of Tesla K40:

• It is targeted for workstations and servers
• Tesla accelerators deliver the best cluster performance while jobs complete with 100-percent reliability and manageability
• More number of double precision units and hence improves double precision performance
• Some Tesla-exclusive features include:
  • NVIDIA GPUDirect RDMA for InfiniBand performance
  • Hyper-Q for MPI (Hyper-Q for CUDA Streams is supported on GeForce GTX TITAN)
  • ECC protection for all internal and external registers and memories
  • Supported tools for GPU and cluster management, such as Bright Computing, Ganglia.

Final word:

• You can decide what you want reliability or cost effectiveness
• Based on the specs for single precision operations Geforce Titan X will give more performance
• ECC support which is not available with GeForce Titan X will not hurt deep learning applications (that is my guess, you can verify more on this and let us all know)

See also

• CUDA
• Scalable Link Interface (SLI)
• Nouveau

Favorite site

• Which GPU to use for deep learning?
• [추천] 딥러닝 공부 가이드 (HW / SW 준비편) ¹
  • Which GPU(s) to Get for Deep Learning: My Experience and Advice for Using GPUs in Deep Learning
  • A Full Hardware Guide to Deep Learning
• Deep Learning을 위해 어떤 GPU를 써야 할까? ²

TL;DR advice

References