by Mathieu Poliquin
NVIDIA’s GTX 1060 shipped in many variants, one maybe less known variant is the 5GB edition intended for Chinese internet cafes. Does it offer a good peformance/price ratio for Machine learning?
Bang for the buck this is not the best gp106 based card you can get for Machine Learning. The marginal cheaper price of the 5GB variant is not enough to justify it’s lower performance mainly due to lower memory bandwidth. That said if you find a GTX 1060 5GB at 2/3 of the price of a GTX 1060 6GB than it might be worth it but currently the price is just a couple of dollars lower.
For gaming thought and people who need to buy graphics card in big batches like for internet cafes, it might be worth it.
Performance tests results
|Test||P106-100||GTX 1060 5GB|
|Cifar10 peak||~8000 examples/sec||~7000 examples/sec|
|PPO2 Atari Pong||~1110 frame/sec||~1000 frames/sec|
|Host to Device||3094.4 MB/s||11205.9 MB/s|
|Device to Host||3207.3 MB/s||12788.0 MB/s|
|Device to Device||152542.9 MB/s||115346.1 MB/s|
The 5GB variant has a difference in memory bandwidth capacity: 168GB/s (5GB card) as opposed to 192GB/s (3GB and 6GB cards) and since these ML benchmarks are bandwidth bound it affects the performance.tags: gtx 1060 - 5GB - gpu - machine learning