I thought I’d share briefly some of our team’s recent experiences in renting time on GPUs for machine learning work.
You’ve probably seen that GPUs are gaining popularity for machine learning because of their inherent parallelism. You may have also heard of renting GPU instances from Amazon for machine learning work. You might suppose, therefore, (as we did!) that renting an Amazon GPU instance would be a good way to gain access to a high-performance GPU.
Here is the key insight I wanted to share:
The motivation for renting a GPU instance at Amazon* is not about getting access to a high performance GPU, but rather for the ability to cheaply gain access to many GPU instances for running parallel experiments.
*This applies to Amazon specifically–I’ll discuss shortly our experience with other providers.
We discovered this insight empirically first. I implemented a benchmark test where we are calculating the L1…
View original post 354 more words