in Link Post

NVIDIA Would Like You to Buy Their More Expensive GPUs for Deep Learning, Please

Source: www.nvidia.com, Hat Tip: WirelessWire.jp

It seems that NVIDIA has recently updated the EULA for it’s GeForce line of consumer GPUs1 to include the following statement:

No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that blockchain processing in a datacenter is permitted.

Why is this important? If you’re any sort of machine learning practitioner you really need to use NVIDIA GPUs. The most cost effective way to do that is with the GeForce line. Instead, NVIDIA would like you to use its much more expensive Tesla line of compute focussed cards.

This certainly looks like a money grab from NVIDIA. That blockchain exception seems to be the biggest clue. NVIDIA’s betting that the ML people care about performance more than they care about value. The blockchain / cryptocurrency people, though? They care about value above all else2.

I can see three outcomes here:

  1. The machine learning community does what its told and gets it’s collective wallet out en masse;
  2. The same community pushes support into OpenCL support for the main machine learning frameworks and makes AMD cards a more viable option;
  3. NVIDIA walks this back.

Personally, I’m betting on 2. I think this change makes it inevitable. I suspect 3 will happen, it’s just a question of how much damage is done first and how much trust has been lost.

  1. Or more specifically the EULA for the driver software.
  2. If it costs you $200 to mine $100 worth of ETH that’s not good business. Unless you’re playing a long game and betting on the price going up, of course.