Monday, November 13, 2017

TensorFlow for AMD Radeon Will Break Nvidia Monopoly

Nvdia has established a virtual monopoly in the artificial intelligence and machine learning arena through its CUDA. Jensen Huang openly acknowledged this. When asked by analysts what Nvidia's advantages are in the AI/ML world, Huang said it's about CUDA. Indeed, the most popular ML framework, TensorFlow, supports CUDA and thus Nvidia GPUs only. Even if you are an AMD loyalist and have bought an AMD VEGA GPU, if you want to do some TensorFlow or Keras work today, you need a NVIDIA GPU. A GTX 1060 seems to be the minimum requirement for AI. But things are changing.

The lack of ML software support in AMD GPUs has attracted attention. A company named Vertex.AI has released an open source machine learning engine called PlaidML. It is based on OpenCL and its initial benchmarks show great promise for AMD Radeon, which has superior compute performance. Unfortunately, plaidML is still in development and lacks support for recurrent neural networks.

Today, AMD announced that its new ROCm 1.7 and MIOpen library will have TensorFlow support. Since Keras runs on top of TensorFlow, Radeon owners can also enjoy their GPU's AI power with a much nicer and easier to use programming interface.

What does this mean for Nvidia? Its near monopoly in AI/ML will be broken. Nvidia will have to share the fledgling AI market with AMD.

This is no less than a sea change. Coupled with Intel's adoption of AMD GPU technology in its mobile platforms, Nvidia's revenue will suffer intolerance decline and its lebensraum will see rapid shrinking.

0 Comments:

Post a Comment

<< Home