The days when GPUs were only about letting you play Crysis with a higher framerate are long over. Many of the most transformative new computational techniques now rely on GPUs’ ability to quickly run certain algorithms in parallel.
One area where GPUs are especially helpful (besides video encoding) is machine learning, and NVIDIA is making a big bet on this. The company is launching two new hardware accelerators today, as well as a suite of tools that will help developers and data center managers use these accelerators to run deep learning software, as well as image and video processing jobs on them.
In addition, NVIDIA today announced that it has partnered with Mesosphere — the container-centric solution for managing large data centers as if they were a single pool of resources. Together, NVIDIA and Mesosphere want to “make it easier for web-services companies to build and deploy accelerated data centers for their next-generation applications.”
Because of the work Mesosphere did with NVIDIA, developers using Apache Mesos (the open-source backbone of Mesosphere’s data center operating system) will be able to use GPU resources in a data center just like they use CPUs and memory. GPU resources will be clustered into a single pool and the software will automatically distribute jobs across all the different machines that offer compatible GPUs.
As for the hardware, NVIDIA is launching two products: the M40 and M4 GPU Accelerators. The M40 is optimized for machine learning and was built and tested for data center use. The M4 was optimized for similar use cases, but with a focus on low-power consumption and video processing.
A number of public cloud vendors, including AWS and Microsoft, now either offer GPU-centric virtual machines or will offer them soon — and for the most part, these data center operators are betting on NVIDIA. Google is making a big bet on machine learning internally, but it hasn’t made GPU instances available on its cloud platform just yet. Chances are it will do so pretty soon — maybe in the context of a dedicated machine-learning service.