At its GPU Technology Conference (GTC) event today, consumer graphics and AI silicon power plants Nvidia announces its next-generation graphical process unit (GPU) architecture, called Ampere and its first Ampere-based GPU, the A100. Nvidia says the Ampere GPUs can offer a 20-fold improvement in performance compared to its previous Volta GPU architecturewhich by itself offers far faster processing times for AI workloads than traditional central processing units (CPUs). For further details, see ZDNet’s Natalie Gagliordi’s coverage of all the Nvidia Ampere related news today.
Light up Spark
What I would like to cover here, however, goes beyond these AI headlines and involves a special nugget just for people doing computer engineering, analysis and machine learning work with Apache sparkles. Specifically, Nvidia announces new GPU acceleration features to come Apache Spark 3.0, whose release is expected in late spring.
Also read: NVIDIA is changing from graphics and games to AI and deep learning
Also read: Nvidia doubles on AI
The GPU acceleration functionality is based on open source RAPIDS package of software libraries built on CUDA-X AI. The acceleration technology, called (logically) the RAPIDS Accelerator for Apache Spark, was developed in collaboration by Nvidia and data Bricks (the company founded by Spark’s creators). It allows developers to take their spark code and, without change, run it on GPUs instead of CPUs. This provides much faster training times for machine learning models, especially if the hardware is based on the new Ampere generation GPUs, which itself offer 5 times + faster training and initial / scoring times than their Nvidia Volta predecessors.
Faster training times allow for greater amounts of training data, which is necessary for greater accuracy. But Nvidia says the RAPIDS accelerator also improves performance dramatically Spark SQL and DataFrame operations, which also benefits the GPU acceleration for non-AI workloads. This means that the same spark cluster hardware can be used for both computer technology / ETL workload as well as machine learning jobs.
This, in turn, avoids the need to provide a separate spark cluster dedicated to AI work, and allows the whole load process train test pipeline to run together in a single job, on a single cluster. Finally, says Nvidia, the RAPIDS accelerator also increases data transfer performance across nodes in a spark cluster by leveraging open source Unified Communication X (UCX) framework, which allows data to move directly between GPU memory.
In addition, as the RAPIDS accelerator is designed for open source Apache Spark, it will not only benefit users of the Databricks platform, but also users of machine learning platforms offered by major public cloud providers. In an advanced briefing for press members NVidia CEO Jensen Huang explained that spark cluster users were enabled Azure Machine Learning or Amazon SageMaker can also benefit from the GPU acceleration.
Adobe is experiencing GPU acceleration
Adobe – An Nvidia partner who is also a Databricks customer has tested the GPU-accelerated Spark 3.0 technology, saying it has achieved 7x performance improvement and 90% cost savings. “We see significantly faster performance with NVIDIA-accelerated Spark 3.0 compared to running Spark on CPUs,” said William Yan, senior director of Machine Learning, Adobe. “With these game-changing GPU performance gains, brand new opportunities to improve AI-powered features are opened in our full suite of Adobe Experience Cloud apps.”
At this point, Apache Spark is such a pervasive platform – both in standalone and integrated form – for analysis and machine learning that the RAPIDS Accelerator has the potential to drive mainstream adoption of GPU technology. It will be up to the public cloud providers to make the economy of GPU-based infrastructure sufficient to support such widespread GPU deployment. For now, though, you can bet that with so many teams and organizations fighting a vaccine and effective treatments for COVID-19, there will be a significant cohort of clear, willing and skilled early adopters for GPU-accelerated spark.