Relevant Paper(s):
Abstract:
For the past decade, enterprise AI stayed in the model-centric comfort zone. Leveraging the
wide-spreaded user expertise on commodity hardware, models and data, practitioners can
easily develop thousands of model variations for various tasks. The recent trend of foundation
models opens the unprecedented opportunity to push enterprise AI into an uncharted
data-centric AI territory. These large foundation models are usually pretrained using
self-supervision at scale. It enables practitioners to use only a handful of model backbones, and
focus on the data-centric tasks for many downstream applications. In this tech talk, I am very
excited to share the full-stack innovations at SN for building enterprise foundation model
solutions. At the ML application layer, I want to highlight the 0-1 accuracy breakthrough in High
Resolution 3D segmentation enabled by the large memory capacity in SambaNova's own RDU.
In the layer of model training, I will highlight our recent endeavor in accelerating domain specific
large language models which leads to a 6X faster time-to-accuracy for pretraining large
language models over standard A100 GPUs recipes. Lastly, I will showcase how our team built
the prototype of learned RDU performance optimization methods, which pushes further towards
fully unleashing the hardware with significantly reduced engineering cost than rule-based
methods. We hope these innovations could push the boundary of enterprise AIs in the
data-centric era.
Bio: Jian Zhang is an Director of Machine Learning at SambaNova Systems. He leads the ML team which builds the deep learning foundations for SambaNova's large-scale enterprise AI solutions. With the mission of democratizing modern foundation model systems, the ML team at SambaNova innovates on both the machine learning and the system aspects, including productionalizing large foundation models and ML/hardware co-design on emerging hardware. Before joining SambaNova Systems, Jian got his PhD in machine learning from Stanford University focusing on machine learning and natural language processing systems.