Neural Magic is excited to wrap up a year of innovation that included new versions of our machine learning libraries and tools—DeepSparse, SparseML, and SparseZoo, which are designed to accelerate inference using the power of software.
Here are some of Neural Magic’s greatest hits for 2022:
Innovation: We proudly launched commercial support for our blazing fast DeepSparse inference runtime, DeepSparse Enterprise 1.0, which enables customers of all sizes—from start-ups to global Fortune 100 brands—to deploy deep learning models confidently. Neural Magic also demonstrated significant performance improvements, both by creating models with higher levels of sparsity through the use of novel sparsity techniques such as AC/DC and OBS, and by developing more optimizations in DeepSparse. We created sparse versions of more models in the NLP and CV domains, including oBERT, BERT-Large, YOLOv5, and YOLACT. Our revamped documentation website provides integrated experiences and step-by-step guides, and our Labs initiative offers specialized consulting services. To learn more about how to use all of our latest tools, visit the blogs and news archive.
Partnerships and Integrations: Neural Magic announced partnerships and integrations with companies such as AMD, AWS (Lambda and SageMaker), Azure, Argilla, Google (Cloud Run and Kubernetes Engine), Hugging Face, MLCommons, and Weaviate.
Industry Conferences: We attended over a dozen industry conferences and events over the past year, including Ai4, CVPR, Edge AI Summit, ICIP, Intel Vision, NeurIPS, and re:MARS.
Research: Our research team presented or had accepted papers at leading conferences:
- CVPR—How Well Do Sparse Imagenet Models Transfer?
- EMNLP—The Optimal BERT Surgeon: Scalable and Accurate Second-Order Pruning for Large Language Models
- NeurIPS—Software-Delivered AI: Using Sparse-Quantization for Fastest Inference on Deep Neural Networks
- NeurIPS—Optimal Brain Compression: A Framework for Accurate Post-Training Quantization and Pruning
- Sparse*BERT: Sparse Models are Robust
Webinars: We presented on several different topics, from optimizing deep learning models to achieving the fastest CPU inference performance. If you missed a webinar, catch the recording from our video playlist. We also launched a monthly virtual Lunch and Learn open to the public - take a break and learn how well sparse models transfer.
2022 was a busy and productive Neural Magic year as we continue to drive innovation in the field of machine learning and sparsity. We thank all our software users, staff, partners, advisors, and vendors for getting us to this point. With an exciting roadmap ahead and new partnerships to announce, we look forward to working with you in 2023!