Explore the cutting-edge advancements in sparsity and how they can make CPUs perform as fast as GPUs in this 50-minute interview with Nir Shavit. Dive into the world of sparse models and their efficient handling using plain CPUs, without the need for specialized hardware. Learn about Neural Magic's innovative algorithms for pruning and forward-propagation of neural networks, and understand the potential applications and significance of sparsity in the Deep Learning community. Discover the links between sparsity and the human brain, the challenges of building sparse models, and the advantages of CPUs over GPUs for sparse computations. Gain insights into the future of sparse architecture, its implications for various neural network types, and the potential synergies with blockchain technology and distributed computing.
How to Make Your CPU as Fast as a GPU - Advances in Sparsity with Nir Shavit
Overview
Syllabus
Introduction
Sponsor: AssemblyAI
Start of Interview
How the NIR company was founded?
What is Sparsity about?
Link between the human brain and sparsity
Where should the extra resource that the human brain doesn't have go?
Analogy for Sparse Architecture
Possible future for Sparse Architecture as standard architure for Neural Networks
Pruning & Sparsification
What keeps us from building sparse models?
Why are GPUs so unsuited for sparse models?
CPU and GPU in connection with memory
What Neural Magic does?
How do you deal with overlaps in tensor columns?
The best type of sparsity to execute tons of CPU
What kind of architecture would make the best use out of a combined system of CPUs and GPUs?
Graph Neural Networks in connection to sparsity
Intrinsic connection between the Sparsification of Neural Networks, Non Layer-Wise Computation, Blockchain Technology, Smart Contracts and Distributed Computing
Neural Magic's target audience
Is there a type of model where it works particularly well and the type where it doesn't?
Taught by
Yannic Kilcher
