Choosing Tensor Accelerators for Specific Tasks: Compute vs Memory Bound Models, Arithmetic Intensity, and Model Quantization

As an infrastructure provider, we often see that potential clients are somewhat disoriented when it comes to choosing the right accelerators. The technology evolves rapidly, and questions like “What’s out there?”, “How do these cards differ?”, and “Which ones are better?” are completely valid. But the selection criteria are far from simple. In this talk, we’ll explain key concepts such as compute-bound vs memory-bound models, arithmetic intensity, and model quantization—factors that are crucial when choosing the right hardware for your workload.

Hennadiy Karpov
De Novo, CTO
  • In IT since 1995, holding a Ph.D. in Physical and Mathematical Sciences.
  • One of the founders of De Novo (2008) and among the authors of the business concept and high-level architecture of the first commercial data center in Ukraine (2007-2010) and Ukraine's first cloud (2010-2012).
  • Currently, his professional interests are focused on developing PaaS-class cloud services (managed Kubernetes, scalable object storage, database as a service, AI/ML workspace, and more) for the Ukrainian cloud and for Ukraine.
  • LinkedIn, Facebook
Sign in
Or by mail
Sign in
Or by mail
Register with email
Register with email
Forgot password?