IBM – Boosting In-Core AI and Analytics Performance on the Example of Power10

25.10.2023

Fr. Dr. Silvia M. Müller, Distinguished Engineer im IBM Forschungs- und Entwicklungszentrum Böblingen, hat bei uns in der HTL Donaustadt einen Vortrag über moderne Prozessoren und deren kontinuierliche Verbesserung mittels Spezial-Architekturen gehalten. Was für ein technischer und interessentar Vortrag! Unterstützt wurde die Veranstaltung von GSE Guide Share Europe.

Die Agenda war spannend:

Modern processors use a set of techniques to boost their in-core compute performance so that they can tackle compute intensive business applications including a large set of AI tasks.

On the example of the Power10, the presentation will explain state-of-the art concepts including:

  • Instruction level parallelism (more instructions per unit of time):
    • Pipelining
    • Superscalar
    • Out-of-order
    • Multithreading
  • Data parallelism (more computation per instruction):
    • Scalar, vector (SIMD), matrix 
    • Quantization for AI applications

It will also address design trade-offs, for example between SIMD-width and degree of superscalar, design guidelines (e.g.: vector vs matrix) and coding guidelines.

Dr. Silvia M. Mueller
IBM · IBM Germany Research and Development Lab
PhD, 82 Publications

Silvia M. Mueller received the M.Sc. degree in mathematics and the Ph.D. degree in computer science from the University of Saarland, Saarbrücken, Germany, in 1989 and 1991. Silvia M. Mueller works at the IBM Germany Research and Development Lab. As an IBM Distinguished Engineer, she is leading the development of competitive arithmetic units for POWER and z Systems. She has been driving a shift in IBM’s Systems value towards stack solutions based on hardware/software co-design and co-optimizations, specialized arithmetic accelerator engines and hardware differentiation. She works closely with research on all topics like AI and security.

Wir freuen uns sehr, dass wir diesen Vortrag für unsere SchülerInnen und auch externe PartnerInnen ermöglichen konnten.