Accelerator-centric systems design - ACACES summer school seminar

Lev Denisov
DEIB PhD Student
DEIB - PT1 Room (Building 20, Ground Floor)
September 9th, 2022
12.00 pm
Contacts:
Lev Denisov
Research Line:
Advanced software architectures and methodologies
DEIB PhD Student
DEIB - PT1 Room (Building 20, Ground Floor)
September 9th, 2022
12.00 pm
Contacts:
Lev Denisov
Research Line:
Advanced software architectures and methodologies
Sommario
On September 9th, 2022 at 12.00 pm Lev Denisov, DEIB PhD Student, will hold a seminar on "Accelerator-centric systems design" in DEIB PT1 Room.
This seminar is based on the series of lectures by Mark Silberstein about the accelerator-centric systems design read at ACACES 2022 summer school.
The challenge to scale general computing hardware performance made the research community to look for other ways to speed up the computations. While the more disruptive technologies such as quantum computing and neuromorphic computing are still not mature enough to be used in the real applications the most realistic way to achieve the speedup is using specialized accelerator hardware designed for the specific workloads. The existing accelerators such as GPUs have shown that they can significantly decrease the computation of parallelizable programs. Unfortunately, according to the Amdahl’s law, the total program execution cannot be faster than the slowest part of the program, limiting what one accelerator type can achieve. This calls for using accelerators for IO, security, communication, etc., which leads to increased complexity and inefficiency caused by the integration of different accelerators in the existing computing architectures. In this seminar we will discuss a way to design systems that can efficiently utilize heterogeneous accelerators.
This seminar is based on the series of lectures by Mark Silberstein about the accelerator-centric systems design read at ACACES 2022 summer school.
The challenge to scale general computing hardware performance made the research community to look for other ways to speed up the computations. While the more disruptive technologies such as quantum computing and neuromorphic computing are still not mature enough to be used in the real applications the most realistic way to achieve the speedup is using specialized accelerator hardware designed for the specific workloads. The existing accelerators such as GPUs have shown that they can significantly decrease the computation of parallelizable programs. Unfortunately, according to the Amdahl’s law, the total program execution cannot be faster than the slowest part of the program, limiting what one accelerator type can achieve. This calls for using accelerators for IO, security, communication, etc., which leads to increased complexity and inefficiency caused by the integration of different accelerators in the existing computing architectures. In this seminar we will discuss a way to design systems that can efficiently utilize heterogeneous accelerators.