Nathaniel Bleier
University of Michigan, MI·Computer Science
Specialized computer chips called accelerators run voice recognition in phones, driver assistance in cars, medical imaging machines and data centers. Designing one is hard because choices about which operations to build into silicon, how to lay out circuits and how to translate software onto the chip all affect one another.
- Fields
Computer Science
Engineering
- Topics
hardware-software co-design
hardware accelerators
neuro-symbolic reasoning
- Grants
- 1 known grant, through Aug 2030
What they work on
- Combine learning-guided program synthesis with automated logical reasoning to generate instruction sequences, accelerator control codes and microcoded expansions that are formally proven equivalent to the original program.
- Use equality saturation, a compact representation of many equivalent programs, with integer-programming extraction to jointly optimize a chip's instruction set, microarchitecture parameters and software transformations.
- Apply inductive proof techniques from hardware verification to detect redundant circuitry when a chip runs a known class of workloads, automatically pruning general-purpose designs into compact application-specific ones.
- Validate the framework on application-specific instruction-set processors and domain-specific accelerators, measuring area, power, performance and correctness on standard embedded benchmarks.
More about the research
Designers either explore these choices by hand for years or use automated tools that can produce inefficient or incorrect designs. As artificial intelligence strains electricity supplies, these mistakes grow more costly.
Why it matters
Faster, provably correct chip design tools could help engineers build smaller, more energy-efficient accelerators for cars, medical devices and AI data centers, where errors and wasted power are costly. Open-source release could lower barriers for chip designers outside large companies.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Bo Yuan
Rutgers University-New Brunswick, NJ·Engineering·Early-career
Develop design principles for building domain expertise into learning-based robot motion planners, improving their reliability and performance in complex, resource-limited real-world settings.
Josep Torrellas
University of Illinois Urbana-Champaign, IL·Computer Science
Develop approximate algorithms for graph clustering in static and dynamic settings, graph construction that preserves key information, and machine learning, for computational biology, social network and software verification problems.
Soner Onder
Michigan Technological University, MI·Computer Science
Develop Vectorized Instruction Space (VIS), which converts control dependences into data dependences using an executable dynamic single-assignment form at the machine-instruction level, with lower overhead than traditional predication.
The school
University of Michigan: 124 labs on PI Pages, $2.1B a year on research, 963 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.