Uploaded November 2026 | Updated November 2026, 3 days ago
Speakers:
Dr. S. K. Nandy, Deepak Shapeti, Dr. Ranjani Narayan & Dr. Nils Voss.
----------
Abstract:
GPUs were built for dense-tensor throughput, but the fastest-growing work in both HPC and AI — sparse solvers, mixture-of-experts inference, long-context and agentic workloads — is irregular and bound by memory bandwidth and latency. Both fields now reward hardware that adapts its parallelism to the program.
REDEFINE is a software-defined many-core processor built on that premise: a dynamic-dataflow execution model, a fabric of Compute Resources built from RISC-V based compute cores and function accelerators based on PoSiCore elements — extending the open ISA rather than replacing it — and a C-with-pragmas programming framework to exploit task level and data level parallelism; We present the motivation, execution model and architecture, programming framework, and the status of the REDEFINE chip and IP, closing with a composable REDEFINE for agentic AI at the edge.
----------
Further information:
bsc.es/es/research-and-development/research-seminars/sors-redefine-software-defined-dynamic-dataflow-the-convergence-hpc-and-agentic-ai-0
Speakers:
Dr. S. K. Nandy, Deepak Shapeti, Dr. Ranjani Narayan & Dr. Nils Voss.
----------
Abstract:
GPUs were built for dense-tensor throughput, but the fastest-growing work in both HPC and AI — sparse solvers, mixture-of-experts inference, long-context and agentic workloads — is irregular and bound by memory bandwidth and latency. Both fields now reward hardware that adapts its parallelism to the program.
REDEFINE is a software-defined many-core processor built on that premise: a dynamic-dataflow execution model, a fabric of Compute Resources built from RISC-V based compute cores and function accelerators based on PoSiCore elements — extending the open ISA rather than replacing it — and a C-with-pragmas programming framework to exploit task level and data level parallelism; We present the motivation, execution model and architecture, programming framework, and the status of the REDEFINE chip and IP, closing with a composable REDEFINE for agentic AI at the edge.
----------
Further information:
bsc.es/es/research-and-development/research-seminars/sors-redefine-software-defined-dynamic-dataflow-the-convergence-hpc-and-agentic-ai-0










