Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
A heterogeneous RISC-V based accelerator chip (EPAC) designed for high-performance computing (HPC) and AI workloads, featuring specialized vector and stencil processing tiles.
Utility
citations
0
co_authors
39
EPAC (European Processor Accelerator) represents a significant multi-year sovereign silicon effort by the European Processor Initiative (EPI). Its defensibility score of 8 is derived from the immense physical and capital barriers to entry inherent in custom silicon development (tape-out on GF22FDX, 0.3B transistors). This is not a software project that can be easily cloned; it is a validated hardware implementation involving multiple partners like BSC, ETH Zurich, and Semidynamics. While the 22nm node is trailing-edge compared to consumer 3nm/5nm chips, the moat lies in the domain-specific architecture (VEC and Stencil tiles) and the geopolitical mandate for European technological sovereignty. Frontier labs (OpenAI/Google) pose low risk here as they focus on general-purpose AI accelerators (TPUs/LPUs) or massive-scale cloud infrastructure, rather than specialized RISC-V HPC chips for European research clusters. The 39 forks in just 3 days for an academic/hardware project indicate high interest within the RISC-V and HPC research communities. The main risk is the competitive gap in raw TFLOPS compared to commercial giants like NVIDIA or AMD, but within its niche of sovereign, energy-efficient HPC, it is category-defining for the region.
TECH STACK
INTEGRATION
hardware_dependent
READINESS
The reusable building blocks distilled from this project — each a mechanism you could lift into your own.
TileMemoryRequest -> NoCTransaction
Route transactions from domain-specific accelerator tiles (vector, stencil, variable-precision) through a unified Network-on-Chip (NoC) using standardized bridge protocols.
L1CacheMiss -> CoherentMemoryResponse
Maintain memory coherence across diverse tile-specific L1 caches (vector, stencil, scalar) via a directory tracking system hosted at shared L2 cache slices.