vTen: Tensor-Centric Verification Framework for Domain-Specific Accelerators

2026-10-06

AI Summary

This study addresses the software-hardware semantic gap in Design Space Architecture (DSA) verification and the efficiency bottlenecks caused by Cocotb synchronization overhead. We propose a data-centric declarative verification framework that pioneers an architecture strictly decoupling verification intent from execution mechanisms. By leveraging a domain-specific language (DSL) and a tensor-centric abstraction model, the framework enables kernel-granularity batch processing, significantly reducing interaction frequency between the host and the simulator. Experimental evaluation on a 3D U-Net accelerator demonstrates that the proposed approach reduces simulation latency by 50% and code complexity by 60.3%, thereby effectively enhancing hardware verification productivity.

Abstract

The semantic gap between tensor-centric software models and signal-level hardware testbenches creates significant productivity bottlenecks in verifying Domain-Specific Accelerators (DSAs). Existing frameworks like Cocotb suffer from prohibitive synchronization overheads due to fine-grained interactions. To address this, we propose vTen, a data-centric framework that strictly decouples verification intent from execution mechanics. By leveraging a declarative DSL and kernel-granular batching, vTen minimizes host-simulator interaction frequency. Evaluation on a production-scale 3D U-Net accelerator demonstrates that vTen achieves a 2x performance improvement in simulation latency and a 60.3% reduction in code complexity compared to Cocotb.

Problem

• Domain-Specific Accelerators
• Verification
• Tensor-centric
• Semantic gap
• Synchronization overhead

Innovation

• Tensor-Centric Verification
• Domain-Specific Accelerators
• Declarative DSL
• Kernel-Granular Batching
• Simulation Overhead Reduction