Search papers, labs, and topics across Lattice.
Georgia Institute of Technology
4
0
5
Industry-standard GPU memory sanitizers harbor major blind spots across core memory spaces, finally quantifiable through 149 targeted CUDA stress tests.
Remote HBM traffic can vary by 90x for the same GEMM dimensions, and a 2D block-swizzle traversal can slash this traffic by over 5 times.
Achieving up to 24.7x reduction in remote HBM traffic for LLM GEMM tasks could redefine efficiency standards in multi-chiplet GPU architectures.
Recovering type information from untyped GPU register files is the key to enabling effective binary analysis, unlocking reverse engineering and security analysis of proprietary GPU code.
