Author of "Network and Storage Benchmarks for LLM Training on the Cloud"
How media typically covers Henry Zhu
Henry Zhu as author
Adding 'cutlass' to CUDA/Triton kernel names triggers additional compiler optimization passes in ptxas, yielding up to 20% performance improvements in some cases like CuTE sgemm2.cu.
“Author of "Maybe Consider Putting "Cutlass" in Your CUDA/Triton Kernels"”
InfiniBand networking provides 10x faster training than standard Ethernet, and optimal storage selection can double checkpointing speed, delivering 6-7x combined end-to-end speedup for LLM training on cloud infrastructure.
“Author of "Network and Storage Benchmarks for LLM Training on the Cloud"”