CUDA

A deep technical review of Core PyTorch sessions at PyTorch Conference North America 2026, covering compiler architecture, distributed communication, device portability, and low-latency LLM serving.

An in-depth technical analysis of executing NVIDIA CUDA binaries and frameworks on AMD Radeon GPUs under Windows using translation layers like ZLuda, HIP SDK, and SCALE, alongside cloud API fallbacks.

An in-depth technical exploration of hardware acceleration, compiler infrastructure, Triton kernels, and distributed compute scaling sessions featured at PyTorch Conference North America 2026.

PyTorch 2.14 introduces NVGEMM with CuTeDSL-generated CUTLASS kernels in TorchInductor, delivering major epilogue fusion upgrades, enhanced dynamic compilation, and lower latency for modern deep learning workloads.

A deep dive into building a high-performance LLM runtime from scratch, covering CUDA graphs, weight packing, and H100 optimization strategies.

Discover how building a custom GPU-resident Top-K CUDA kernel eliminates PCIe transfer bottlenecks in Agentic RAG pipelines, delivering microsecond-level retrieval for high-performance LLM applications.

A deep dive into how vLLM uses PagedAttention to eliminate memory fragmentation and increase LLM inference throughput by up to 24x.

Explore TitanCore Core-1, a high-performance C++/CUDA infrastructure designed for trillion-parameter LLM training using ZeRO-3 and custom fused kernels for 2.6x speedup.

While the world focuses on Nvidia's H100 and Blackwell GPUs, the real secret to their trillion-dollar dominance lies in CUDA. This deep dive explores how software, not just silicon, created an unassailable moat for AI development.

Explore how modern LLMs like Claude 3.5 Sonnet and OpenAI Codex are revolutionizing GPU programming by generating high-performance Triton and CUDA kernels.

An in-depth technical exploration of how Claude 3.5 Sonnet is revolutionizing low-level GPU programming and serving as a high-fidelity teacher for open-source model distillation.

Discover how fused Triton kernels can drastically reduce memory overhead in the final LLM layers, preventing OOM errors during training and fine-tuning.