Youngbin Lee·Nov 8Under the Hood of JAX on TPU: From Asynchronous Dispatch to Custom Kernels with PallasHow JAX hides host latency with futures, why naive benchmarks fail, and how to explicitly program TPU VMEM and DMA using Pallas
Youngbin Lee·Nov 8Benchmarking JAX on Cloud TPU v5e vs. PyTorch on NVIDIA A100: Why Peak TFLOPs/s LiesA Roofline analysis and real-world hardware experiment showing how batch size, XLA fusion, and host transfers dominate inference latency
Youngbin Lee·Nov 8How to Think About TPUs vs. GPUs: TensorCore, MXU, VMEM, and Compiler-Driven ExecutionWhy NVIDIA’s “Tensor Core” and Google’s “TensorCore” are completely different architectural units — and how VMEM scratchpads change the…
Youngbin Lee·Nov 8Tracing Y = ReLU(XW + b) from Python to Silicon: JAX (OpenXLA) vs. PyTorch 2How a single line of linear algebra transforms through Jaxpr, StableHLO, and XLA HLO fusions compared to TorchDynamo, AOTAutograd, and…
Youngbin Lee·Apr 29GPT Writes One Token at a Time. This Model Writes Entire Blocks at Once.Building a Diffusion Language Model from Scratch with Flax NNX on TPU