A hardware-software co-design prototype called PIM-HBM explores zero-copy memory view fusion and fault telemetry for LLM training, using JAX/XLA integration and branchless CUDA kernels to mitigate com…
A developer ported Karpathy's nanochat to run on a TPU v6e-8 using JAX, achieving a CORE score that reproduces the original quality but with a model FLOPs utilization (MFU) of about 24%, roughly half …
NVIDIA and Google researchers have developed a host offloading technique in JAX that reduces high-bandwidth memory bottlenecks during LLM training, achieving up to 57% throughput improvements on NVIDI…
Google Cloud demonstrated elastic training on TPUs, where a worker failure during multi-node LLM training was recovered in under two minutes without restarting the job. Using the JAX AI stack (MaxText…
A developer built a hybrid GNN and decision-tree architecture for predicting XLA compiler runtime, achieving high speed and low memory footprint. The solution outperforms standard deep GNNs by shiftin…
The OpenXLA compiler stack and JAX now run upstream on AMD ROCm, with XLA gating every pull request on real AMD Instinct silicon through GitHub Actions and JAX running hardware tests on every ROCm PR.…
Google's TPU v5p and upcoming Ironwood TPU7x offer higher raw matrix throughput and simpler interconnects than NVIDIA H100 GPUs for large-scale AI training, but GPUs remain superior for inference with…
MLIR, a compiler infrastructure framework, has become the foundation for numerous machine learning compilers including XLA, Triton, Mojo, Torch-MLIR, IREE, and ONNX-MLIR. It provides a reusable IR con…
Google's TPU uses a systolic array architecture optimized for tensor algebra, offering higher throughput and energy efficiency than GPUs for dense matrix operations, but requires XLA compilation and i…
Google launched the TPU Developer Hub, a centralized educational resource for guidance on Google Cloud TPU infrastructure and software stack. The hub covers the full developer lifecycle from pre-train…
Google Cloud launched the TPU Developer Hub, a centralized educational resource providing technical guidance on TPU hardware, software stack, and optimization strategies for AI developers. The hub off…
A developer benchmarked quantum circuit simulation backends using a 20-qubit VQE workload on an NVIDIA RTX 5090 GPU, finding JAX/XLA 12.4x faster than PyTorch and 15.7x faster than TorchQuantum for po…
NVIDIA GPUs handle variable-length inference requests dynamically without recompilation, while TPUs and AWS Trainium require fixed shapes compiled ahead of time, causing crashes or stalls on mismatche…