A new field report from OpenAI reveals that coding agents, primarily Codex and Claude Code, enabled small teams to modernize critical scientific software in genomics and other fields across eight proj…
Google's open-source TPU compiler, shipped as the Mosaic TPU dialect inside JAX, documents eight explicit memory spaces that programmers must manage, including vector memory (VMEM), scalar memory (SME…
AMD has published a machine-readable ISA for its Instinct GPUs and partnered with Anthropic and OpenAI to let frontier models natively write and optimize low-level kernels, claiming a 38% inference sp…
XLA, the compiler behind JAX, TensorFlow, and PyTorch/XLA, optimizes array programs by freezing shapes, statically allocating buffers, and fusing operations against a global cost model, which makes it…
XLA, the compiler under JAX, TensorFlow, and PyTorch/XLA, uses two intermediate representations: classic HLO (a hand-built C++ IR) and MLIR dialects such as StableHLO and CHLO, with a translation laye…
Deep Q-Networks (DQN) replace the Q-table in Q-learning with a neural network that outputs Q-values for all actions, enabling generalization across similar states. To stabilize training, DQN uses expe…
Researchers released four natural-language-to-MLIR benchmarks totaling 410 in-scope pairs and a three-layer schema-derived constraint stack that lets SmolLM2-1.7B match or exceed 15B-34B code LMs on s…
A developer built a neural network training loop in five lines using JAX, demonstrating gradient descent from scratch and achieving a 32x speedup with JIT compilation on a MacBook CPU. The project ver…
A new open-source project simulates 36-qubit quantum circuits with a 549 GB state-vector footprint at ~0.01ms per gate using pure JAX, accelerated on NVIDIA GPUs and Google Cloud TPU v6e-64 and v5e cl…
A new model compression method called requential coding compresses a generative model by coding a training process built from self-generated data, achieving code length equal to the cumulative teacher…
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…
LFortran and Enzyme enable exact gradients from legacy Fortran code without rewriting, turning validated physics solvers into differentiable layers for JAX. The experimental stack compiles Fortran to …
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 …
George Hotz's tinygrad claims to map PCIe Base Address Registers directly for NVIDIA and AMD GPUs, a low-level runtime approach that bypasses vendor driver APIs like CUDA and ROCm. The tinygrad runtim…
LFortran, LLVM, and Enzyme have been combined to automatically differentiate legacy Fortran simulation code, producing exact gradients through a multi-step time loop that match an analytic answer. The…
Hugging Face announced that custom GPU kernels on its Hub are now signed artifacts governed by a trusted publisher model, requiring a dedicated repository type that replaces the old model-type format.…
A developer building a GPT-2 implementation in JAX discovered that token embeddings and the output head account for nearly half of the model's 163 million parameters, while attention layers use fewer …
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…
Giles Thomas completed building and training a GPT-2 small model from scratch using JAX, achieving a test loss of 3.418784, outperforming both his PyTorch model (3.538161) and the original GPT-2 small…
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…