CUB algorithms are now easier to use with single-phase APIs. Previously, developers had to manually query and allocate temporary storage. New overloads accept a memory resource directly, automating the management of intermediate scratch space.

NVIDIA CCCL 3.1 now offers three determinism modes for floating-point reductions. Developers can trade performance for precision, ranging from "not-guaranteed" (fastest) to "gpu-to-gpu" (slowest, but guarantees bitwise-identical results).

cuSOLVER sees significant speedups in eigen-decomposition. Batched SYEV shows roughly 2x performance gains on Blackwell compared to the L40S. The GEEV hybrid CPU/GPU algorithm also demonstrates improved speeds across various matrix sizes.

CUDA 13.1 optimizes block-scaled FP4, FP8, and BF16 matrix multiplications on Blackwell. Benchmarks indicate that B200 and GB200 products typically deliver 2x the speed of the H200, with even higher performance gains observed on B300 models.

New library features include an experimental Grouped GEMM API in cuBLAS for Blackwell (supporting FP8/BF16) and a faster sparse matrix-vector multiplication API in cuSPARSE. cuFFT also adds a device API to simplify code generation and metadata querying.

Nsight Systems 2025.6.1 now supports system-wide CUDA tracing and host function node tracing. Hardware-based tracing is now the default setting where supported. Additionally, green context timelines now display tooltips showing SM allocation usage.

Compute Sanitizer 2025.4 introduces compile-time patching with the -fdevice-sanitize=memcheck flag. This integrates error detection directly into NVCC, enabling faster debugging runs and better detection of illegal memory accesses between adjacent allocations.

NVIDIA Nsight Compute 2025.4 adds full profiling support for CUDA Tile kernels. The tool now distinguishes between Tile and SIMT kernels in results and includes a "Tile Statistics" section. It also enables profiling for CUDA graph nodes launched from the device.

Recent cuBLAS updates boost double-precision (FP64) matrix multiplication performance through emulation on Tensor Cores. This is vital for architectures like the NVIDIA GB200 NVL72, allowing efficient handling of FP64 workloads on hardware optimized for AI.

For Ampere and newer GPUs, MPS now supports static SM partitioning via the -S flag. This feature allows developers to create exclusive SM partitions for clients, ensuring deterministic resource allocation and improved workload isolation.

Multi-Process Service (MPS) now features Memory Locality Optimization Partition (MLOPart) for select Blackwell GPUs. This splits a single GPU into multiple logical devices. It improves performance by assigning specific compute and memory resources to distinct partitions.

Green contexts, previously only in the driver API, are now available in the runtime API. These lightweight contexts allow for fine-grained spatial partitioning. You can dedicate specific Streaming Multiprocessors (SMs) to high-priority tasks to isolate latency-sensitive work.

The initial release includes CUDA Tile IR (a virtual instruction set) and cuTile Python (a DSL for kernel authoring). Currently, support is exclusive to NVIDIA Blackwell GPUs (compute capability 10.x/12.x), with a C++ implementation planned for future updates.

To modernize GPU programming, NVIDIA has launched CUDA Tile. This model works a layer above SIMT, allowing developers to define data "tiles" and math operations rather than managing individual threads. The compiler handles the details, abstracting hardware like tensor cores.

NVIDIA CUDA 13.1 marks the most significant update to the platform in 20 years. This release focuses on massive performance gains and new tools for accelerated computing. Key headlines include the new CUDA Tile programming model and runtime API access for green contexts.

My new instrumental, "The Corner Bistro," is out now! This track blends smooth jazz with a chill vibe, perfect for when you need to relax and unwind. Turn up the volume and let the mood settle in. 🎧 Listen to "The Corner Bistro" here: https://t.co/pm8S7Kgl78

#FlowerOnRonin

U.S. farmers are on track for a record corn and soybean harvest. However, the current export outlook for the two products could not be more opposite.

Realty Income Corporation (NYSE: #O) increased its monthly dividend to $0.2695 per share from $0.2690, to be paid on October 15, 2025, to shareholders recorded as of October 1, 2025. The annualized dividend is now $3.234 per share, up from $3.228.

El Salvador added 7 $BTC to its reserves in the last week.