A developer compared Apple's Mac Studio M5 Ultra and NVIDIA's DGX Spark as local AI machines, finding the Mac Studio offers 4.4x the memory bandwidth for faster LLM inference, while the DGX Spark prov…
NVIDIA introduced shadow engine recovery, a preview feature in NVIDIA Dynamo, that restores LLM inference capacity in seconds by keeping a fully initialized shadow engine idle on the same GPUs, using …
Apple's Metal GPU glossary 'metalworking' targets developers familiar with CUDA but new to Apple Silicon, mapping every concept from GPU cores to MLX's architecture with CUDA equivalents and real kern…
Nokia's chief technology officer Pallavi Mahajan said the company's $1 billion investment from Nvidia will enable GPU-based radio access network (RAN) software that can run compute-intensive machine-l…
Nvidia's Jetson Orin edge AI module, delivering up to 275 TOPS, is widely used in robotics and embedded systems, including drones and computer vision applications, according to Nvidia's official speci…
Nvidia announced at Hot Chips 2026 that CUDA will officially support RISC-V CPUs, making RISC-V the third supported host architecture alongside x86 and ARM, with strict requirements including the RVA2…
Nvidia is extending CUDA support to RISC-V CPUs, requiring server-grade features such as RVA23 compliance, ACPI, PCIe coherency, and peer-to-peer PCIe, according to a Hot Chips 2026 talk. The company …
A developer's deep dive on tamiz.pro examines the shift from high-performance LLM serving engines like SGLang to local-first logic frameworks such as OpenLogi, which prioritize data privacy, determini…
A developer has released jina-embeddings-v4, a self-hosted server for the jina-embeddings-v4 embedding model with an OpenAI-compatible /v1/embeddings endpoint. The server runs on a single NVIDIA GPU a…
A bug is a violation of a specification, and no specification exists that prefix caching's variable logits violate, according to an analysis of vLLM and SGLang issues. The vLLM PR #34046 adds an opt-i…
Higgsfield AI, a generative AI company with $500M in annual revenue run rate and 25M+ users, is hiring an ML Systems Performance Engineer (MFU) for its Almaty, Kazakhstan office. The role focuses on o…
A technical essay argues that mastering Physics-Informed Neural Networks (PINNs) requires a hybrid background rather than a single major, recommending Computer Science or Mathematics paired with elect…
Ferrox v0.9.1, a pure-Rust GGUF inference engine, shipped today with MoE prefill on Apple Metal 2.4x faster, closing the gap to llama.cpp from 2.62x behind to 1.11x on OLMoE-1B-7B. The update also imp…
A developer's final part of a CUDA/ROCm optimization series details implementing Mixture of Experts (MoE) inference, including a custom top-k gating kernel, expert parallelism with All-to-All communic…
A new CUDA kernel correctness harness compiles kernels with nvcc, loads them into Python via ctypes, and checks outputs against NumPy baselines, with timing via CUDA events. All 23 tests pass on an NV…
At ModCon 2026, Modular unveiled new capabilities for its MAX platform, emphasizing openness and accelerator portability, allowing AI workloads to run across diverse hardware. The company also announc…
Nvidia CEO Jensen Huang told customers on August 13, 2026, that A100 GPUs will remain effective through 2029, while the company simultaneously pushes its Blackwell, Hopper, and upcoming Rubin architec…
Cerebras Systems claims its Wafer-Scale Engine 3 (WSE-3) delivers 1,800 tokens per second at 23 kW for Llama 3 70B inference, versus about 850 tokens per second at 56 kW for an 8×H100 DGX system, yiel…
A developer detailed a CUDA-based approach to finding the closest island to any ocean point using spherical geometry and GPU acceleration. The method leverages the Haversine formula and spatial indexi…
A developer benchmarked Hashcat on an NVIDIA GB10 GPU inside a DGX Spark system, reporting speeds of 81.16 GH/s for MD4, 53.54 GH/s for MD5, and 16.70 GH/s for SHA1. The benchmark used Hashcat v7.1.2-…