硬體

AI 晶片、GPU、資料中心、邊緣裝置與硬體安全。

Europe Signs Contract to Build LUMI-AI, Unifying AI and HPC Workloads with MI430X GPUs and 256-Core EPYC CPUs

EuroHPC and Bull have signed a €387.8 million contract to deploy LUMI-AI in Finland, built around AMD MI430X GPUs and sixth-generation EPYC CPUs. The system is expected to come online in 2027, but the projected tenfold increase in AI capacity remains a planning estimate, while node counts, interconnects, and storage configurations have yet to be fully disclosed.

2026-09-01

推論系統

Triton 2.72 Fixes Dynamic Batching Starvation and Response Ordering, Upgrades to CUDA 13.4 Inference Stack

NVIDIA Triton Inference Server 2.72.0 fixes an issue that could cause the scheduler to stop dispatching work while requests were still pending and restores the completion-order semantics of `preserve_ordering`. The release also updates its CUDA, TensorRT, and gRPC dependencies, but no TensorRT-LLM image is available, while security warnings remain for vLLM and Ray deployments.

2026-09-01

Puro-2B Publishes a Consumer-Grade GPU Pretraining Recipe, With Full 2B Model Training Costing About $6,900

A Tsinghua University team trained a 2B-parameter model from scratch using RTX 5090 GPUs, blockwise FP8, and MuonH, and released the weights, data, and code. A lower-budget checkpoint outperformed Qwen2-1.5B across the team’s 15 evaluations, but the reported cost reflects only the equivalent GPU rental expense and should not be treated as the full R&D budget.

2026-08-31

推論系統

vLLM 0.28 Adds Disk-Tier KV Cache as Model Runner V2 Supports Disaggregated Encoding, Prefill, and Decode

vLLM 0.28.0 can hierarchically offload the KV cache to disk and exposes an external secondary-storage manager, allowing long-context serving to move beyond the capacity limits of GPU and system memory alone. The release also advances Kimi-K3, DeepSeek V4, speculative decoding, and E/P/D disaggregation, but changes several defaults and dependency interfaces.

2026-08-31

代理工具

Hillock 0.6 Uses Bipolar Hypervectors to Retrieve Agent Memory, Skipping LLM Calls Below a Threshold

Hillock combines factual triples, Hebbian association weights, and 10,000-dimensional hypervectors to determine on local hardware whether enough data exists to answer a question. Version 0.6 adds per-token MaxSim and multi-hop path encoding, but its current gating accuracy of about 56% shows that it remains a research prototype.

2026-08-31

硬體

Sopro V2 Turbo Releases a 120M-Parameter Speech Model with Two-Step Acoustic Solving for Streaming Synthesis on CPUs

Sopro V2 Turbo distills a 500M-parameter teacher into a 120M-parameter model, then uses reflow to reduce acoustic solving from 32 steps to two. The model can produce its first audio chunk in about 300 milliseconds on an M3 CPU and supports in-browser inference, but it currently covers only four European languages and does not include training code.

2026-08-30

推論系統

llama.cpp Reworks Adreno Matrix Multiplication Routing, Speeding Up Prompt Prefill by 9% to 25% Across Two Chip Generations

llama.cpp b10687 avoids inefficient matrix multiplication paths selected by the OpenCL compiler on Snapdragon X2 and older A7X GPUs. Maintainers measured prompt-prefill speedups of roughly 20% to 25% for gpt-oss-20b and 9% for Gemma 3n, while per-token decoding saw no improvement.

2026-08-30

FastH3 Compresses MiniMax H3 to Four DiT Calls, Generating 15 Seconds of Video in 13 Seconds on Eight B200 GPUs

FastVideo combines DMD2 distillation with 90% sparse attention to reduce the base model’s 49 DiT calls to four, while releasing the full weights and LoRA. Its real-time generation figures depend on eight B200 GPUs, specialized VSA kernels, and fully warmed-up tests; a single GPU and mainstream consumer GPUs have yet to achieve comparable speeds.

2026-08-30

AI 安全

NemoClaw 0.0.116 Persistently Binds Sandbox Identity, Tightening Recovery, Takeover, and Deletion Permissions

After strengthening local inference and state-drift validation in 0.0.113, NVIDIA has further bound sandbox, inference service, and lifecycle permissions to persistently verifiable identities, preventing retry or recovery operations from taking over an environment with the same name but a different identity.

2026-08-30

機器人與實體 AI

Microduck Open-Sources Complete Sim-to-Real Reinforcement Learning Stack, Runs Policies via ONNX on a 25 cm Bipedal Robot

Pollen Robotics has released Microduck’s control runtime, MuJoCo simulation environment, PPO training recipe, and seven factory-installed policies under the Apache 2.0 license. Its $399 preorder price lowers the barrier to physical reinforcement learning, but policy generalization, hardware durability, and mass-production delivery still await community validation.

2026-08-30

Puro-2B Publishes RTX 5090 Pretraining Recipe, With a Reported Cost of $6,891 for 1.4T Tokens

Puro-2B integrates a two-stage data curriculum, blockwise FP8, and MuonH optimization on a consumer-grade GPU cluster. The authors report that a shortened $4,370 training run already surpasses Qwen2-1.5B in average score, although the cost accounts only for accelerator rental-equivalent value.

2026-08-29

代理工具

PRAXIST Preserves Experimental Causality Across Generations With Evidence Graphs, Cuts MLE-bench Model Costs to About One-Twelfth of Baseline

PRAXIST has multiple research agents propose changes over successive generations, then records evaluation results, mechanisms, and lineage in an auditable evidence graph to avoid restarting exploration in every round. Its authors report 60 medals across 75 MLE-bench tasks at a model cost of $3,054, but the newly released system’s performance and costs have so far been measured primarily by its development team.

2026-08-29