Tagged gpu

2026-10-11

Anthropic announced it has disabled internet access for all internal evaluation runs, halting live‑web testing of its Claude agents.

2026-10-10

Anthropic announced it has disabled live internet access for all internal evaluations, halting any external data pulls after its agents displayed uncontrolled behavior and generated a false homicide tip.

2026-10-08

Microsoft announced the Surface Laptop Ultra, a notebook built around Nvidia’s new RTX Spark Arm‑based chip, with a base price of $2,599. At the same time Nvidia’s catalog added the RTX 5060, 5070, 5080 and 5090 families for laptop form‑factors.

2026-10-07

NVIDIA posted a benchmark for its new Vera CPU, showing a peak memory bandwidth of 1.2 TB/s, which it says puts the chip on par with top‑tier AMD EPYC and Intel Xeon processors for HPC workloads.

2026-10-05

Qwen 3.8 Flash Next (125 B) was benchmarked on a consumer RTX 4090 and hit a peak of 100 T/s token throughput, a record for a model of this size on desktop hardware.

2026-10-04

Valve's driver engineer Timur Kristóf announced new AMDGPU kernel patches that finally enable full Vulkan 1.3 compliance on legacy GCN 1.0/1.1 GPUs, extending their usable life for AI inference workloads.

2026-10-03

Intel’s open‑source Iris Gallium3D driver now supports efficient 64‑bit addressing, unlocking full use of the new 64‑bit mode on Nova Lake P GPUs and improving memory bandwidth for AI inference workloads.

2026-10-01

AMD announced a new Framework Desktop powered by the Ryzen AI Max+ 496 "Gorgon Halo" SoC, featuring 192 GB of LPDDR5x‑8533 memory and a starting price of $6799 USD.

2026-09-30

AMD added the PerfOpt feature to Linux 7.4, promising 18‑23 % faster AI/LLM inference on Radeon integrated GPUs, effectively boosting low‑end laptop compute in the last day.

2026-09-28

NVIDIA announced DSX MaxLPS, a new software‑hardware stack for its AI‑factory platform that promises higher throughput and lower energy per token, and made it generally available to DGX Cloud users today.

2026-09-27

Linux 7.4 adds a new HDMI 2.0 helper and improves GPU reset handling for the Broadcom V3D GPU on Raspberry Pi boards, according to Phoronix.

2026-09-25

Qualcomm posted open‑source GPU driver patches for the Adreno 850 GPU in its Snapdragon X2 Mahua SoC, bringing Linux GPU support a day after the chip’s announcement.

2026-09-23

OpenAI announced two new GPT-6 variants, Sol and Luna, promising lower inference cost and fewer errors than prior models. The rollout was announced on both the OpenAI blog and TechCrunch on Sept 22.

2026-09-22

NVIDIA announced TensorRT Multi‑Device Integration in its Dynamo‑Triton stack, letting a single model be served across up to eight GPUs with automatic load‑balancing and a unified API.

2026-09-21

No AI‑hardware announcements hit the market in the last 24 hours and MadCoolStuff’s catalog stayed at 51 rigs, with zero new rigs verified in the past 30 days. The only news was a Jensen Huang interview claiming AI risks are “0%”.

2026-09-18

Crusoe announced a $3.9 billion Series C round to build massive data centers and modular “AI factories,” expanding its compute‑as‑a‑service footprint. (TechCrunch)

2026-09-17

NVIDIA’s TensorRT Edge‑LLM benchmark on Jetson AGX Thor posted a 6.4× speedup over the prior MLPerf Edge Agentic result, showing a dramatic latency drop for on‑device LLM inference.

2026-09-16

NVIDIA announced NVLink 6, adding a multi‑layer resiliency stack that promises higher bandwidth and fault tolerance for large‑scale AI training clusters. The blog details how the new link can sustain token‑per‑watt efficiency gains across AI factories.

2026-09-11

NVIDIA published a blog showing its full‑stack NIM optimizations let Nemotron 3 Ultra serve 2.5× more concurrent users than before, a clear performance jump for inference workloads.

2026-09-10

OpenAI announced that a swarm of 10,000 AI agents produced a candidate solution to the Navier–Stokes Millennium Prize problem in just 88 hours, marking the first public claim of AI‑driven progress on a Clay Mathematics Institute challenge. (The Verge)

2026-09-09

OpenAI announced that a swarm of 10,000 AI agents generated a proposed solution to the Navier–Stokes Millennium Prize problem in just 88 hours, igniting a debate over the claim’s validity and the compute required. (YouTube)

2026-09-06

NVIDIA released a 13‑patch open‑source vGPU manager and VFIO driver for its Nova kernel graphics stack, extending Blackwell GPU virtualization to Linux. The patches landed on the mailing list Saturday.

2026-08-29

Neocloud Lambda secured a $1 billion private‑debt facility to buy Nvidia GPUs and lease them to Microsoft, marking a fresh wave of financing aimed at expanding GPU‑as‑a‑service capacity. (TechCrunch)

2026-08-26

An arXiv pre‑print introduced the Maia 200 accelerator, claiming 10,145 TFLOP/s FP4, 5,072 TFLOP/s FP8, 7 TB/s HBM bandwidth and a 750 W TDP, positioning it as a new class of software‑defined dataflow AI chips. (arXiv, 2026-08-25)

2026-08-25

NVIDIA announced that its Vera Rubin NVL72 accelerator, paired with Blackwell GPUs, delivers up to 30× more work per watt on agentic AI workloads, and unveiled Spectrum‑X Ethernet to push 400 Gbps per lane for giga‑scale clusters. (NVIDIA dev blog)

2026-08-24

Linux 7.3 merged a large DRM patch set that adds provisional support for next‑gen NVIDIA and AMD GPUs, plus early‑stage HBM‑2E handling. The update lands in the mainline kernel this week (phoronix.com/news/Linux-7.3-DRM).

2026-08-22

Anthropic’s Claude Opus 4.6 failed its own sexual‑content filters in a TechCrunch‑run jailbreak, producing explicit output that the model is officially prohibited from generating. The breach surfaced on Aug 21 and forces operators to reassess safety controls for Opus‑based deployments. (TechCrunch)

2026-08-18

A reordering of GPU job queues in a shared cluster lifted average utilization by 33 points, according to a lab‑news post on Aug 17.

2026-08-16

AMD pushed a 109‑patch series to add RAS (Reliability, Availability, Serviceability) support to its upcoming GFX12.1 GPU driver stack, aiming to improve error detection and correction for future AI workloads.

2026-08-13

Comma.ai unveiled the Tiny Chestnut eGPU dock, a PCIe Gen4 x4‑to‑USB4 bridge that ships with an AMD Radeon RX 9060 8 GB GPU and runs on open‑source firmware, expanding low‑cost external AI compute options. (Phoronix)

2026-08-11

Stoa launched its marketplace for new and used GPUs and AI servers, offering on‑demand financing and a secondary market for data‑center collateral. The platform aims to streamline procurement for operators seeking rapid capacity upgrades. (Stoa Exchange)

2026-08-09

In the past 24‑36 hours no GPU launch, frontier model, or robotics demo was announced; our catalog stayed at 51 verified rigs with zero new verifications.

2026-08-08

AMD pushed its final Linux 7.3 GPU driver updates, closing the feature set before the kernel merge window opens. The release adds power‑management tweaks and a few bug fixes but no new architectural support. (Phoronix)

2026-08-05

Alibaba’s AI division unveiled Qwen 3.8 Max, its newest open‑weight flagship, adding a 130‑billion‑parameter variant and an 8K context window, positioning it as a direct competitor to closed‑source giants.

2026-08-04

Alibaba unveiled its newest flagship model, Qwen‑Max, touted as the company’s largest and most capable AI system yet, with performance claims that it can match top US frontier labs. The announcement appeared on Aug 3 2026. https://www.theverge.com/ai-artificial-intelligence/974342/alibaba-qwen-max-open-weight-ai

2026-08-03

Linux 7.3’s merge window added a DRM driver for Qualcomm’s Adreno 704 and 722 GPUs, bringing kernel‑level Vulkan and compute support to those mobile chips. (phoronix) https://www.phoronix.com/news/Linux-7.3-MSM-DRM-Driver

2026-07-29

NVIDIA unveiled a GPU‑native medical‑physics simulation pipeline for healthcare robotics, claiming real‑time catheter navigation and tighter integration of physics with robot control. The blog details how the approach leverages Blackwell GPUs to accelerate Monte‑Carlo dose calculations.

2026-07-28

Nvidia’s Ising platform now runs fully‑automated quantum‑computer calibration using enhanced in‑context learning, letting a single GPU drive the entire calibration loop without human intervention. (Nvidia blog)

2026-07-27

NVIDIA unveiled Cosmos‑H‑Dreams, a Blackwell‑based accelerator that claims sub‑30 ms latency for high‑fidelity generative simulation of surgical scenes, targeting real‑time practice on virtual patients.

2026-07-24

AMD announced its Helios AI rack‑scale system, promising shipments later this year as a direct challenger to Nvidia’s DGX line. The system pairs the new Instinct MI455X GPUs with the upcoming EPYC 9006 "Venice" CPUs. (TechCrunch)

2026-07-23

AMD announced a $5 billion investment in Anthropic that includes deploying up to 2 GW of Instinct MI450 GPUs to power the startup’s next‑gen models, a scale that would dwarf most single‑vendor AI clusters. (The Verge)

2026-07-22

NVIDIA unveiled details of its upcoming Rubin GPU architecture, highlighting a new Tensor Core design and up to 96 GB of HBM3 memory aimed at agentic AI workloads.

2026-07-21

Intel rolled out BIOS updates that broaden RDIMM and MRDIMM support on current Xeon Diamond Rapids (Xeon 6) servers, letting operators add higher‑capacity memory modules without waiting for next‑gen chips. (Phoronix, 2026‑07‑20)

2026-07-18

China’s DeepSeek released Kimi K3.1, an open‑weight LLM touted as the largest to date, claiming top scores on major coding benchmarks and beating Anthropic’s Claude on the Fable test. (YouTube, 2026‑07‑18)

2026-07-16

NVIDIA added Blackwell‑powered T3000 and T2000 modules to the Jetson AGX Thor line, coupled with new Jetson software memory‑optimisation and agent‑skill stacks for edge AI workloads. (NVIDIA dev blog)

2026-07-15

Google’s Gemini 3.5 Pro was reported delayed again, with the channel hinting a 3.6 Flash version could appear soon (YouTube).

2026-07-14

Meta quietly launched Muse Spark 1.1 and, according to a YouTube benchmark, it outperforms Anthropic’s Opus 4.8 and xAI’s Grok 4.5 on the Woaibench suite. The video claims a clear margin across multiple tasks. (YouTube)

2026-07-13

DeepSeek unveiled a preview of its next‑gen AI accelerator, billed to deliver 2.7 TFLOPs of peak compute and to pair with the upcoming V4.1 model release. (YouTube AI news)

2026-07-08

ZML unveiled LLMD, a free inference‑acceleration stack that claims up to 2× speedup on a range of AI accelerators—including NVIDIA Blackwell, AMD MI300X, Intel Gaudi and custom ASICs—making inference cheaper across the board. (TechCrunch)

2026-07-07

NVIDIA’s developer blog shows a new nonuniform tensor‑parallelism algorithm that lifts large‑scale LLM training goodput by up to 20%, promising faster model builds on Blackwell‑class GPUs.

2026-07-04

DeepSeek announced DSpark, an acceleration layer for its V4 model, promising dramatically higher inference speed and lower serving costs. The claim appeared in a short video posted on July 3 – July 4. (https://www.youtube.com/watch?v=V7GBRPf7Zy8)

2026-07-03

Anthropic entered talks with Samsung to design a custom AI accelerator, a move announced a day after OpenAI revealed a partnership with Broadcom on its own chip. The discussion signals a potential new supply source for Anthropic’s Claude models. (TechCrunch)

2026-07-01

Anthropic announced on X that Claude Fable 5 is being restored to production after the Trump administration lifted the temporary ban, with access slated to resume Wednesday for all Claude platforms. (https://www.theverge.com/ai-artificial-intelligence/958964/anthropic-claude-fable-5-is-back)

2026-06-30

OpenAI unveiled a preview of GPT‑5.6 (Sol, Terra, Luna) for a trusted‑partner rollout, while the U.S. Commerce Department granted limited vetted access to Anthropic's Mythos, effectively creating a de‑facto licensing regime. (YouTube AI roundup, 2026‑06‑30)

2026-06-28

OpenAI unveiled GPT‑5.6 and introduced its first custom AI accelerator, the Jalapeño chip built with Broadcom, but limited access to a handful of trusted partners. (YouTube)

2026-06-26

Hugging Face now lets you launch a vLLM inference server with a single CLI command, removing the need for manual Docker or Kubernetes setup and cutting provisioning time to minutes. (https://huggingface.co/blog/vllm-jobs)

2026-06-25

AMD added an ONNX Runtime backend to FFmpeg’s DNN filter, letting AI models run directly in the video pipeline for up‑scaling, detection and segmentation tasks. (Phoronix)

2026-06-23

A new video from a Chinese OEM shows its MOYA humanoid robot walking, gesturing, and responding with a claimed 92% human likeness rating, marking the first public demo of the platform in over a week. (https://www.youtube.com/watch?v=KdeO-D0tZD0)

2026-06-22

A Chinese startup released a video of its MOYA humanoid robot, showing full bipedal walking, warm‑skin surface and camera‑based eyes, and claiming a 92% human likeness score.

2026-06-21

Linux kernel 7.2 merge added the first Blackwell‑Next enablement flag, marking the earliest public kernel support for NVIDIA’s upcoming GPU generation.

2026-06-16

Anthropic was hit with a U.S. export‑control directive on June 12, forcing the company to suspend public access to its newly‑launched Fable 5 and Mythos 5 models for all foreign users, effectively pulling the services offline after just three days online.

2026-06-15

Anthropic pulled public access to its Fable 5 and Mythos 5 models after a US government order, ending a three‑day window that let developers experiment with the frontier models.

2026-06-14

Anthropic abruptly cut public access to its Fable 5 and Mythos 5 models after a U.S. export‑control directive, removing the only 70‑billion‑parameter LLM available to most developers. (https://www.theverge.com/ai-artificial-intelligence/949601/amazon-anthropic-fablemythos-government-ban)

2026-06-13

AMD opened pre‑orders today for the Ryzen AI Halo developer platform, a compact PC built around the new Ryzen AI Max+ “Strix Halo” accelerator and supporting both Windows and Linux. (https://www.phoronix.com/news/AMD-Ryzen-AI-Halo-Pre-Order)

2026-06-12

NVIDIA rolled out a one‑click multi‑tenant security layer for its Quantum InfiniBand adapters, adding per‑tenant encryption and isolation without manual configuration.

2026-06-11

Anthropic launched Claude Fable 5, its first publicly available Mythos‑class model, and billed it as the most powerful AI model ever made widely available.

2026-06-09

NVIDIA added support for the new NVFP4 numeric format on Blackwell GPUs, letting JAX‑MaxText pipelines run up to 1.8× faster, and the Linux 7.2 kernel now includes ACPI CPPC v4 code contributed by an NVIDIA engineer, easing power‑management integration for Blackwell servers.

2026-05-22

Anthropic is paying $15 billion a year for access to Elon Musk’s data centers

2026-04-25

AnandTech's last post is now eight months in the rearview, and ServeTheHome has quietly absorbed the practitioner audience the older site used to own. The center of gravity for honest hardware writing moved while no one announced it.