All news with #nvidia tag

🔒 Researchers at the University of Toronto demonstrated GPUBreach, a GPU-targeted Rowhammer technique that flips bits in GDDR6 to corrupt GPU page tables and subvert device memory controls. An unprivileged CUDA kernel can obtain arbitrary read/write access to GPU memory and then exploit NVIDIA driver flaws to escalate to CPU privileges and spawn a root shell. The work, due at IEEE S&P 2026, includes technical materials and shows impacts from key leakage to ML model manipulation.

⚠️ New research describes GPUBreach, a set of GDDR6 RowHammer techniques that corrupt GPU page tables to gain arbitrary GPU memory read/write and, in GPUBreach's case, full host control. The work shows chained GDDR6 bit-flips can corrupt trusted driver state and trigger kernel memory-safety bugs in NVIDIA drivers even with the IOMMU enabled. Related efforts (GDDRHammer, GeForge) also achieve GPU-side arbitrary read/write, though some require IOMMU to be disabled. Enabling ECC reduces risk but is not a guaranteed mitigation for all platforms.

⚠️ A new attack called GPUBreach demonstrates that Rowhammer-induced bit flips in GDDR6 memory can corrupt GPU page tables and allow an unprivileged CUDA kernel to gain arbitrary GPU memory read/write access. The University of Toronto team showed this capability can be chained into CPU-side privilege escalation by exploiting memory-safety bugs in the NVIDIA driver, potentially yielding a full system compromise up to a root shell. Critically, the attack works with IOMMU enabled and remains unmitigated on consumer GPUs without ECC. Full technical details and a reproduction package will be published on April 13.

⚛️ Microsoft announces an AI for nuclear collaboration with NVIDIA to deliver an end-to-end, AI-powered foundation for nuclear project delivery. The initiative pairs Microsoft Azure, generative AI for permitting, and NVIDIA simulation and AI stacks to speed design, streamline licensing, and improve operations via Digital Twins. Early adopters — including Aalo Atomics, Southern Nuclear, and Idaho National Laboratory — report major time and cost reductions while preserving regulatory traceability and security.

🚀 Google Cloud and partners announced that llm-d has been accepted into the CNCF Sandbox to promote open, accelerator-agnostic standards for distributed LLM inference. As a founding contributor alongside Red Hat, IBM Research, CoreWeave, and NVIDIA, Google emphasizes running any model on any accelerator in any cloud without vendor lock-in. GKE Inference Gateway now integrates the llm-d Endpoint Picker (EPP) to enable model-aware routing that optimizes for KV-cache hits, inflight requests, and queue depth, yielding concrete production gains in Vertex AI tests. Complementary work on the Kubernetes LeaderWorkerSet (LWS) API and vLLM extensions for Cloud TPUs targets scalable multi-node orchestration and up to 5x throughput improvements.

⚡ This technical guide explains how to extract peak training performance on Ironwood TPUs using the JAX and MaxText ecosystems. It highlights native FP8 support, Tokamax kernels for on-chip efficiency, offloading collectives to SparseCore processors, VMEM tuning, and sharding strategies (FSDP, TP, EP, CP, and hybrids). Practical flags and libraries such as Qwix and Tokamax are recommended for implementation.

⚡ AWS now supports NVIDIA Inference Xfer Library (NIXL) with Elastic Fabric Adapter (EFA) on all EFA-enabled EC2 instances and regions. This integration accelerates disaggregated LLM inference by increasing KV-cache throughput, lowering inter-token latency, and optimizing KV-cache memory use between prefill and decode nodes. NIXL interoperates with frameworks such as NVIDIA Dynamo, SGLang, and vLLM. Supported versions are NIXL 1.0.0+ and EFA installer 1.47.0+, available at no extra cost.

🔐 At the Nvidia GTC conference CEO Jensen Huang introduced NemoClaw, a secure runtime for running OpenClaw-style agents built on the Nvidia Agent Toolkit and the broader NeMo ecosystem. Central to the offering is the open-source OpenShell runtime, which provides kernel-level sandboxing and a “privacy router” to monitor and block unsafe communications. Nvidia says NemoClaw is hardware-agnostic though optimized for its own microservices, and aims to make edge agent deployment viable for enterprises while researchers inspect it for CVE-level flaws.

🚀 Microsoft and NVIDIA announced deeper integration at NVIDIA GTC, extending Microsoft Foundry to support NVIDIA Nemotron models and to simplify building production agents. New Azure AI infrastructure optimized for inference and reasoning will bring Vera Rubin NVL72 into liquid‑cooled datacenters and add initial support on Azure Local. Foundry Agent Service, Control Plane observability and a Voice Live API preview aim to accelerate prototype‑to‑production paths, while Fabric–Omniverse links and a public Physical AI Toolchain support simulation‑to‑operations workflows.

🔒 Check Point has integrated with NVIDIA DSX Air’s cloud-based testing environment to let organizations pre-validate security-aware AI data center designs before deploying hardware. The capability enables large-scale simulation and end-to-end validation of AI Factory deployments across compute, networking, orchestration and security. By validating integrations, configurations and automation in advance, teams can reduce resource intensity and accelerate secure rollouts.

🚀 At NVIDIA GTC 2026, Google Cloud announced an expanded co‑engineering partnership with NVIDIA centered on the new Google Cloud AI Hypercomputer, designed to address the infrastructure demands of agentic and large-scale MoE workloads. The updates include momentum for G4 VMs powered by NVIDIA RTX Pro 6000 Server Edition, a preview of fractional G4 VMs using NVIDIA vGPU, and planned support for NVIDIA Vera Rubin NVL72 rack systems. Software integrations such as NVIDIA Dynamo with GKE Inference Gateway, Vertex AI Model Garden additions, and a public sector AI startup accelerator target lower latency, higher throughput, and more flexible consumption for inference and training.

🚀 Starting today, Amazon Web Services has made EC2 G7e instances, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, available in Asia Pacific (Seoul) and Europe (Spain). These instances deliver up to 2.3x inference performance versus G6e and support up to eight GPUs (96 GB each), 192 vCPUs, and 1,600 Gbps of networking for demanding LLM, multimodal, and spatial computing workloads. G7e also supports NVIDIA GPUDirect P2P and GPUDirect RDMA with EFA for accelerated multi-GPU and multi-node performance and can be purchased as On-Demand, Spot, or via Savings Plans. Provisioning is available through the AWS Management Console, CLI, and SDKs.

🚀 Amazon Web Services has launched EC2 G7e instances in the Asia Pacific (Tokyo) region, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. These instances deliver up to 2.3x inference performance compared to G6e, support up to eight GPUs with 96 GB per GPU, and provide up to 192 vCPUs and 1600 Gbps networking. They include NVIDIA GPUDirect P2P and GPUDirect RDMA with EFA in EC2 UltraClusters, and are available as On-Demand, Spot, or via Savings Plans.

🚀 Amazon Web Services has launched Amazon EC2 G6e instances powered by NVIDIA L40S Tensor Core GPUs in the Middle East (UAE) Region. These instances offer up to eight L40S GPUs with 48 GB each, third-generation AMD EPYC processors, up to 192 vCPUs, 1.536 TB system memory, 7.6 TB local NVMe storage, and up to 400 Gbps networking. G6e is optimized for ML workloads including LLMs, diffusion models for image/video/audio generation, and large-scale spatial computing and digital twins. Instances are available across multiple regions and purchasable via On-Demand, Reserved, Spot, and Savings Plans.

🚀 Amazon Web Services has made EC2 G6e instances powered by NVIDIA L40S Tensor Core GPUs available in the Middle East (UAE) Region. These instances support up to eight L40S GPUs (48 GB each), third-generation AMD EPYC processors, up to 192 vCPUs, 1.536 TB memory, 400 Gbps networking and 7.6 TB NVMe storage. They target machine learning and spatial computing workloads — from deploying large language and diffusion models to building immersive 3D simulations and digital twins. G6e instances are available via On‑Demand, Reserved, Spot and Savings Plans.

🚀 Amazon EC2 G7e instances, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, are now available in US West (Oregon). These instances deliver up to 2.3x inference performance compared to G6e and offer up to eight GPUs with 96 GB per GPU, 5th Gen Intel Xeon processors, 192 vCPUs, and up to 1600 Gbps of networking. They target LLMs, agentic and multimodal generative AI, spatial computing, and workloads that require combined graphics and AI processing, while supporting GPUDirect P2P and RDMA in UltraClusters for low-latency multi-GPU scaling.

🤖 Maia 200 is Microsoft’s new first‑party AI inference accelerator, fabricated on TSMC’s 3nm process and optimized for low‑precision tensor compute. It combines native FP8/FP4 tensor cores with 216GB of HBM3e at 7 TB/s, 272MB on‑chip SRAM and specialized data‑movement engines to raise token throughput and utilization. Microsoft positions Maia 200 as delivering over 10 petaFLOPS (FP4), about 30% better performance‑per‑dollar versus its prior fleet, and higher FP8 performance than competing hyperscaler accelerators. Deployed in US Central with Azure integration and an SDK preview, Maia 200 targets inference for services such as Microsoft 365 Copilot and OpenAI GPT‑5.2.

🚀 This post describes two validated deployment recipes for serving large Mixture-of-Experts (MoE) models on Google Cloud's A4X machines using NVIDIA Dynamo. The recipes provide throughput- and latency-optimized configurations that exploit the 72‑GPU GB200 NVL72 rack fabric, WideEP/DeepEP parallelism, global KV cache, and GKE-aware rack-level scheduling. Performance validation reports >6K tokens/sec/GPU for the throughput recipe and a 10ms median inter-token latency for the latency-optimized recipe.

🚀 Amazon EC2 G7e instances are now generally available, powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. G7e delivers up to 2.3x inference performance versus G6e and supports configurations with up to 8 GPUs (96 GB each), 5th Gen Intel Xeon processors, 192 vCPUs, and up to 1600 Gbps of Elastic Fabric Adapter networking. Designed for LLMs, multimodal and spatial computing workloads, G7e includes NVIDIA GPUDirect P2P and RDMA support in EC2 UltraClusters and is available in US East (N. Virginia) and US East (Ohio) as On‑Demand, Spot, or via Savings Plans.

⚠️ Researchers at Palo Alto Networks' Unit 42 disclosed critical weaknesses in the NeMo, Uni2TS and FlexTok Python libraries used with Hugging Face models, where malicious code can be hidden in model metadata and executed automatically when a manipulated file is loaded. The root cause is the use of Hydra's instantiate(), which accepts arbitrary callables and arguments and can therefore permit remote code execution if metadata is untrusted. Vendors including NVIDIA, Salesforce and the maintainers of FlexTok have issued fixes and CVE assignments; users should upgrade affected libraries and audit models before loading.