Aurimas Griciūnas’ Post

I just had my Christmas moment! 🎁 𝗡𝗩𝗜𝗗𝗜𝗔 𝗗𝗚𝗫 𝗦𝗽𝗮𝗿𝗸 has arrived and I finally got my hands on one of these beasts a few days ago. Just a reminder why it is a big deal: 🚀 NVIDIA GB10 Grace Blackwell Superchip. 💾 128 GB of Unified System memory - you can run models with up to 200 billion parameters on GPU locally! ☁️ Supports the same software stack that runs on datacenter GPUs - build on NVIDIA DGX Spark, deploy to cloud seamlessly. 𝗙𝗶𝗿𝘀𝘁 𝗶𝗺𝗽𝗿𝗲𝘀𝘀𝗶𝗼𝗻𝘀. 𝘏𝘦𝘳𝘦 𝘪𝘴 𝘸𝘩𝘢𝘵 𝘴𝘵𝘰𝘰𝘥 𝘰𝘶𝘵 𝘢𝘧𝘵𝘦𝘳 𝘴𝘰𝘮𝘦 𝘵𝘦𝘴𝘵𝘪𝘯𝘨. After some early testing, what really stood out was how smooth the 𝗗𝗲𝘃𝗘𝘅 is: ✅ Everything just works out of the box. ✅ NVIDIA Sync enables remote work on the machine as if I was running everything on my laptop. ✅ Tons of advanced examples that can be easily adjusted for your own workloads. 𝘐𝘯 𝘭𝘦𝘴𝘴 𝘵𝘩𝘢𝘯 𝘵𝘸𝘰 𝘥𝘢𝘺𝘴 𝘰𝘧 𝘵𝘦𝘴𝘵𝘪𝘯𝘨 𝘐 𝘸𝘢𝘴 𝘢𝘣𝘭𝘦 𝘵𝘰: ⚡ Run gpt-oss-120b on the device with as much as ~30 tokens/s throughput. Moved one of my projects to run locally on the device using the model endpoint. 🤖 Migrate two of my multi-agent projects to run on device using LLMs deployed on the machine. Currently fine-tuning performance with new models in the background 𝘞𝘩𝘢𝘵’𝘴 𝘯𝘦𝘹𝘵: 🕐 I have few projects lined up that require fine tuning of reasonably large models, can’t wait to try the machine for this use case. Want to get your own? 👉 https://nvda.ws/47nDdqG Huge shout-out to NVIDIA AI for the collaboration and support! #LLM #AI #DGXSpark

NVIDIA is redefining scalable LLM deployments with the new Run:ai v2.23 and Dynamo integration 🚀 Key Highlights ⚡ Gang Scheduling: All critical pods launch together—no more partial deployments, maximized GPU utilization. 📈 Topology-Aware Placement: Components are co-located for minimal latency and blazing-fast throughput, ideal for multi-node clusters. 💡 Dynamo's Innovative Features: Disaggregated prefill/decode, dynamic GPU scheduling, LLM-aware routing, and multi-tier KV cache management for efficient AI inference at scale. 🔧 Easy Setup: Step-by-step Kubernetes and Helm guide, with sample YAML annotations for preferred/required placement strategies. ✨ “Run:ai + Dynamo make distributed AI deployments predictable, fast, and cost-effective—atomic launches, low-latency networking, and always-on GPUs.” Check out the full blog for tips on deploying Gen AI models at scale! https://lnkd.in/geecvbHC Anish Maddipoti Harry Kim Arun Raman

NVIDIA Unveils DGX Spark — The World’s Smallest AI Supercomputer NVIDIA has launched DGX Spark™, a compact desktop supercomputer delivering 1 petaflop of AI performance and 128GB of unified memory, powered by the Grace Blackwell architecture. Designed for developers, DGX Spark integrates NVIDIA GPUs, CPUs, networking, CUDA libraries, and AI software into one system, enabling local development of AI agents, large language models (up to 70B parameters), and inference workloads (up to 200B parameters) — without relying on cloud infrastructure. Jensen Huang, NVIDIA’s CEO, personally delivered the first unit to Elon Musk at SpaceX, echoing the 2016 delivery of the original DGX-1 to OpenAI. Early adopters include Meta, Microsoft, Google, Hugging Face, Anaconda, and JetBrains, with systems now available through Acer, ASUS, Dell, Lenovo, HP, MSI, and GIGABYTE. “DGX Spark puts a supercomputer on every developer’s desk,” said Huang. “It’s built to power the next wave of AI breakthroughs.” DGX Spark is available now on NVIDIA.com and through global partners.

I believe we’ve just crossed a line in AI hardware… With NVIDIA‘s new Spark system, a full-scale AI supercomputer will now fit on a desk, runs on a normal outlet, and can handle models with up to 200 billion parameters. A few years ago, that setup would have sounded absurd. Now, it ships for $3,999. To me, Spark marks a turning point: the line between personal computer and data-center hardware is dissolving. The same compute power once rationed out by cloud credits or research grants can now sit beside a monitor in a lab, startup, or university office. And with Acer, Asus, Dell, and others releasing their own versions, this feels less like a niche product and more like the beginning of a new category. Call it personal-scale supercomputing. When this level of power becomes accessible, innovation stops being centrally planned. Experiments multiply. So do risks. Frontier-scale capability in millions of hands means breakthroughs and missteps at the same time. I find that exhilarating and unsettling in equal measure. The AI race may soon be defined not by who owns the biggest cluster, but by who learns fastest from their own. Check out the full article here: https://lnkd.in/e3ZRJj4Q #NVIDIA #GenAI #AgneticAI #Infrastructure #LLMOps

NVIDIA and partners （CoreWeave, Microsoft, Nscale） will deploy up to 120,000 Blackwell Ultra GPUs and invest up to £11B in UK data centers by 2026—Britain’s largest AI infrastructure rollout, serving leading models including OpenAI. Nscale will scale 300,000 Grace Blackwell GPUs globally （60,000 in the UK） and, with OpenAI and NVIDIA, launch Stargate U.K. running on Blackwell Ultra by 2026. Microsoft and Nscale plan the UK’s most powerful Azure supercomputer in Loughton with 24,000+ Blackwell Ultra GPUs; CoreWeave adds a Scotland data center; BlackRock to invest up to £500M upgrading NVIDIA-ready sites. NVIDIA accelerates the UK’s quantum stack: OQC/Digital Realty quantum‑AI center on CUDA‑Q, hybrid photonic‑AI work with ORCA/Imperial/Poznan, GPU‑accelerated error correction at Edinburgh, AI‑controlled hardware at Oxford, and SEEQC/NQCC QPU‑GPU integration. New techUK–NVIDIA R&D hub and QA-led upskilling bring DLI courses and DGX Cloud access; Isambard‑AI fuels national projects from UK‑LLM to sovereign health and science models. 🔔 Follow us for daily AI updates! 📘 Facebook: https://lnkd.in/gxDt7PJa 📸 Instagram: https://lnkd.in/gmYfWDbF #NVIDIA #UKAI #AIInfrastructure #OpenAI #AIGenerated #CreatedWithAI

NVIDIA B200 vs. H200: A Guide to Choosing the Right Model for Your Needs https://lnkd.in/gBdg2D3p 🚀 Unleashing AI Potential: NVIDIA’s B200 GPU 🚀 The launch of NVIDIA’s B200 GPU is revolutionizing the AI landscape. Featuring the innovative Blackwell architecture, this powerhouse boasts 208 billion transistors and an astonishing 20 PFLOPS of FP4 compute power. But is it worth the upgrade? Key Highlights: Performance Leap: 2.25x throughput over the H200. 4x faster training for large language models (LLMs). 15x higher inference for real-time applications. Rigorous Testing: Compute Performance: Nearly 2x throughput advantage in GEMM benchmarks. Scaling Efficiency: Enhanced communication in large-scale training setups. Real-World Impact: 50% higher throughput for 7B models. Over 70% increase for 70B models. If you’re involved in large-scale AI deployments, the B200 is a game-changer. 👉 Dive into the details and experience the power of the B200 on HPC-AI.COM! Share your thoughts and let’s explore how to fuel AI innovation together! 🌟 Source link https://lnkd.in/gBdg2D3p

NVIDIA's AI Supremacy: Navigating the Surge of a Tech Titan The central narrative is NVIDIA's strategic pivot and technological dominance in the AI chip market. Originally known for its GeForce gaming GPUs, the company's CUDA architecture proved exceptionally well-suited for the parallel processing required by AI and machine learning models. This foresight positioned NVIDIA as the primary supplier of the 'picks and shovels' for the AI gold rush. The insatiable demand for its data center GPUs, like the H100 and the upcoming Blackwell platform, from cloud providers and enterprises has led to exponential revenue and profit growth, propelling its market capitalization into the trillions. The market is currently gripped by a powerful combination of FOMO (Fear Of Missing Out) and a genuine belief that AI represents a paradigm shift on par with the internet. NVIDIA is viewed not just as a semiconductor company, but as the foundational enabler of this new era. This sentiment has led investors to prioritize growth potential over traditional valuation metrics, willingly paying a premium for a stake in the company's future. The narrative is so compelling that it has created a 'gravity' effect, pulling in capital from all corners of the market. Source - https://lnkd.in/dc2c8BPR

🤑⚡ Nvidia Fuels OpenAI’s $100B Compute Push OpenAI and Nvidia have announced what’s being called the biggest AI infrastructure project in history — a plan to deploy 10 GW of Nvidia systems (millions of GPUs) to power the next era of intelligence. The Rundown: ✨ $100B phased investment tied to GPU deployment. ⚡ First gigawatt goes live in H2 2026 on Nvidia’s Vera Rubin platform. 🌍 Nvidia becomes OpenAI’s “preferred strategic compute & networking partner.” 📊 Deal complements existing ties with Microsoft & Oracle. Why it matters: ⚙️ Secures compute + power capacity for OpenAI’s next-gen models. 🧠 Locks Nvidia into AI’s most valuable customer pipeline. 🚀 Raises debate: true innovation funding or “infinite money loop” cycling capital across AI giants? Quote: “The biggest AI infrastructure project in history” — positioning Nvidia & OpenAI at the center of the AI compute race. ❓Is this the dawn of AI’s industrial scale — or the start of an unsustainable AI arms race? #AI #OpenAI #Nvidia #GPUs #AIInfrastructure #FutureOfAI

Nvidia just flipped the AI script: $100B to power 10 GW of OpenAI compute—first 1 GW goes live in H2’26 on the new Vera Rubin platform. Think millions of GPUs lighting up the next model wave. � This isn’t a capex plan—it’s an AI factory blueprint. Jensen Huang says it’s 4–5 million GPUs worth of scale, the largest AI infra build ever. Markets agreed, with a sharp pop as the deal hit the tape. �� The spicy bit: “circularity.” OpenAI could use Nvidia’s cash to buy Nvidia systems—governance questions, sure—but it also locks the ecosystem to Team Green for years. � Ripple effects are massive: 10 GW means grid‑level power, multi‑billion networking, liquid cooling everywhere, and supply chains stretched from optics to substations. �� If AI is the new industrial revolution, this is the moment the assembly line went online. Winners will speak compute, power, and bandwidth natively. �Save this for earnings season: watch GPU units, network attach, and power deals as the leading indicators of who’s building real AI factories. �#AI #Nvidia #OpenAI #GPUs #DataCenters #Rubin #Blackwell #AIInfrastructure

Mediatek x NVIDIA Power the Next Era of Desktop AI MediaTek co-designed the GB10 Grace Blackwell Superchip for the NVIDIA DGX Spark, delivering 1 PFLOP of AI performance in a desktop-sized supercomputer. DGX Spark enables developers to prototype and fine-tune large AI models with unmatched power efficiency and performance. Launching October 15, it’s set to redefine desktop AI computing. #AIInnovation #MediaTek #NVIDIA #Supercomputing #EdgeAI https://lnkd.in/gSg7Bc-f