Yunan Jiang

We released an improved speech-to-speech model and new API capabilities to help build production grade voice agents. Highlights: - Realtime API is now GA - gpt‑realtime: our most advanced speech-to-speech model is more natural, expressive, and better at following complex instructions (even mid-sentence language switches) and tool calling - New integrations: support for remote MCP servers, image input, plus phone calling via SIP - Two fresh voices: Cedar and Marin - 20% price reduction https://lnkd.in/gnzHiNjd

866

23 Comments

From rockets to AI. Nine years after the original NVIDIA DGX-1 handoff, Jensen Huang delivered a brand-new DGX Spark to Elon Musk at Starbase. The DGX Spark is a petaflop powerhouse built for creators, researchers, and developers — a desktop-sized supercomputer with five times the performance of the first DGX-1. 🚀✨ Filmed at SpaceX rocket factory in Texas, this moment captures a conversation between the two visionaries. ➡️ https://bit.ly/4o7oqGw 📹 https://bit.ly/42GxidS

26

The KV Cache bottleneck has been one of the biggest challenges in scaling inference. NVIDIA’s latest Dynamo + NIXL update changes the game — and we're proud to be at the heart of it. Here’s what stands out: ⚡ We're 8x faster than competitors in NVIDIA’s new NIXL benchmarks 🧠 WEKA is the only vendor delivering DRAM-level performance 🔌 We contributed the GDS plugin framework that NIXL runs on This isn’t just about faster numbers. It’s about making inference at scale more sustainable, more efficient, and ready for what’s next in AI. Read this blog post from NVIDIA to see how we're helping solve one of the toughest challenges in AI infrastructure: https://weka.ly/42AI1pS

35

We’ve teamed up with Google for Developers to launch a new recipe for disaggregated inference with NVIDIA Dynamo. NVIDIA Dynamo separates the key phases of inference on distinct GPUs—boosting performance and cutting costs. Try the Dynamo recipe on A3 Ultra GPU-accelerated instances, explore the configs, and share your feedback to help shape the future of scalable LLM inference. 🔗 https://nvda.ws/4nuSADb

232

4 Comments

🚀 Behind the Magic: How NVIDIA Tensor and Ray Tracing Cores Power the Future of AI & Graphics We often hear about GPUs breaking benchmarks-but what’s really happening under the hood? 💡 In my latest technical deep dive (Part 3 of my GPU architecture series), I explore NVIDIA’s Tensor Cores and Ray Tracing (RT) Cores, two of the most revolutionary hardware components in modern GPUs. 🔬 What’s inside: ▪️ Tensor Cores: → Mixed-precision computing (FP16, TF32, FP8, INT8) for AI acceleration → The MMA (Matrix Multiply-Accumulate) engine’s pipeline in depth → Hopper’s Transformer Engine: Dynamic precision and sparsity exploitation for LLMs ▪️ Ray Tracing Cores: → BVH traversal hardware: How GPUs perform billions of ray-triangle tests/sec → Full ray shading pipeline: Closest hit, miss shaders, and secondary rays → AI-powered denoising using Tensor Cores for real-time photorealism ▪️ DLSS Synergy: → How Tensor + RT Cores work together to upscale & reconstruct high-fidelity images → A look at DLSS 3.0 frame interpolation and the future of AI-native graphics Why does this matter? These cores are pushing the boundaries of: ✅ AI model training & inference (GPT, vision transformers) ✅ Gaming realism (RTX & cinematic lighting) ✅ Scientific simulation & content creation I’ve packed this blog with diagrams, links, and architectural insights to help developers, engineers, and enthusiasts truly understand the silicon behind the magic. 👇 Full deep dive in comments! 🔗 #NVIDIA #GPUArchitecture #TensorCores #RayTracing #DLSS #DeepLearning #MachineLearning #CUDA #Hopper #H100 #Transformers #AI #GamingTech #ComputerVision #EdgeAI

217

35 Comments

🎉 Exciting News from the CUTLASS team at NVIDIA! 🎉 CUTLASS 3.8, delivering day 0 support for Blackwell, has just been released on GitHub 🚀 💡 Unlike our Hopper 3.0 release, where features rolled out over time, CUTLASS 3.8 arrives fully loaded with everything you need to make your workloads go brrr: - Out of the box support for optimized kernels like Flash Attention ⚡ - GEMMs supporting new block-scaling FP8, FP6, and FP4 data types 📊 - Various example kernels showing off CUTLASS+Blackwell features 💅 - Near peak performance of the hardware across the board for every data type 💪 - Full exposure of all building blocks in CUTLASS and CuTe for all new PTX instructions 🧰 We can’t wait to see what you all from the community will build! Share your feedback, ask questions, and let’s push the boundaries of GPU performance together! 💻🤝💻 👉 Check out the full release log now for a detailed list of all the new goodies! https://lnkd.in/gQWYs7fu #NVIDIA #CUDA 💚 PS: this post was written with the help of ChatGPT (4o)

324

2 Comments

The ExecuTorch 1.0 GA release is here, and it’s redefining what’s possible for AI at the edge. Built on a unified PyTorch workflow and optimized for Arm-based platforms, ExecuTorch 1.0 makes developing and deploying edge AI faster, with higher performance and less fragmentation across billions of edge devices. By using the same PyTorch models end-to-end across mobile, embedded, and edge platforms, workflows are streamlined and time-to-market is accelerated, alongside automatic performance optimizations through Arm KleidiAI, TOSA, and CMSIS-NN integrations. And the best part? It’s available right now, so you can bring edge AI to life everywhere, for everyone: https://okt.to/mS5u8R

75

2 Comments

This morning, we announced our Cube 3D foundational model (https://lnkd.in/gdxVTYkh) for generative AI with the key technical insight of cross attention between 3D and text/image tokens. This week, we’re releasing an open-source version of this model for use by anyone on or off the Roblox platform. In addition, we’re rolling out the first capability of this model with our new Mesh Generation API, which saves hours on each object created, giving creators back time to experiment with new ideas. In the long term, we plan to extend this technology to complex objects, even complete scenes.

144

2 Comments

Jensen Huang: I don’t need to change the world overnight “Having a simple idea that you can execute perfectly is sometimes better than having a grandiose idea that your company can’t execute on.” Jensen reflects on how he’s constantly bombarded with ideas and feature requests from customers and engineers at Nvidia. But he believes it’s prudent to keep things simple. As he puts it: “I don’t need to change the world overnight; I’m going to change the world over the next 50 years. I don’t need to build a killer product overnight; I just need to build a winning product. And the goal of winning is so that you can play again. It’s just like pinball… Most companies need to realize that it is a long road and you can’t build that perfect product [overnight].” He continues: “Once you do that and keep the project’s scope confined — you have a long-term vision but your product definition is rather simple — your people can now execute flawlessly on a simple plan. Then you come back and do it again.” Video source: Stanford Technology Ventures Program (STVP) (2003)

69

1 Comment

NVIDIA Research team just released a model to predict small molecule retrosynthesis path using generative AI technology. The full synthesis pathway is represented as a text sequence, and the model uses an Encoder-Decoder Transformer-based model architecture and Chain-of-Reaction (CoR), inspired by Chain-of-Thought (CoT) in LLMs, to generate the synthesis pathways for the target product molecule or its closest synthesizable analog(s). Read more from the blog or find the github code from there: https://lnkd.in/gUefdRFd

45

📣 Blackwell sets a new inference speed world record — A single NVIDIA DGX B200 server with eight #NVIDIABlackwell GPUs can generate over 1,000 tokens per second (TPS) per user on the Llama 4 Maverick model, the largest and most powerful model in the @AIatMeta Llama 4 family. ⚡⏱️ Additionally, a system with eight Blackwell GPUs can also deliver up to 72,000 tokens/second in a maximum throughput scenario. 🏆 Blackwell is the first platform to achieve this model performance, demonstrating how it delivers the best combination of throughput, output speed and accuracy for LLM token generation. 🏎️🏁 See our tech blog for details.

10

Zibra AI is looking for partners to build the next generation of cloud rendering infrastructure. Have you ever tried using cloud rendering infrastructure for VFX production? In theory, it offers compelling advantages: on-demand compute and storage scaling, centralized resource management, and high availability of processed data. In practice, however, it remains impractical for many high-end production pipelines. The main technical barrier? Volumetric VFX file sizes. No one wants to wait hours for a terabyte-sized VDB project to download from the cloud—or be hit with a massive data transfer bill at the end of the month. Even today, we still hear about teams physically transferring data between locations using hard drives. But what if your volumetric data took up only 5% of its original size—without compromising quality? With that kind of efficiency, you can simulate, render, and distribute content seamlessly across any location. This is now possible thanks to Zibra AI’s compression technology: ZibraVDB. It replaces OpenVDB as the new standard for volumetric visual effects, offering state-of-the-art compression rates, fast real-time decompression, and real-time rendering. ZibraVDB supports all the functionality required by render farms and integrates smoothly with major VFX production tools. Adding it to your pipeline is as simple as inserting a caching node in your Houdini project. If data storage and transfer have been holding you back from building a cloud rendering infrastructure, send me a direct message. With ZibraVDB, cloud rendering finally becomes practical.

27

As a prelude to the AI Infra Summit starting today at the Santa Clara Convention Center (Sept 9-12 2025), I published a paper highlighting the architecture challenges facing the industry. You can download the paper from the following link: https://lnkd.in/gz3-3cTn Artificial Intelligence is no longer a niche workload—it is the engine driving hyperscale data centers, autonomous vehicles, and mission-critical systems worldwide. Yet behind the promise lies a crisis in design: unpredictable workloads, skyrocketing silicon costs, interconnect bottlenecks, thermal hotspots, and the fragility of heterogeneous chiplet integration. This paper cuts through the noise, showing why traditional tools and benchmarks are failing us and why predictive, system-level exploration is now essential. From chiplets to cloud, it demonstrates how early modeling uncovers bottlenecks, reduces risk, and saves millions in time and cost, while enabling safer, faster, and more energy-efficient AI infrastructure. If you care about building AI systems that actually deliver in the real world—reliably, at scale, and ahead of competitors—this is the paper you cannot afford to ignore #AISystemArchitecture #ErchitectureExploration #PredictiveModelingForAISystems, #AIInfrastructureDesign #ChipletSystemDesign #SoCArchitectureExploration #DataCenterAIArchitecture #VisualSimArchitect #AIWorkloadModeling #ChipletIntegrationChallenges #SystemLevelSimulationTools #CacheHierarchyOptimization #MemorySystemArchitecture #InterconnectBottlenecksInAI #AIPowerAndThermalManagement #AIPerformanceVSReliabilityTradeOffs #EarlyStageSystemDesignExploration

49

4 Comments