Jeff (Yuhao) Tang

⚡️ Frontier open models run faster on NVIDIA. New benchmarks show the NVIDIA Blackwell platform maximizes inference ROI by optimizing user interactivity and throughput per GPU. ⚙️ OpenAI gpt-oss-120b reaches over 60,000 tokens per second (TPS) per GPU. ⚙️ AI at Meta Llama-3.3-70b tops 15,000 TPS per GPU. ⚙️ DeepSeek AI DeepSeek-R1 achieves over 10,000 TPS per GPU. Our continuous software improvements and extreme hardware-software co-design enable real-world AI inference to balance cost, efficiency, throughput, and responsiveness. 📈 Explore the interactive performance charts: https://bit.ly/4ons9zL

2

📝 Learn how to use #GenAI-Perf to benchmark the Meta Llama 3 model when deployed with NVIDIA NIM, including popular metrics and parameters. Technical deep dive with a step-by-step guide ➡️ https://bit.ly/42Sm3zt This is part 2 in our #LLM benchmarking series.

2

Object Detection in an Urban Environment using Waymo Open Dataset on AWS 【Core Architecture & AWS Implementation】 Developed an end-to-end detection pipeline on AWS, leveraging Amazon S3 for large-scale Waymo Open Dataset management and Amazon SageMaker for distributed training. By optimizing pipeline.config across multiple backbones and hyperparameters, I successfully streamlined the process from data ingestion to SavedModel exportation. 【Algorithm Performance Evaluation & Analysis】 I conducted a comparative analysis of three detection architectures at the same training stage: ResNet50 V1 FPN: Demonstrated the most stable performance with an mAP of 0.1375, showcasing superior feature extraction and rapid convergence. EfficientDet-D1: Utilizing a BiFPN architecture, it achieved the lowest Total Loss (0.3256), indicating high precision in object localization. MobileNetV2 FPN-lite: Designed for lightweight deployment, providing a critical baseline for inference latency in edge computing. 【Cross-Domain Algorithmic Insights: From Biosignals to Computer Vision】 Coming from a background in biosignal processing, I gained deep insights from Data Augmentation techniques in CV. While signal processing often relies on heavy filtering to extract "clean" data—which can lead to overfitting—CV uses random scaling and transformations to enhance model robustness. I plan to apply this "data variance" logic to signal algorithms (e.g., temporal shifting or synthetic noise injection) to build more generalized models. 【Future Work】 To enhance accuracy and real-time performance, I have mapped out two paths: Accuracy Tuning: Improving small-object recall by refining multiscale_anchor_generator density and scales, while unfreezing Batch Normalization for fine-tuning with Cosine Learning Rate Decay. Model Compression: Implementing Pruning and Quantization to significantly boost inference speed and energy efficiency on automotive hardware without sacrificing detection precision. #waymo #MachineLearning #ComputerVision #AWS

12

Etched just raised ~$500M in a major funding round (valued at ~$5B) to accelerate its custom AI silicon — a solid alternative to Nvidia-centric stacks. https://lnkd.in/dttxqJPJ

6

“I think the next 10 years will be a cyber resiliency marathon for all of us. I’m not talking just about cybersecurity. Cyber resiliency is much more than that.”

1

🚀 Excited to announce the launch of Shamsi Calendar API! 🎉 After weeks of development, I'm proud to release a comprehensive RESTful API for the Afghan Shamsi (Solar Hijri) calendar - now live at https://lnkd.in/g7TNWXys Why I built this: As Afghan developers, we've long struggled with the lack of reliable tools for our calendar system. This API changes that. 🌟 Key Features: ✅ Accurate Shamsi ↔ Gregorian date conversion ✅ Full calendar views (day, week, month, year) ✅ Complete holiday management system ✅ Trilingual support: Dari, Pashto & English ✅ Interactive Swagger documentation ✅ 151 tests with 85.6% code coverage ✅ Free and open-source 🛠️ Built with: TypeScript, Express.js, tested with Jest, and deployed with enterprise-grade security (rate limiting, CORS, Helmet) 📊 What you can do: Build Afghan calendar apps without reinventing the wheel Integrate Shamsi dates into your existing applications Access accurate holiday information Use it in any language: TypeScript, Python, Java, etc. 🎯 Perfect for: Afghan tech startups, government systems, educational platforms, mobile apps, and anyone serving the Afghan community worldwide. Try it out at the interactive documentation and let me know what you think! Contributions and feedback are welcome. #OpenSource #API #AfghanTech #ShamsiCalendar #WebDevelopment #TypeScript #SoftwareEngineering #Afghanistan

22

1 Comment

TL;DR: Cascade RL and even Cascade smaller LR SFT for mitigating catastrophic forgetting. Nvidia recently showed that Cascade RL can mitigate catastrophic forgetting. Really excellent and exciting work! (https://lnkd.in/gDwKxG5f) We also observed similar effects in healthcare domains (clinical trial tasks) in our recent paper: Developing Large Language Models for Clinical Research Using One Million Clinical Trials https://lnkd.in/gNCf9M4i. Beyond Cascade RL, we also find that Cascade SFT with a smaller learning rate can significantly suppress catastrophic forgetting, even across highly heterogeneous clinical trial tasks. This is based on the findings in my previous paper below: SFT Doesn't Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMs https://lnkd.in/gqWJdvmX.

58

2 Comments