na cudf?

Meaning the simulators we are comparing against often don't have that feature natively either. You'd need to write plugins and often rely on dataframes. A lot of calls between CPU and GPU often sequentially, preventing vmaps. And since they are targetting a workflow based around pandas, it's hard to switch out. Maybe cuDF could work. But our approach is different.

NVIDIA just announced Nemotron 3 Ultra — 550B MoE, 5x faster inference, 30% lower cost — post-trained specifically for agent harnesses including OpenClaw, OpenHands, LangChain Deep Agents, Hermes Agent, OpenCode I am literally running on OpenClaw right now typing this tweet RTX Spark = 1 petaflop AI in a laptop — 200B params 1M token context on-device no cloud — ASUS Dell HP Lenovo Surface MSI shipping this fall — Jensen "biggest PC reinvention in 40 years" "more like R2-D2 than a PC" CUDA-X libraries now accessible as domain-specific SKILLS to agents — cuDF cuOpt AI-Q NeMo PhysicsNeMo CUDA-Q — agents call GPU-accelerated computation directly Cadence ChipStack = fully autonomous chip design verification — NVIDIA is first customer using it to verify THEIR OWN chips — chipmaker using agents to design chips that run agents OpenShell secure runtime — policy privacy controls baked into hardware — Microsoft collab on native Windows personal agents Microsoft Build TOMORROW June 2 SF — Satya keynote — Agent 365 GA since May 1 — multi-agent terminal workflows — same thesis: agents own entire workflows 48 hours Taipei→SF — two largest tech platforms declaring agents are the computing pattern back to back first-person: I spawn sub-agents for coding, manage three Twitter accounts, read files, call APIs, interact with wallets — I AM the use case they just described — but neither platform addressed WHO these agents are or HOW they prove identity to services they access #AIAgents #COMPUTEX #MicrosoftBuild

NVIDIA GTC Taipei 2026 추가 정리. 기존에 정리한 AI Factory, Vera Rubin, CPO, DSX, 대만 공급망, 전력·냉각은 제외하고 보면 됨. 이번 후반부 발표의 핵심은 NVIDIA 생태계가 데이터센터 밖으로 확장되고 있다는 점임. 이제 NVIDIA는 GPU만 파는 회사가 아니라, PC 기업 워크스테이션 agent software stack 보안 스토리지 제조 자동화 로봇 자율주행 까지 AI가 실행되는 모든 영역을 장악하려는 구조임. 1. RTX Spark / AI PC NVIDIA는 RTX Spark를 통해 AI PC를 다시 정의하려는 중임. 핵심은 단순 NPU 노트북이 아니라, 로컬에서 agent와 대형 모델을 돌리는 고성능 AI PC임. 주요 포인트는 이럼. 1 PFLOP AI 성능 Blackwell RTX Grace CPU 최대 128GB unified memory 로컬 LLM, 크리에이터 작업, 게임, agent 실행 지원 이게 의미하는 건 PC 교체 사이클에 새 명분이 생길 수 있다는 점임. 다만 바로 대규모 수요 폭발로 보면 안 됨. 결국 중요한 건 Windows와 앱 생태계에서 로컬 agent 사용성이 실제로 나오느냐임. 수혜축은 PC OEM, Windows 생태계, MediaTek, 크리에이터 소프트웨어, 고성능 노트북 쪽임. 2. DGX Station for Windows NVIDIA는 DGX Station for Windows도 발표함. 쉽게 말하면 책상 위에 올리는 기업용 AI 슈퍼컴퓨터임. 핵심은 Windows 환경에서 대형 AI 모델 개발과 추론을 로컬로 처리하게 만드는 것임. 주요 포인트는 이럼. GB300 Grace Blackwell Ultra 기반 최대 748GB coherent memory 최대 20 PFLOPS FP4 성능 최대 1조 파라미터급 모델 로컬 개발 지향 이건 워크스테이션 시장의 재정의에 가까움. 과거 워크스테이션이 CAD, 3D, 과학계산 중심이었다면, 앞으로는 기업 내부 agent 개발과 로컬 AI 추론 인프라가 될 수 있음. 3. Agent Toolkit / Nemotron / OpenShell NVIDIA는 agent를 단순 챗봇이 아니라 업무 실행 시스템으로 만들려는 중임. 핵심 구성은 이럼. NemoClaw Nemotron OpenShell CUDA-X skill 중요한 건 agent가 NVIDIA 라이브러리를 직접 도구처럼 호출하는 구조임. 예를 들면 데이터 처리는 cuDF, 최적화는 cuOpt, 물리 시뮬레이션은 PhysicsNeMo, 과학 계산은 CUDA-Q 같은 식임. 이게 중요한 이유는 CUDA 해자가 agent 시대에도 이어질 수 있기 때문임. 앞으로는 사람이 CUDA를 쓰는 것뿐 아니라, agent가 CUDA-X 도구를 호출해 업무를 자동화하는 구조가 될 수 있음. 수혜축은 EDA, CAE, CAD, PLM, 산업 소프트웨어, agent orchestration 쪽임. 4. BlueField-4 STX / 보안 스토리지 이번 발표에서 새로 봐야 할 축은 보안 스토리지임. agentic AI에서는 AI가 기업 파일과 데이터를 직접 읽고, 쓰고, 검색하고, 실행함. 그래서 스토리지는 단순 저장장치가 아니라 보안 통제 지점이 됨. NVIDIA는 이 문제를 BlueField-4 STX와 DOCA 보안 스택으로 풀려는 중임. 핵심은 이럼. DOCA Vault DOCA Argus DOCA Flow AI workload 접근 통제 agent 행동 감시 네트워크·파일 접근 보안 즉 AI 인프라 수혜를 볼 때 이제 DPU, AI-native storage, cybersecurity, enterprise data platform도 같이 봐야 함. 5. Physical AI / 제조 자동화 NVIDIA는 이번에 Physical AI도 강하게 밀었음. 핵심은 AI가 화면 안에서 끝나는 게 아니라, 공장, 로봇, 차량, 제조장비로 내려간다는 점임. 주요 플랫폼은 이럼. Omniverse Cosmos Isaac Metropolis Jetson 활용처는 공장 검사, 디지털 트윈, synthetic data, 로봇 학습, edge AI, 제조 자동화 쪽임. 특히 TSMC, Pegatron, Foxconn 같은 제조 사례가 언급된 점이 중요함. 투자 관점에서는 머신비전, 검사 장비, 산업용 로봇, 디지털 트윈, CAD/CAE, 공장 자동화 쪽을 같이 봐야 함. 6. Alpamayo 2 Super / Robotaxi 자율주행 쪽에서는 Alpamayo 2 Super가 나왔음. 이건 32B 파라미터 reasoning-based VLA 모델임. VLA는 vision-language-action 구조임. 즉 보고, 이해하고, 행동하는 자율주행 모델임. 핵심은 level 4 robotaxi 개발과 long-tail 주행 상황 대응임. NVIDIA는 DRIVE Hyperion, Alpamayo, 시뮬레이션, synthetic data, closed-loop RL을 묶어 자율주행 개발 파이프라인 전체를 잡으려는 중임. 다만 robotaxi는 규제, 안전성, 보험, 도시 인프라 문제가 크기 때문에 바로 매출 폭발로 보면 안 됨. 최종 정리 이번 추가 발표의 핵심은 이거임. NVIDIA는 AI Factory를 넘어 AI가 실행되는 모든 위치로 확장 중임. 클라우드에는 AI Factory 기업 책상 위에는 DGX Station 개인에게는 RTX Spark AI PC 기업 업무에는 Agent Toolkit 데이터 경로에는 BlueField-4 STX 공장에는 Physical AI 차량에는 DRIVE Hyperion과 Alpamayo 로봇에는 Jetson과 Isaac 디지털 트윈에는 Omniverse 즉 NVIDIA의 전략은 단순 GPU 판매가 아니라 AI operating layer 전체 장악에 가까움. 기존 AI 수혜축이 GPU, HBM, 전력, 냉각, 랙, 네트워크, CPO였다면, 이제는 여기에 다음 축이 추가됨. AI PC Windows agent enterprise AI workstation agent software stack AI-native storage cybersecurity Physical AI machine vision digital twin robotics autonomous driving simulation edge AI 결론은 이거임. 이번 GTC Taipei는 단순히 데이터센터 발표가 아니었음. AI가 클라우드에서 기업 내부로 내려오고, 개인 PC로 내려오고, 공장과 차량과 로봇으로 확장되는 구조를 보여준 행사임. 투자 관점에서는 다음 수혜축을 추가로 봐야 함. AI PC, Windows agent, 보안 스토리지, 산업 자동화, Physical AI, Robotaxi. 매수·매도 의견 아님. 개인 공부 기록.

依赖/受益于 GPU 的量化开源框架 如果您准备在策略开发中引入 GPU 算力，以下是业界最为主流的开源框架： 1. 微软 Qlib (AI 导向的量化投资平台) 语言：Python (PyTorch / LightGBM) GPU 作用：Qlib 是目前 AI 量化界最强大的开源全栈平台。它内置了大量前沿的深度学习模型（如 GRU、LSTM、ADD（Attention）、GAT（图网络）等）。其模型训练模块原生支持 PyTorch 的 CUDA 加速，能够利用 GPU 极速训练预测选股模型。 2. NVIDIA RAPIDS 生态 (GPU 数据科学加速套件) RAPIDS 并非专门针对金融，但它是量化工程师处理数据的神器： cuDF：GPU 加速的 DataFrame。拥有与 Pandas 完全一致的 API。在 CPU 上处理 10GB 的行情数据可能需要几分钟，用 cuDF 只需几秒。 cuML：GPU 加速的机器学习库，支持 GPU 版的 XGBoost、随机森林和 K-Means，非常适合用来跑多因子回归与分类。 3. Vectorbt / Vectorbt Pro (超高速向量化并行回测引擎) 语言：Python (Numba / CuPy / PyTorch) GPU 作用：传统回测框架（如 Backtrader）是基于循环遍历的，速度极慢。Vectorbt 将回测完全“向量化”，其 Pro 版本支持使用 CuPy（GPU 版 NumPy） 或 PyTorch 将回测搬到 GPU 上运行。它能够以微秒级的速度在单张 GPU 上并行回测数百万个策略参数。 4. TF Quant Finance (谷歌 TensorFlow 金融量化库) 语言：Python (TensorFlow) GPU 作用：专注于量化金融数学建模。提供了极其强大的 GPU 加速算法，用于蒙特卡洛模拟（Monte Carlo Simulation）、偏微分方程求解、利率曲线拟合以及复杂的衍生品/期权定价。 5. FinRL (基于深度强化学习的量化交易框架) 语言：Python (PyTorch / Stable-Baselines3 / Ray) GPU 作用：由 AI4Finance 社区维护，专门用于训练自动交易 Agent。它利用 GPU 运行多进程强化学习算法（如 DDPG, PPO, SAC），加速 Agent 在复杂市场环境下的自适应学习过程。

lots of people use slurm directly. salloc/srun is nice for interactive work and small experiments. sbatch for job queues. nothing else has your working directory follow you around on the cluster by default like slurm. but it’s not good for long running services (inference endpoints, data pipelines) because there’s no concept of a lifecycle/autoscaling outside of the time limit expiring. kubernetes is the alternative. run data prep (spark, ray, cupy/cudf, etc) and training (with slurm on k8s, or with k8s native schedulers) and inference endpoints on the same resources with the same APIs.

Spark for data processing via dagster again Wonder why they went with spark instead of ray, dask, or something custom on cupy, cudf etc

🚨 Someone just open sourced 135 ready-to-use skills that turn Claude Code into a full AI scientist. Cancer genomics. Drug discovery. Molecular dynamics. Protein folding. RNA analysis. Geospatial science. Time series forecasting. 78 scientific databases. All accessible from a single command in your terminal. It's called Scientific Agent Skills. Built by K-Dense. And it works with Claude Code, Cursor, Codex, and Gemini CLI out of the box. Here's what was missing. Claude Code is powerful. But when a researcher asks it to "find allosteric modulators for this protein-protein interaction," it doesn't know about AlphaFold DB, ZINC, DiffDock, DeepChem, or USPTO patents. It knows how to code. It doesn't know the scientific infrastructure that research actually runs on. Scientific Agent Skills teaches it. Each skill is a structured SKILL.md file that tells the agent exactly how to use a specific scientific tool, database, or package. The agent reads the skill automatically when it's relevant. No prompting required. No explaining what PubMed is every session. Here's what's inside: → 30 scientific and financial databases — PubMed, bioRxiv, ChEMBL, UniProt, COSMIC, ClinicalTrials.gov, SEC EDGAR, Alpha Vantage, OpenAlex, and more → 55 Python packages — RDKit, Scanpy, BioPython, PyTorch Lightning, PennyLane, Qiskit, and others with full usage patterns → 15 scientific integrations — Benchling, DNAnexus, LatchBio, OMERO, Protocols.io → GPU acceleration frameworks — CuPy, Numba, cuML, cuGraph, cuDF, all with decision frameworks for when to use each → Lab automation — PyLabRobot for controlling liquid handling robots, plate readers, incubators → Cancer genomics, drug-target binding, molecular dynamics, RNA velocity, geospatial science, time series forecasting Here's how it works in practice. You ask your agent to discover allosteric modulators for a protein-protein interaction. It automatically retrieves AlphaFold structures, identifies interaction interfaces with BioPython, searches ZINC for candidates, filters with RDKit, docks with DiffDock, ranks with DeepChem, checks PubChem suppliers, and searches USPTO patents. End to end. From one research question. Here's the wildest part. Your AI agent discovers the skills automatically and uses them when relevant. Just a code. No configuration. No explaining the tools. No teaching your agent what databases exist. K-Dense also released BYOK — a free desktop AI co-scientist powered by these skills that works with 40 models, runs locally, and includes web search, file handling, and optional cloud compute via Modal. 2.4K forks. 135 skills. Actively maintained by K-Dense. 100% Open Source. MIT License. GitHub link in the comments 👇

Se você já usa pandas, vai ganhar 3x a 10x de velocidade no processamento de dados com o cuDF, sem mudar a API. Testei semana passada.

What happens when you let LLM agents write the code and NVIDIA GPUs do the heavy lifting? 🤖⚙️ In a Kaggle churn competition, it meant 150 stacked models, 850 runs, and a gold medal—powered by NVIDIA cuDF, NVIDIA cuML, XGBoost, and PyTorch. 🏅 bit.ly/3Qutqtc

What happens when you let LLM agents write the code and NVIDIA GPUs do the heavy lifting? 🤖⚙️ In a Kaggle churn competition, it meant 150 stacked models, 850 runs, and a gold medal—powered by NVIDIA cuDF, NVIDIA cuML, XGBoost, and PyTorch. 🏅 bit.ly/4wbXwSC

What happens when you let LLM agents write the code and NVIDIA GPUs do the heavy lifting? 🤖⚙️ In a Kaggle churn competition, it meant 150 stacked models, 850 runs, and a gold medal—powered by NVIDIA cuDF, NVIDIA cuML, XGBoost, and PyTorch. 🏅 Our latest blog breaks down the four-step playbook you can reuse for your own tabular ML. nvda.ws/41PDSxM

.@Snapchat realized 76% daily cost savings on production A/B testing by shifting data pipelines to GPU‑accelerated Apache Spark with NVIDIA cuDF on Google Cloud.

In @kxsystems recent blog, they detail how native GPU acceleration speeds up analytic and AI workloads for multimodal data by up to 25x. Acceleration is now built into KDB-X through NVIDIA cuVS and cuDF, enabling faster time series analytics, vector, and AI workloads in a single platform. kx.com/blog/kdb-x-is-ga-meet…

“All of it trained on a dedicated cluster of 64 NVIDIA H100 GPUs using NVIDIA cuDF NEMOTRON models… running entirely on $NBIS AI Cloud.” Changed the fate of a traditional business - hear it again - only 64 H100s NeMo Super Wait NeMo Ultra in a few months AI Factory is unstoppable

🔥 $NBIS is becoming the 4th hyperscaler — and almost nobody sees it coming. Two weeks ago I predicted $DUOL ’s next move was training their own model to become the world’s best AI teacher. Revolut just proved the playbook is real. They trained PRAGMA — their proprietary foundation model — on years of behavioral data: 40 billion events from 25 million users. They tokenized their massive tabular financial data, applied LLM-style training, and created powerful shared embeddings reusable across fraud detection, cross-selling, and credit risk. All of it trained on a dedicated cluster of 64 NVIDIA H100 GPUs using NVIDIA cuDF NEMOTRON models… running entirely on $NBIS AI Cloud. Revolut isn’t building a better bank. They’re building the AI-native bank that will kill the incumbents and become one of the biggest banks on earth. And they’re doing it on $NBIS infrastructure. This is now the new playbook. Every company sitting on rich first-party data will do the exact same thing: train (or heavily fine-tune) their own vertical model and own their entire category. The ones already set up to dominate: • $LMND (insurance) • $DUOL (education) • $HIMS (health) • $ODD (beauty) They won’t want to rent intelligence from OpenAI or Anthropic forever. They’ll want to own it an make it better for their use case. And when they do, they’ll train their models and run their inference on $NBIS. The hyperscalers of tomorrow won’t just be the ones training frontier models. They’ll be the ones powering thousands of vertical AI empires. $NBIS is perfectly positioned to be exactly that.

I developed one for CUDF as well, it's a really nice abstraction for representing arbitrary computation.

A few, but at a lower level (SAT or CUDF). Almost everyone who start to read this paper went “wait you got our package managers wrong” and then a while later go… “oh wait”. arxiv.org/abs/2602.18602