Bodun Hu

GPUs still optimize compute and communication separately. That leaves performance on the table. CUCo is our agentic, training-free framework that rewrites CUDA+NCCL programs to jointly co-design compute + communication on device, unlocking overlap, fusion, and pipelining. Up to 1.57× end-to-end latency reduction on MoE, KV-cache transfers, FlashAttention+context parallelism, and GEMM+AllGather. 📄 arxiv.org/abs/2603.02376 🌐 ut-infraai.github.io/cuco/ 💻 github.com/UT-InfraAI/cuco #CUDA #MLSystems #GPU #HPCL @BodunHu @utnslab

For decades, OSes have relied on hand-tuned heuristics. What if they could write their own? 🔥 Meet PolicySmith, our LLM-driven framework that generates high-performing, interpretable system code. 📄arxiv.org/abs/2510.08803 at #HotNets25 #LDOS #MLforSystems #CodeGeneration

Check out our hot take on policies in computer systems, especially machine learned ones. As we dig deeper into LDOS implementation, we're seeing a foundational role for learned heuristic code driving resource management, allocation, and planning policies. In the limit, code learned on the fly can tailor an entire OS to optimally serve currently running workloads while meeting complex system objectives. This work will appear at HotNets 2025. Led by the brilliant Rohit Dwivedula and @divytweet with collaborators @swarat and @daehyeok_kim .

Thrilled to hear USENIX ATC will now officially continue under ACM SIGOPS sponsorship! Same name, agenda, and abbreviation. This is a testament to the community's strength. 👏🏽

KV caching in LLMs, clearly explained (with visuals):

Introducing ChartMuseum🖼️, testing visual reasoning with diverse real-world charts! ✍🏻Entirely human-written questions by 13 CS researchers 👀Emphasis on visual reasoning – hard to be verbalized via text CoTs 📉Humans reach 93% but 63% from Gemini-2.5-Pro & 38% from Qwen2.5-72B

USENIX ATC is no more usenix.org/blog/usenix-at…

News: the latest ChatGPT-4o (2025-03-26) jumps to #2 on Arena, surpassing GPT-4.5! Highlights - Significant improvement over the January version (+30 pts, #5->#2) - Tied #1 in Coding, Hard Prompts. Top-2 across ALL categories - Matching or surpassing GPT-4.5 with 10x cheaper price. Congrats @OpenAI for the new milestone! View below for more insights ⬇️

BREAKING: Gemini 2.5 Pro is now #1 on the Arena leaderboard - the largest score jump ever (+40 pts vs Grok-3/GPT-4.5)! 🏆 Tested under codename "nebula"🌌, Gemini 2.5 Pro ranked #1🥇 across ALL categories and UNIQUELY #1 in Math, Creative Writing, Instruction Following, Longer Query, and Multi-Turn! Massive congrats to @GoogleDeepMind for this incredible Arena milestone! 🙌 More highlights in thread👇

Applications opened today for the LDOS PhD Research School: ldos.utexas.edu/events/researc…. Please apply or share with your friends and/or students!

DeepSeek R1 is 100% Opensource and 96.4% cheaper than OpenAI o1 while delivering similar performance. OpenAI o1: $60.00 per 1M output tokens DeepSeek R1: $2.19 per 1M output tokens People with $200 ChatGPT subscription, let that sink in.

Nobel Prize in Physics is awarded to AI this year. Incredible!

@utnslab at our recent end-of-semester social.

Super happy about this!! ❤️ But to be honest, with exceptional faculty colleagues and students, @UTCompSci feels like -- and deserves to be -- a top 5 department 🙂

@utnslab hike at Barton Creek with @VenkatArun95 @daehyeok_kim and our amazing students - @jitaotweet @Jane_JiayiChen @BodunHu @divytweet @brian0296 @RohitDwivedula Jiaxin, Geon-Woo, Jeongyoon, Yeonju, Haoyu, Darius, and Donghyun.

@aureliomjr @TravelGov I can’t seem to upload any photos

@TravelGov Greetings. Is the ceac.state.gov down? I need to retrieve a DS-160 confirmation, but apparently, the website is down for a few hours.

@TravelGov Well I can’t upload any photos because the website just says “error: there was an error processing the request”

@DrupalThoughts @racknerd @racknerd their Dallas data center is also constantly experiencing outage. I’ve never experienced so many outages with Vultr in 2 years compared to racknerd in one month

Generative LLMs are slow and expensive to serve. Their much smaller, distilled versions are faster and cheaper but achieve suboptimal generative performance. We show it is possible to achieve the best of both worlds. Code: github.com/flexflow/FlexF… Paper: cs.cmu.edu/~zhihaoj2/pape…

A series of AI announcements by Meta: - MTIA v1: an AI chip for fast inference: ai.facebook.com/blog/meta-trai… - RSC: 5 exaflops, 16,000 GPU Research Super Cluster for AI research: ai.facebook.com/blog/supercomp… - AI-focused data centers: ai.facebook.com/blog/meta-ai-i…