About
Software engineer at AMD interested in compilers, AI, and Machine Learning.
Activity
531 followers
-
Ian Wood reposted thisIan Wood reposted thisWe have day 0 support for the new Radeon RX 9070 XT and RX 9070 GPUs in IREE / MLIR, with both ROCm and Vulkan. Combined with SHARK AI, we got Stable Diffusion XL fp8 running on these cards: https://lnkd.in/g5QX8Z-k Demo: https://lnkd.in/gDaEnBPm
-
Ian Wood reposted thisIan Wood reposted thisThis summer, I interned at Deloitte in Seattle as a tax intern in business tax services. I am truly grateful to have had this opportunity to work with Deloitte professionals who supported me in gaining knowledge and experience in a Big 4 Accounting Firm. I am super excited to announce that I have accepted a full-time position as a tax consultant! A special thank you to Isabella Wagner, Grace Hodge, MST, CPA, and Jenna Rogers, CPA for giving me hands on experience with engagements throughout my internship.
-
Ian Wood liked thisIan Wood liked thisI am proud to announce that I have passed all 4 CPA exams on my first try! I am super thankful to everyone who supported me throughout this journey at EY and the University of Arizona, I could have not done it without you. I am excited to continue the CPA certification process and excited for what the future holds!
-
Ian Wood liked thisIan Wood liked thisMay 15th, 2026 was my last day at AMD — an incredibly hard decision to step away from colleagues I'd worked with for over 7 years and a project I'd been part of since its early days. I have no words for the gratitude I feel toward Stella Laurenzo, my manager across multiple companies and, more than that, a mentor who has had a tremendous personal impact on me. Thank you, Stella, for your technical guidance and for building a place where people could thrive — everyone knew you had their back. To the many phenomenal engineers I had the opportunity to work with: I can't name you all, but I'm grateful for everything you taught me. And thanks to Anush E. for pulling us along the remarkable journey from Google to Nod to AMD. Today (June 15th, 2026) I started at NVIDIA — a full-circle moment, back on Vinod Grover's team almost exactly 12 years since I first joined. Vinod always has a ton of great ideas, and with the new age of computing upon us, I'm excited to explore how AI tools can build better compilers, generate better kernels, and unlock what comes next. Thanks Vinod and Luis Ceze for the opportunity. A lot has changed in 7 years, but it was good to be back at the familiar Redmond office today (same espresso machine, I think). If you're there too, I'd love to grab a coffee and hear what you're working on.
-
Ian Wood liked thisIan Wood liked thisI’m so excited to share that I was selected as an Event Planner on the executive board for Psi Chi, the International Honor Society in Psychology, at the University of Arizona! As a Psychology major pursuing a future career in child psychology, I’m excited for the opportunity to help create events that bring students together, encourage involvement, and support professional and academic growth within the psychology community. I’m especially looking forward to strengthening my leadership, communication, organization, and collaboration skills while working with an amazing team to plan engaging experiences for members throughout the year. I am so greatful for this opportunity and so excited for next semester!
-
Ian Wood liked thisIan Wood liked thisI finally gave in and spent the several hours to upgrade my project setup, after spending the last month watching some of the folks I work with running setups where they can just be running a dozen different little pods of workers off of the same git repos. Confession: I don't usually "set up a dev environment" -- typically preferring to keep things in a mostly default state so I can machine hop. So this is more time than I've spent on setup in, maybe, decades. And it was somewhat annoying. The main pieces are: * beads_rust (https://lnkd.in/gFARsKV5) : A simplification of yegge's original beads * beads_viewer: https://lnkd.in/gJCSRWzi * wezterm: Lua scriptable terminal that I can just claude my way to efficiency with I have one main workspace with research, docs, top level skills, and all the normal things. Then I nest reference projects as submodules (pretty much all of the standard AI things you want an agent to always have available for reference). And I keep the projects I'm contributing to under a gitignore'd projects/ dir in the super-workspace. There's a script in there to worktree-fork the whole ball of it (which is surprisingly hard with submodules and loose bits), set up new direnv, initialize a worktree specific beads database, etc. Then there is a hotkey to pop a new wezterm 4-pane grid with two claudes, beads viewer and a terminal for a worktree. The hardest part was getting the claude sandbox working and navigating around the many bugs (pro tip: stay on v2.1.25 -- .29 regressed many things about the sandbox). I didn't have time to import my co-worker's cross-review skill, which has been working out pretty well (it does an in-depth cross review task based loop between Claude, Gemini, and Codex and has been producing pretty good results on things he's sent me). Sandboxing, isolated worktrees, and beads are pretty much the backbone of running pods of agent-loops semi-autonomously, and I'm looking forward to going into that mode and being less single threaded. If my co-worker's experience translates, we think this kind of rig is really all you need in terms of scaling yourself.
-
Ian Wood liked thisIan Wood liked thisSome of you know me as a pretty good engineer. Like a lot of us, I've been figuring out my relationship with AI. In my personal life, I'm basically always using it for all kinds of things. And I use it a lot transactionally for the job. But one advantage of being hands on and managing a large team is that you get to see a larger spectrum. And I've come to believe that just as there has always been stratification between the skills you learn at each step of the software engineering journey, use of AI follows similar stair steps. Used well, it is an accelerant at all levels, but it is the senior engineers using it in very advanced ways that is catching my interest. I'd go so far as to say that if you're a senior engineer doing meaty platform work, and you're not using a thousand dollars equivalent of GPU capacity a month to scale yourself, you're probably leaving a lot of effectiveness on the table. My senior engineers who have really embraced AI with advanced workflows were typically already highly effective. But now I'm watching them running entire virtual teams and having the bandwidth to also deliver the stuff that usually falls through the cracks (docs, tests, proven correctness, linters, CI). Another thing I've noticed is that so much of our software is just fluff and layers that we created so that humans can form tribes and argue about limited scopes -- and we're drowning in most of it. But AI agents have no problem cutting down through all of that and getting to the meat of the matter. I think this is going to have a long term, positive effect, letting us simplify our software and focus on the basics at the bottom again vs all of the affordances and preferences that were put in place for people. We are living in very interesting times. AI magnifies engineering skills at all levels when used extensively. It is also a bit terrifying when it magnifies the slop instead -- this is where I'm finding that those who had good fundamentals already are evolving the fastest. In key ways, it puts more pressure on the humans in order to get high quality results.
-
Ian Wood reacted on thisIan Wood reacted on thisI’m happy to share that I’ve obtained a new certification: Google Advanced Data Analytics from Google!
Experience
Education
-
Northwestern University
-
-
Relevant Coursework: Operating Systems, Compiler Construction, Kernel & Low-Level Development, Computer Networking, System Security, Artificial Intelligence (AI) Programming, and Machine Learning (ML).
Projects
-
Compiler with LLVM Backend
- Extended my compiler developed in my Compiler Construction class to support additional language features (e.g. struct, classes, references, etc.)
- Utilized LLVM as a backend for target code generation as well as optimization passes.
- Implemented additional modern compiler features, such as type checking, to enhance the production-readiness of the compiler -
Port Filesystem to Nautilus Kernel
-
- Ported FatFs, a filesystem designed for embedded systems, to Nautilus which lacked robust filesystem support.
- Implemented Nautilus’ filesystem interface, managed state, and wrote disk I/O functions that connect FatFs to Nautilus’ block device drivers.
- Tested my implementation by validating FatFs’ ability to create new files, read existing files, create new directories, etc. -
Compiler Project
-
- Developed a compiler using C++ to compile a C-like language into an executable x86-64 binary.
- Implemented basic compiler techniques such as register allocation, variable spilling, instruction selection (tree tiling), and linearization of the control flow graph.
- To improve performance some optimizations were included such as dead-code elimination, constant folding, and coalescing/freezing of variables. -
AI Chess Engine and Single-Player Game
-
- Developed logic for the game using Python and implemented the GUI using pygame.
- Wrote a basic alpha-beta pruning AI algorithm that allows the user to play against the computer.
- Converted the project into C++ using a Northwestern game API and added improvements that better determined the score of the game state. These improvements led to a more advanced opponent which could analyze the board at a greater depth.
Explore more posts
-
Sergio Rivera Bustamante
AMD is hiring Rust engineers to develop compilers, firmware, and runtime environments for its GPUs. Not as an experiment. As a core technical direction. NVIDIA is exploring Rust in CUDA Oxide. The community is building Rust GPU and Rust CUDA. Vectorware is betting on native shaders in Rust. The industry is not "trying a new language". It is responding to a structural constraint that the C/C++ ecosystem can no longer solve at scale. The reason is simple: A CPU has a few powerful cores. A memory error is serious, but contained. A GPU has thousands of simple cores executing in parallel. A memory error is not a bug. It is a silent corruption that propagates across thousands of simultaneous threads. In that environment, memory safety is not a style preference. It is an operational requirement. And Rust does not provide that through a team convention or a linter. It provides it at compile time. The borrow checker acts as the first deterministic gatekeeper before the code touches hardware. That is why AMD is not looking for "developers who know Rust". It is looking for engineers who understand that the correctness of GPU firmware cannot be delegated to human code review. It must be guaranteed structurally. The GPU tooling ecosystem is still immature and fragmented. That is not a hardware weakness. It is an opportunity for whoever designs the software layers with the right guarantees from day one. The question is no longer "Rust on GPUs?". The question is: who will design the validation and memory safety layers for the parallel computing of the next decade? Are you observing this convergence, or are you already building on it? Let me know in the comments. 👇 #RustLang #GPU #SystemsEngineering #AMD #MemorySafety #SoftwareArchitecture
-
Drew Steinberg
We recently introduced #NVIDIA NV-Tesseract, a family of models designed to unify #anomalydetection, classification, and forecasting within a single framework. NVIDIA NV-Tesseract-AD builds on this foundation, introducing diffusion modeling stabilized through curriculum learning and paired with adaptive thresholding methods—addressing noisy, high-dimensional signals that drift over time. Learn more. #DataCenter #Industrial #AI
-
Yuvraj S.
Wrote up a quick blog post exploring OS scheduling and thread mapping using some C++ benchmarks. It's a fun look at how physical cores vs. logical threads impact actual throughput vs. just time-slicing. Big thanks to Amrutansh and Priyanshu Sir for the nudge to document this! Always open to feedback or corrections from the systems folks on here! 🔗 https://lnkd.in/ghHJhW-W
10 Comments -
Vijay Shekhawat
This is wild. Anthropic let multiple AI agents work in parallel to build a C compiler from scratch - no human compiler engineers involved. In about two weeks, they produced ~100k lines of Rust code and a compiler that can compile the Linux kernel and even Doom. Cost: around $20k in API credits. What’s exciting isn’t just that it worked. It’s the shift from “AI helps write code” to “AI systems can autonomously build complex software by collaborating.” Feels like the early days of distributed systems - but for intelligence. Definitely worth a read: https://lnkd.in/eu4SJVSw #AI #Engineering #AutonomousAgents #SoftwareDevelopment
2 Comments -
Rob Kemp
NVIDIA DGX Spark keeps getting faster. ✨⚡ New software, models, and open-source optimizations deliver up to 2.6× higher performance with NVFP4, run 235B-parameter models locally, and allow multiple workloads to run simultaneously on the desktop. Read the technical blog ➡️ https://bit.ly/4jqPFdG
-
Joe Pierce
🎯 Similar LLM accuracy, half the memory on NVFP4. VentureBeat highlights how NVIDIA researchers trained a 12-billion parameter Hybrid Mamba-Transformer model on 10 trillion tokens with training loss within 1% of FP8. Results achieved utilized NVFP4 features: ⚙️ Hadamard transforms to handle block-level outliers ⚙️ 2-D quantization aligned across forward/backward passes ⚙️ Stochastic rounding with selective high-precision layers NVFP4 is in active development with AI leaders AWS, Cohere, Google Cloud, OpenAI, and Perplexity. 🔗 Read the full article to learn more.
1 Comment -
Francesco Carobolante
In my last post, I argued that realized memory bandwidth matters more than peak specs for AI inference. Positron AI’s "Asimov" ASIC is the architecture that implements this thinking. The thesis is simple: If decode-phase inference is memory-bound at <1 GOPS/GBPS, design the chip around the memory subsystem, not peak FLOPS. It generates some penalty on first token arrival, but not on the overall response time. The Headline: 2.3 TB of memory per Chip. Nvidia’s upcoming Rubin tops out at 288 GB of HBM4. That’s an 8x capacity difference, allowing massive models to reside on a single chip. While others attack the "Memory Wall" differently—Groq/Cerebras with ultra-fast but low-capacity SRAM, and SambaNova with a tiered HBM+DDR5 cache—Positron is betting on Memory Disaggregation. The "Gearbox": Credo’s Weaver Chiplet. The secret isn't HBM; it’s commodity LPDDR5x. To overcome LPDDR’s lower per-channel bandwidth, Positron and Credo developed the Weaver chiplet. The Technology: Using 112G SerDes at ~1 pJ/bit, it pushes memory 250mm away from the die. The Result: No expensive CoWoS interposers. No HBM allocation wars. Just standard PCBs and DRAM from any major vendor. This is a supply chain argument as important as the architectural one: by sidestepping TSMC’s CoWoS and HBM constraints, Positron ensures everything - from silicon to assembly - can be sourced and built reliably. There's also a forward-compatibility benefit: when LPDDR6 arrives, they can swap the Weaver chiplets while the Asimov die doesn't change. Efficient Compute & Cooling The compute side is just as agile. The TransWarp systolic array dynamically reconfigures between 512x128 (Prefill GEMMs) and 128x512 (Decode GEMVs). Thermal Envelope: 400W air-cooled. Infrastructure: No liquid cooling or 1kW+ TDP required. I was initially skeptical of LPDDR5x for AI. The industry has conditioned us to believe HBM is the only path. But looking at realized utilization and supply chain logic, it's a hard argument to ignore. There is one critical challenge this architecture creates that hasn't been discussed yet. I’ll let you guess it before my next post. #AI #Semiconductors #Inference #ChipDesign #ASIC #Positron #HBM #LPDDR5x #Chiplets #HeterogeneousIntegration #SupplyChain #DataCenter
5 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content