Over the past few weeks, I have added several new pages to Xe-Diagrams, with more details about GPU systems and how different parts of the stack work together.
Main page: https://xwang186.com/xe-diagram/
Before getting into the updates, I have to give some credit to Codex with GPT-5.6 Sol Ultra.
I have been working on this project with Codex for quite a while. For this round of updates, I asked the new model to work on several pages I wanted to add, and I was surprised by how well it did.
One thing that surprised me even more was that it figured out a much better way for me to update the website instead of manually uploading everything each time. I had been using the “High” model for similar tasks before, and its solution was basically to take over the browser and make the necessary clicks. The stronger model somehow dug deeper into the same public interfaces and found a completely different workflow.
I find that pretty interesting. Same tools and interfaces, but a better model can sometimes discover a much better solution rather than simply execute the obvious one.
So… now I know how much I need the “Ultra” model lol
New major pages
DX12 / Windows System View
https://xwang186.com/xe-diagram/?view=dx12-windows
This page connects the DX12 software stack with the GPU system underneath it, showing how the API, runtime, driver, command submission, and hardware pieces fit together.
Matrix / XMX Execution and Metrics
https://xwang186.com/xe-diagram/?view=matrix-xmx-execution
This page looks at matrix execution from a workload and performance perspective, including how useful metrics can be derived from the underlying execution behavior.
Shader Execution
https://xwang186.com/xe-diagram/?view=shader-execution
This page focuses on how shader workloads move through the GPU execution system and how the different pieces relate to each other.
I have also been expanding the DX12/Vulkan-related material. These pages are meant to be more than architecture diagrams: they are examples of how a system-level architecture map can be used to understand real workloads, software stacks, execution behavior, and performance.
Xe-Diagrams now has five major groups:
- Main Architecture
- Graphics API
- Workload Groups
- Commands & Submissions
- Observability
Each group now contains multiple interactive views and additional material to explore.
I am also planning a separate page about how to use Xe-Diagrams as a learner—not just as a collection of diagrams, but as a way to build a stronger mental model and knowledge base around Intel GPU systems.
Also thinking of whether I should publish my profiling tool. To play safe I have to remove all the decision makings and summaries, downgrading it to an “MDAPI ui wrapper”. I hope it’s still fine to play with, but very likely people can do better with less limitations(unlike me as current employee) and more powerful tokens ;P so..
Leave a Reply