2026 — NOW
Cryogenic GPU Systems
Building a reproducible chain from evidence and parameter synthesis to GPU tracing, simulation, and design-space exploration at 77 K.
YUNXIANG HE / HYPOXANTHINE
Reframing GPU design at 77 K.
Researcher and systems builder at ShanghaiTech VSP Lab. I study cryogenic GPU architecture, build agent tooling, and maintain a long-running archive of technical and literary writing.
77 K GPU ARCHITECTURE ACTIVE / 2026
Explore the field 01 / Current field
Evidence to architecture.
My current work connects cryogenic evidence, parameter synthesis, GPU workload characterization, and simulation into one reproducible design-space exploration chain.
- INPUT
- Evidence at 77 K
- MODEL
- Parameter synthesis
- SYSTEM
- GPU simulation
- OUTPUT
- Design space
02 / Selected systems
Tools that keep becoming infrastructure.
01
Current research
Cryo-GPU
Evidence-guided design-space exploration for GPU architectures at 77 K, connecting device shifts to system allocation through reproducible modeling and simulation. 02
Agent terminal
VSPi
A Chinese terminal UI for Pi coding agents, unifying sessions, model switching, and read-only workflow state in a focused TUI.
03
Recoverable workflow
Hypo-Workflow
A file-first workflow for coding agents, making long tasks discussable, recoverable, and reviewable through Cycles, Plans, Progress, and Memory.
04
AI-first documents
HypoDoc
A structured Markdown specification with parsers, Web/LaTeX renderers, and Desktop and VS Code authoring surfaces.
05
Chinese writing workflow
Hypo-Writer
A local-first Chinese writing system spanning sources, fact cards, outlines, review, and multi-channel publication packages.03 / Research trajectory
From application accelerators to cryogenic systems.
2024 — 2026
GPU architecture and systems
Expanding from individual neural-rendering accelerators toward workloads, memory hierarchies, scheduling, and system modeling.
2022 — 2025
Neural-rendering acceleration
Research spanning algorithm analysis and hardware design for ray representation, sampling, data reuse, and conflict-free encoding.
04 / WeChat & writing
Science writing as a second system.
Latest AI daily issue / 2026-08-23
08·23 - AI 午报：GLM 复活 AMD 老卡推理支持，energygraph 补上三厂 GPU 功耗监控
5 NEWS ITEMS AI 模型观察
王侯将相，宁有种乎？Kimi K3 评测 | AI 模型观察
月之暗面在 7 月 16 日正式发布了 Kimi K3：2.8 万亿参数、原生多模态、100 万 token 上下文，并且承诺完整权重在 7 月 27 日前开放。官方博客里有一句自我评价写得很克制：K3 的整体表现仍落后于最强的闭源模型 Claude Fable 5 和 GPT 5.6 Sol。
AI 焦虑门诊AI 时代，受益最大的，恰恰是「不会编程」的人 | AI 焦虑门诊（1）
:::intro AI 时代受益最大的人，恰恰是原本不懂编程的人。他们之前本就有东西可做、想做，但是缺乏实现这些想法的工具。
AI 模型观察我做了一套自己的 AI 体感评测 | AI 模型观察
:::intro 不需要编程的轻度用户，到底怎么选择适合自己的 AI 模型？那些越来越高的编程榜单、Agent 榜单，真的能反映非专业用户的使用体验吗？
05 / Beyond systems
Literature, teaching, and the games I keep measuring.
Classical poetry, creative writing, course materials, table tennis, and badminton sit beside the engineering work. The archive keeps these threads visible without forcing them into a CV.
POETRY TEACHING WRITING SYSTEMS
About Yunxiang