
Engineer and dad. Curious about Geopolitics, healthcare policy. 🌍 Working @dynamia_ai on HAMi
Joined December 2024
- Tweets52
- Following222
- Followers26
- Likes212
STRONG integrity working again on Android 17 (LineageOS 24, OnePlus 11).
TEESimulator-RS was reporting Android 16 version fields in attestation. My patch adds the KeyMint v5 / SDK 37 fixes.
github.com/fishing-agents/TE…
57
HAMi and Dynamia folks will be at KCD Hangzhou on Nov 28. If you are running GPUs on Kubernetes, go say hi.
1
22
JuiceFS is everywhere in AI infra once you look: BentoML cut LLM load time from 20+ min to minutes, Zhihu runs 3.5 PB of training data on it, Xiaomi 5B+ files. Agent workspaces look like a natural next fit
JuiceFS 应该是这一轮 Agent Infra 里面隐藏的英雄,很多 agent cloud 都基于 JuiceFS 做文件系统,这也是国人做的原创并且比较重要的开源项目了 @suavesu
1
2
38
Fishmantic Coder retweeted
ContainerDays Singapore first day, gave a @HAMiProject workshop and meet the community guys workshop.project-hami.io/lab… @SaiyamPathak @fishmanticoder
Fishmantic Coder retweeted
At #ContainerDays Singapore 2026, Dynamia's @fishmanticoder Reza Jelveh presented “10X Your Neocloud’s GPU Utilization with HAMi.”
The talk covered fractional GPU sharing, heterogeneous scheduling, resource isolation, and DRA.
project-hami.io/landing/cont…
#HAMi #AIInfrastructure
Hey guys I created a graphing engine based on graphviz that makes AI generated graphs more manageable and editable www.teapawt.org/fishmanticoder/status/…
Flowcharts from plain text, styled with CSS themes. Animated traffic, branches that fan out cleanly, dark or light. This is predictygraph.
predictygraph.github.io/
13
Flowcharts from plain text, styled with CSS themes. Animated traffic, branches that fan out cleanly, dark or light. This is predictygraph.
predictygraph.github.io/
Later today at OSS Europe: every tool is hostile, so AI agent limits belong in the tool, not the agent.
Deny-by-default scopes, human approval, microVM sandboxes, where HAMi fits, and do we need GPU passthrough?
osselceu2026.sched.com/event…
#OSSummit
14
Two ways to sandbox an AI agent without a VM:
NVIDIA OpenShell sandboxes the agent: gateway, images, policy prover, credentials injected only for approved hosts.
Sandlock (#OSSummit yesterday) sandboxes every command: no image, no runtime, ~5 ms per launch.
1
1
13
Sandlock: Landlock for fs/net/IPC, seccomp for syscalls, plus a supervisor for HTTP ACLs and copy-on-write workdirs. Rootless.
Cheap enough to wrap every pip install an agent runs. Pair it with an agent-level sandbox, not instead of one.
github.com/multikernel/sandl…
9
Fishmantic Coder retweeted
What a day at the @HAMiProject booth at @KubeCon_ China! 🎉 Non-stop visitors and great conversations from start to finish.
Huge thanks to Mengxuan Li, @jimmysongio, @fishmanticoder, Chris Gao, Xiao Zhang and Jifei Wang, for running the booth 🙌
#KubeCon #CloudNative #HAMi
If GPU memory goes tiered (fast HBM, slow HBF), OXMIQ models 14x capacity at 0.6x bandwidth. HAMi already does topology-aware placement and isolation - tiered memory extends that. vLLM picks what stays hot; the scheduler makes sure it lands there. But the question is will NVIDIA?
🤔 #HBF can pack far more data—but at a cost. OXMIQ’s simulation shows ~14x the capacity of #HBM, but just ~60% of the bandwidth, potentially eroding its cost advantage as inference scales. 💡More: pse.is/9jwupu 🔗
1
43
Good to see more work on ROCm for consumer chips.
Im happy to see ROCm 10 install matrix grow and more importantly make it easy to install.
rocm.docs.amd.com/en/latest/…
If you have feedback on where we can do better please let us know.
2
93
Put simply, the biggest technical bottleneck preventing CXMT from moving to the 1α node and beyond has now been removed.
China’s CXMT Successfully Applies High-k Materials to DRAM Mass Production, Closing In on Samsung and SK hynix
China’s ChangXin Memory Technologies (CXMT), a late entrant to the DRAM industry, is rapidly narrowing the technology gap with market leaders. The company has reportedly succeeded in applying high-k dielectric materials—once considered largely the domain of leading memory manufacturers—to its latest DRAM products in mass production.
In response, Samsung Electronics and SK hynix are accelerating development of next-generation architectures such as 4F² and 3D DRAM in an effort to maintain their technological lead.
Jeongdong Choi, Senior Vice President at TechInsights, discussed the latest roadmap for advanced DRAM technology development during a memory industry trends webinar on the 27th. TechInsights is a semiconductor-focused research and analysis firm.
China Begins Full-Scale Adoption of High-Performance High-k Materials, Rapidly Catching Up With Industry Leaders
The global DRAM market has traditionally been dominated by the “Big Three”: South Korea’s Samsung Electronics and SK hynix, and U.S.-based Micron. These companies have been expanding shipments of high-value AI memory products through the commercialization of 12nm-class DRAM and high-bandwidth memory (HBM).
Chinese companies, however, are catching up quickly. CXMT, currently China’s largest DRAM manufacturer, has successfully commercialized DDR5 and LPDDR5 products based on its G4, or 16nm-class, process.
More recently, CXMT has also begun using advanced materials in DRAM transistors that had previously been employed primarily by leading memory manufacturers.
“The important point is that our analysis shows CXMT has successfully introduced high-k metal gate, or HKMG, technology into its G4 process and LPDDR5X products,” Choi said.
High-k materials are used in insulating layers that prevent leakage current between circuits. Because they can store more charge at the same voltage, they enable more aggressive device scaling than conventional silicon dioxide (SiO₂) gate dielectrics. Samsung Electronics, SK hynix, and Micron began adopting these materials around four to five years ago.
CXMT is also advancing its HBM technology. The company is reportedly in the risk-production stage for HBM2E, the third generation of HBM, using its G3, or 18nm-class, DRAM process. It is also sampling HBM3 based on its G4 process.
“CXMT remains roughly two generations behind the leading players, but it continues to advance its HBM roadmap toward higher-performance memory solutions,” Choi said. “Its G5, or 15nm-class, DRAM process is also under development.”
DRAM Big Three Pursue Technological Shift Toward 4F² and 3D DRAM
Samsung Electronics, SK hynix, and Micron are seeking to maintain their technological advantage by developing next-next-generation DRAM below the 10nm node, known as D0a. In particular, 4F² and 3D DRAM, which fundamentally alter the conventional DRAM cell structure, are emerging as the most likely candidates.
“Based on our current assessment, Samsung Electronics and SK hynix are more likely to pursue 4F² DRAM,” Choi said. “Micron, on the other hand, appears more likely to move directly to 3D DRAM.”
4F² DRAM is a next-generation architecture that changes the orientation of the memory cell—the smallest unit used to store data—from a conventional planar arrangement to a vertical structure.
The “F²” notation refers to the area occupied by a memory cell relative to the minimum feature size, or F, of the manufacturing process.
Conventional DRAM typically uses a 6F² cell architecture. Reducing the cell area increases DRAM density, which can improve data-processing capability and power efficiency. This is why the industry is increasingly focused on 4F² structures.
3D DRAM, meanwhile, vertically arranges cells that have traditionally been integrated horizontally by standing up either the bitlines or wordlines. Bitlines and wordlines are the interconnect structures used to operate the transistors within each memory cell.
With a 3D DRAM architecture, more cells can be packed into the same footprint while increasing spacing between transistors, which can also reduce interference.
However, 3D DRAM is considered significantly more technically challenging than 4F² because it requires the introduction of new materials, bonding technologies, and other complex process innovations.
“3D DRAM remains an extremely challenging technology and would introduce substantial process complexity and yield issues if commercialized at the D0a generation,” Choi said. “For this reason, 4F² DRAM is emerging as the more practical candidate for the D0a generation.”
Fishmantic Coder retweeted
Samsung and SK hynix have been building 12-high HBM4 (36 GB) as the default stack for NVIDIA’s Vera Rubin.
NVIDIA has asked both suppliers to raise the share of 8-high HBM4 in second-half 2026 shipments, and that HBM4E may also lean 8-high. The stated reasons are thermal headroom, packaging yield, and supply flexibility: HBM4 doubles the I/O width to 2,048 pins versus HBM3E, which raises bandwidth but also heat and power in already dense GPU packages. An 8-high part uses fewer DRAM dies per cube, so more cubes can be produced from the same constrained wafer starts.
NVIDIA NVHBM moves the memory controller off the die into the HBM stack. Frees ~25% chip area, up to 30% bandwidth. NVIDIA owns the controller IP every custom XPU now depends on. They basically want a slice of every XPU sale. Smart and terrifying.
Announcing the expansion of NVIDIA NVLink Fusion with NVHBM, a next-generation high-bandwidth memory technology that brings higher memory performance and efficiency to XPUs.
Amazon's @AnnapurnaLabs will be the first to work with us on NVHBM, combining @awscloud custom silicon with our memory technology and the NVLink scale-up architecture to enhance performance and efficiency for AI workloads.
Learn how we're helping hyperscalers and AI innovators build the next generation of AI infrastructure: nvda.ws/4xpBtYN
GLM 5.2 on AMD with 80% the performance of a B200 at 1/3rd the cost.
wafer.ai/blog/glm52-amd































