Research window: 2026-11-16 05:00 ~ 2026-11-17 05:00 (Beijing time, 24 hours) Sources: Google News aggregation (24 query sets in Chinese and English, via proxy), Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International official website and blog, openRuyi official website and main repository, RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep, plus commit and release searches on GitHub across organizations and key repositories including XiangShan/NEMU/difftest/XSCache, CORE-V CVW/CVA6, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, and Radxa Linkr.


In This Issue

  • Today’s Focus: Three memory-access queue fixes in XiangShan Kunminghu and shared ROB enqueue selection (11-16)
  • I. Open-Source Community and Software Ecosystem
    • openRuyi six commits: kernel patch set and config sync refresh, Go backfill of SHA-2 AES optimization (11-16)
    • CORE-V Wally vector branch: basic vector memory access merged, vector LSU still missing (11-16)
    • difftest one commit: instruction progress check made configurable (11-16)
  • II. Product and Ecosystem Developments
    • SingNova packages four RISC-V product lines at security expo, desktop workstation spec raised to 256 TOPS (11-16)
  • III. Domestic Team Deep Dive
    • Sophgo SG2042/SG2044 boot rework: first partition changed to FAT32, kernel built per-chip workspace (11-16)
    • SpacemiT four repos: encoding input changed to borrowing, vision tokenizer normalization, OHOS Float16 fix (11-16)
    • Nuclei Nuclei Studio upgraded to 2026.10, docs site CI follows Ubuntu 26.04 (11-16)
  • IV. Embodied Intelligence and Robotics Special Section
    • SpacemiT realtime-link adds video transmission module, K3 VSLAM guide refreshed in same batch (11-16)
    • Three segments and two carriers of the robotics compute stack: link, SLAM, and model inference each in their place (11-16)
  • V. Summary and Trend Observations
    • Three-day trend: whole-machine architecture switches, open-source core validation fixes, standards race for matrix extensions (11-14 ~ 11-16)
  • Appendix: Complete source list and source verification table

Issue Summary: No major industry news within the window. The only new media piece is SingNova’s repackaging of four RISC-V product lines at a security expo (the same product set was already announced on September 30); the rest of the public information centers on the open-source side and engineering commits from domestic teams: three memory-access queue fixes in XiangShan Kunminghu, six kernel and package-management adjustments in openRuyi, basic memory access merged into the CORE-V Wally vector branch, Sophgo SG2042/SG2044 boot rework, and SpacemiT catching up on robotics links and HarmonyOS.

Today’s Focus: Three Memory-Access Queue Fixes in XiangShan Kunminghu and Shared ROB Enqueue Selection (11-16)

The XiangShan Kunminghu V3 branch landed four commits on November 16, three of them centered on memory-access queues. Two on the store queue side: one keeps unaligned store requests in the unaligned queue until dequeue, preventing in-flight requests from losing tracking after early release; the other fixes a defect where 16-byte cross-boundary requests (cross16B) mistakenly used stale information. One on the load queue side: retain uncache responses and align the timing of delayed release with the response path. The fourth commit is on the reorder buffer (ROB) side, sharing the enqueue head selection within the same bank group to reduce the area and timing overhead of each bank having its own selection logic.

What these four commits share is “boundary conditions.” When unaligned stores cross cache-line or page boundaries, requests linger in the pipeline for multiple cycles; when to release, which copy of information to use, and whether the response or the release comes first—any misalignment only surfaces under high load or specific alignment combinations. Uncache responses bypass the cache; if delayed release is misaligned with the response path, the data read and the data written back can diverge. Such defects never show up in benchmarks, but become intermittent errors in long-duration stress tests, virtualization, and multi-core coherence scenarios.

The ROB commit points in another direction: sharing enqueue selection across banks is a trade-off between area and frequency. XiangShan had just split the ROB’s functional and debug uop buffers on November 12; touching enqueue selection again this time shows that at this stage Kunminghu V3 is doing both “correctness closure” and “backend timing optimization” rather than piling on new features. For those watching open-source high-performance cores, such commits reflect engineering progress better than launch events: only once the defect density in the memory-access queues comes down can frequency and IPC improvements be discussed further.

Date: 2026-11-16 Sources: OpenXiangShan/XiangShan, PR #6930: ROB enqueue head selection shared per bank, PR #6822: unaligned queue retained until dequeue, PR #6770: cross16B requests mistakenly using stale information

I. Open-Source Community and Software Ecosystem

openRuyi: Six Commits — Kernel Patch Sets and Config Sync Refreshed, Go Backports SHA-2 AES Optimization (11-16)

The openRuyi main repository landed six commits on November 16. A morning commit restricted the toolchain major-version macros in rpm-config-openruyi to apply only to toolchain packages, preventing the macros from spreading to unrelated packages; of the five afternoon commits, three were substantive: the Go SHA-2 AES optimization was backported into the distribution, and linux and linux-lts each landed a “refresh patch set and refresh config” commit in sync; subsequently mupdf removed a temporary workaround for python clang, and python-maturin added the missing unicode-xid build dependency.

The Go commit deserves separate mention. Most RISC-V machines lack hardware AES instructions, so TLS and object-storage throughput depends heavily on the software implementation of crypto/sha256; backporting the upstream SHA-2 AES optimization means the distribution is starting to treat “cryptographic performance” as a first-class citizen on RISC-V, rather than waiting for upstream versions to roll forward naturally. The two kernel refreshes maintain the same baseline simultaneously on both the linux and linux-lts lines, indicating that the distribution uses the same patch set for the long-term-support line and the rolling line, so no line will miss patches during upgrades.

The mupdf and python-maturin commits are cleanup of the packaging chain. The former removes a workaround, indicating that the patch previously applied to bypass the python clang issue is no longer needed and the toolchain-side problem has been fixed at the source; the latter is a typical pitfall of Rust extension packaging, where a missing build-time dependency causes compilation to fail on the target machine. Such fixes are more valuable to people doing Python development on RISC-V machines than new features.

Date: 2026-11-16 Source: openRuyi-Project/openRuyi

CORE-V Wally Vector Branch: Basic Vector Load/Store Merged, Vector LSU Still Missing (11-16)

The vector branch of CORE-V Wally (CVW) merged PR #2070 at 17:08 on November 16, with a title that states it plainly: “WIP: add the most basic load/store, waiting for LSU to support vectors.” The author is eoinoconnell04, and it was merged by maintainer georgiatai. In other words, what was merged is the first-layer skeleton of vector memory access; the actual vector load/store unit is not yet in place.

The value of this information lies in positioning. Advancing the RISC-V vector extension on open-source cores typically proceeds in three steps: first the decode and execution units, then the vector register file and scheduling, and finally vector memory access. CVW’s prior work focused on the execution side of RVV instructions; this move into memory access indicates that the first two steps already have a usable foundation; but the simultaneous presence of the qualifiers “WIP” and “waiting for LSU” also indicates that there is still a distance to go before “running real vector workloads.”

A misreading to avoid is treating a branch merge as mainline support. CVW’s main branch had no commits in this window; all vector work is happening on the vector feature branch, and by open-source project convention such branches may persist for a long time or be rewritten at any moment. Teams using CVW as a teaching or tape-out baseline can use this to judge that the vector support timeline is not yet at a stage where products can be planned around it.

Date: 2026-11-16 Source: openhwfoundation/cvw PR #2070: basic vector load/store

difftest: One Commit — Instruction Progress Check Made Configurable (11-16)

The XiangShan differential testing framework difftest landed a commit on November 16 that makes the simulator-side instruction progress check a configurable option. Previously the progress check used a fixed policy, and in long regressions one could only rely on timeouts to judge whether something had “crashed” or was “running slowly”; after making it configurable, the check cadence can be set according to the target program’s instruction count, so hangs can be detected earlier and slow-but-normal test cases are not killed by mistake.

This forms a group with the two commits from November 13: first changing the fork’s checkpoint to 1 second and the packet slot count to 2 to the 22nd power, then making the progress check configurable. Together, the three are moving difftest from “able to compare” toward “able to run stably for long periods” — for a design of XiangShan’s scale, the stability of the verification infrastructure directly determines how many bugs can be converged each day.

Differential testing is a public good in the RISC-V ecosystem: it serves not only XiangShan but is also borrowed by other open-source core projects. The benefit of such changes is not in single-point performance but in lowering the hidden cost of “regressions that won’t run.”

Date: 2026-11-16 Source: OpenXiangShan/difftest PR #1043: configurable instruction progress check

II. Product and Ecosystem Updates

SingNova Packages Four RISC-V Product Lines at Security Expo, Desktop Workstation Spec Rises to 256 TOPS (11-16)

Singapore-based company SingNova (incubated by Nanyang Technological University, with NANYANG SINGTECH as the main entity) issued a press release on November 16, stating that at Safety & Security Asia 2026, held at Marina Bay Sands from November 10 to 12, it showcased four product lines: intelligent video storage servers, the AI Power Studio desktop AI computing system, AI virtual employees, and the Storage Station intelligent NAS. The specifications given in the release are: AI Power Studio uses a 12-core RISC-V CPU plus a dataflow architecture, with INT8 compute up to 256 TOPS, up to 128GB LPDDR5 ECC unified memory, can decode 128 channels of 1080p@30fps, and includes PCIe 4.0 x16; the video storage server uses chip-level intelligent encoding to compress video size to one-third; Storage Station is an all-in-one NAS based on a multi-core RISC-V architecture, supporting RAID-TP and 13-level file permissions.

What needs to be made clear is how new this press release actually is. The same set of four product lines had already appeared in full in the September 30 Tech Week Singapore press release, when the AI Power Studio spec was 200 TOPS INT8 plus 32 TFLOPS FP16; this time it only says “up to 256 TOPS INT8,” with no FP16 figure and no explanation of whether the difference between 200 and 256 comes from hardware configuration, sparse compute, or a different counting method. Therefore, this issue does not treat it as a new product launch, but rather as “the same product line repackaged for a security industry scenario.”

What is genuinely informative is the form factor. None of these four devices is a development board: they have chassis, management interfaces, industry protocols (RTSP/ONVIF, H.265/H.264), redundancy and auditing, and what is being sold is complete machines that can go directly into a server room. RISC-V here serves as the foundation of the complete machine, not a bare board for demonstration; the configuration of 128GB unified memory and 128-channel decoding also points to real workloads such as local large-model inference and video analytics, rather than benchmarks. The gaps are equally obvious: the press release does not say which company supplies the SoC, how much it costs, or when it ships, and the combination of a 12-core RISC-V with 256 TOPS is uncommon among publicly available chips; whether it is self-developed or an externally purchased platform cannot be determined from public information.

Date: 2026-11-16 Source: GlobeNewswire: SingNova showcases RISC-V AI computing and intelligent video storage product lines at Safety & Security Asia 2026, GlobeNewswire: September 30 Tech Week Singapore press release for the same product lines

III. In-depth Look at Domestic Teams

Sophgo SG2042/SG2044 Boot Overhaul: First Partition Switched to FAT32, Kernel Built Per-Chip Workspace (11-16)

Sophgo’s RISC-V boot repository bootloader-riscv landed two commits on November 16, both submitted by SophgoChaoWei. The first switches the first partition of SG2042 and SG2044 to FAT32; the second adjusts the build scripts so that openEuler kernel RPMs are produced in per-chip workspace directory trees rather than sharing a single output directory.

Switching the first partition to FAT32 is a firmware compatibility issue. The UEFI specification requires that partitions readable by firmware use a FAT-family filesystem, and RISC-V server platforms generally follow a UEFI-plus-GRUB boot chain; if the first partition is not FAT32, the firmware stage may fail to find the boot files. This kind of change looks like nothing more than a formatting parameter, but it actually determines “whether the official image can be recognized directly by firmware”—a hard threshold for users installing distributions like openEuler and openRuyi.

Building kernel RPMs per-chip workspace reflects the reality of maintaining multiple product generations in parallel. SG2042 is the previous-generation 64-core part, while SG2044 is the current server-class platform; the two have different kernel configurations and patch sets, and sharing one output directory easily leads to version cross-contamination. After splitting into per-chip workspaces, a single CI pipeline can produce kernel packages for both generations in parallel, and when problems arise they can be traced to a specific chip line.

In the same window, Sophgo also had a batch of commits on the edge side (13 for sophon-tools, libsophav 2.4.0, libsophon 0.4.16, cv84x6-llm-server 0.6.38), but per the repository notes those artifacts are arm64 debs belonging to the ARM edge product line. They are not counted as RISC-V items in this issue and are only recorded in the source verification table.

Date: 2026-11-16 Source: sophgo/bootloader-riscv

SpacemiT’s Four Repos: Encoder Input Switched to Borrowing, Vision Tokenizer Unified, OHOS Float16 Fix (11-16)

SpacemiT touched four repositories in this window. The multimedia library mpp added “synchronous input borrowing” for the encoder at 15:08, changing the input buffer’s lifetime from copy to borrow; the vision model library model-zoo-vision merged PR #121 at 17:24, converging the CLIP and Gemma tokenizers onto a common implementation; the AI SDK updated its vision submodule in the early hours of the next day; and the spacemit backend of llama.cpp merged PR #70 at 20:19, fixing a Float16 RTTI loading failure on OpenHarmony.

Encoder input borrowing is a typical zero-copy optimization. Every input frame in video encoding crosses the user space/driver boundary; if it is fully copied each time, both bandwidth and latency double. With borrowing, the driver uses the buffer held by the application directly, at the cost of stricter lifetime management—the application must guarantee the buffer is not released before encoding completes. This kind of change only shows its value at resolutions above 1080p and with multiple concurrent streams, and it also means the semantics of the MPP interface have changed, so downstream code needs to follow suit.

Merging the CLIP and Gemma tokenizers points to the maintenance cost of “one model library supporting multiple models.” Previously each model carried its own tokenizer, duplicating vocabulary and preprocessing logic, and a fix had to be made in multiple places; after converging on common, adding a new model only requires wiring up configuration. The Float16 RTTI loading failure on OHOS is an old C++ cross-ABI problem: HarmonyOS’s compiler and runtime handle type information for the half type differently from regular Linux, and dynamic library loading fails when the type information cannot be found. This fix shows that spacemit’s llama.cpp backend is already being used for real on HarmonyOS devices, not just adapted for Linux.

Date: 2026-11-16 Source: spacemit-com/mpp, spacemit-com/model-zoo-vision PR #121, spacemit-com/llama.cpp PR #70

Nuclei Studio Upgraded to 2026.10, Docs Site CI Follows Ubuntu 26.04 (11-16)

Nuclei’s Nuclei Studio documentation repository had two updates on November 16, both from fanghuaqi: at 14:54 the IDE version on the documentation homepage was updated to 2026.10, and at 15:21 the MkDocs deployment on the Ubuntu 26.04 runner was fixed. It should be noted that this repository hosts Nuclei Studio’s supply documentation, user guide, wiki, and discussions; the IDE itself is still released as a binary package. The README also notes that 2026.10 is the latest public version and lists the libraries needed to run riscv64 gdb and QEMU under Ubuntu 20.04.

Putting documentation and release notes in a public repository and accepting external issues and PRs is rare among domestic RISC-V IP vendors: most vendors only ship installers, hang documentation on their official websites, and users who find problems can only go through business channels. The Ubuntu 26.04 runner deployment fix looks trivial, but it is really about “whether the documentation can be published automatically”—GitHub Actions host image upgrades bring changes to Python and MkDocs dependencies, and if you don’t keep up, the publishing step fails silently.

For teams building products on Nuclei cores, this kind of maintenance cadence matters more than any single version update. The long-term usability of a toolchain and IDE depends on whether the documentation and dependency notes for the corresponding version can still be found three to five years later; putting supply documentation under version control and CI is a low-cost solution to this problem.

Date: 2026-11-16 Source: Nuclei-Software/nuclei-studio

IV. Special Chapter on Embodied Intelligence and Robotics

SpacemiT’s robotics repository spacemit-robotics/realtime-link committed the v1.0.0 “video transmission module” on the feat/zhangrou/add-video-transmission branch at 21:46 on November 16, authored by zhangrou651. Within the same window, spacemit-com/docs-ros merged PR #236 at 11:37, refreshing the SpacemiT VSLAM guide for the K3 platform.

Judging by its name, realtime-link is a real-time link component on the robotics side, with prior capabilities concentrated in control and status channels; adding video transmission means it is beginning to handle the return of perception data. For robotic systems, video links and control links have opposite requirements: control demands deterministic low latency and small data volumes; video demands high bandwidth, tolerates dropped frames, and requires adaptive bitrate. Putting both into a single component means either scheduling them over separate channels or implementing priorities at the transport layer—the specifics will have to wait for more code to be made public.

The K3 VSLAM guide refresh is a signal of software maturity. On RISC-V, VSLAM is not a question of whether it runs, but whether frame rate, accuracy, and power consumption can all meet targets simultaneously; a vendor willing to write up K3’s VSLAM configuration as a guide indicates that this pipeline has already been run on real hardware and that customers can be given reproducible steps. Viewed together, the two materials show that what SpacemiT is filling in on the robotics side is the “video in, pose out” segment—precisely the perception front end ahead of VLA models.

Date: 2026-11-16 Source: spacemit-robotics/realtime-link, spacemit-com/docs-ros PR #236: Refresh K3 VSLAM guide

Arranging today’s several commits along the data flow, robotics-side compute can be divided into three stages. The first stage is the video link: frames captured by the camera must be delivered to the compute unit with low latency, corresponding to the video transmission newly added to realtime-link. The second stage is SLAM and geometry: turning frames into poses and maps, corresponding to the K3 VSLAM guide. The third stage is model inference: vision-language-action models running locally, corresponding to the spacemit backend of llama.cpp and the tokenizer normalization in the vision model library. The three stages place different demands on hardware—the link needs I/O and real-time capability, SLAM needs single-core performance and memory bandwidth, and inference needs matrix compute throughput and capacity.

Form factors are also diverging. One category is desktop or rack-style workstations: SingNova’s AI Power Studio pairs a 12-core RISC-V with 128GB of unified memory and 128-channel decoding, with the selling point of running large models plus video analytics locally; the other is embedded boxes: SOPHGO’s CV84X6 LLM Server is packaged as an arm64 deb, offering an OpenAI-compatible interface plus a web management page, with the selling point of being usable right out of the box. The former relies on capacity and bandwidth, the latter on power consumption and integration—neither follows the “one chip solves everything” approach.

RISC-V’s position in this layering needs to be stated precisely: it is currently strong in the link and SLAM—platforms like the K3 have real multi-core and vector capabilities and can handle front-end processing; it is weak on the inference-side software stack—llama.cpp’s backend is still fixing ABI-level issues, and the model library’s tokenizer has only just been normalized. In other words, robotics requires “all three stages to work,” and RISC-V currently has two working, with the third being filled in.

Date: 2026-11-16 Source: spacemit-robotics/realtime-link, spacemit-com/docs-ros, spacemit-com/llama.cpp, GlobeNewswire: SingNova product line press release

V. Summary and Trend Observations

There was no launch event today; all four items are “catch-up” work: XiangShan is closing boundary defects in the load/store queue, openRuyi is polishing kernel and crypto performance, CORE-V is filling in the first-layer skeleton of vector load/store, and SpacemiT is fixing HarmonyOS ABI issues and the robot video pipeline. Submissions like these do not make headlines, but they determine whether a chip can be developed on directly once it is in hand, and whether it can be maintained over the long term.

The first judgment is that open-source high-performance cores have entered the “correctness closure” phase. XiangShan’s submissions this week shifted from CI step splitting and documentation governance back to specific load/store queue defects, indicating that the feature surface of Kunminghu V3 has been frozen and that what comes next is a contest over defect density and frequency. For teams using open-source cores to build products, this phase is actually a good time to get involved: interfaces are stable, defects are being systematically cleaned up, and the verification infrastructure is strengthening at the same time.

The second judgment is that distributions have begun to treat performance as a first-class citizen. openRuyi backported Go’s SHA-2 AES optimization and synchronously refreshed the patch sets for two kernel lines, an approach consistent with mature distributions: rather than waiting for upstream to roll naturally, it pulls the paths it cares about forward. RISC-V cannot catch up in general-purpose compute in the short term, but on “being fast enough on specific paths” it can rely on distributions to fill the gap, and this is where domestic distributions can differentiate themselves.

The third judgment is that on the robotics side, the competitive focus is shifting from chips to pipelines. Both of SpacemiT’s items today concern the “video in, pose out” segment rather than another new chip; the two form factors from SingNova and Sophgo also show that the overall machine form factor and degree of software integration determine whether something can be deployed more than peak compute does. For teams working on embodied intelligence, the first question to ask during selection is no longer “how many TOPS” but “among the three segments of video pipeline, SLAM, and inference, which one do I need to fill in myself.”

Trend over the past three days: complete machines switching architectures, open-source cores filling in verification, standards contending over matrix extensions (11-14 ~ 11-16)

Looking at November 14 to 16 together, three lines each advanced. On the complete-machine side, Purism launched the Librem 5 v2 crowdfunding campaign on November 14, switching the SoC of its next-generation Linux phone to SpacemiT K1, with both the $1,599 crowdfunding price and the 2028 delivery timetable made public; on November 16, multiple foreign media outlets continued to follow up with wording consistent with the official page, so this issue does not repeat the coverage.

On the open-source core side, FURSOY published fursoy-cpu on November 15, an out-of-order RV64 core that can boot Linux 6.6.50 and closes at 500MHz on 45nm, with deliverables including SBI firmware, device trees, and Sail differential verification scripts; on the same day, XiangShan was closing load/store defects and CVW was filling in vector load/store. The three items point to the same conclusion: competition among open-source cores has shifted from “whether it can run” to “verification and reproducibility.”

On the standards side, FPRox on November 15 compared five proposals for the RISC-V matrix extension side by side—three layout categories, dot product, fat inner product, and outer product, corresponding to IME, VME, AME and Zvdot4a8i, Zv[b]dota. A large number of proposals is not itself a problem; incompatibility is: if things ultimately fail to converge on the mainline, operator libraries and compiler backends will have to maintain a separate path for each combination. The common constraint across these three lines remains the software stack, and the progress of the software stack is precisely reflected in seemingly trivial submissions like today’s.

Appendix

Source Verification Table

Source Verification Result
Google News aggregation (24 Chinese and English query sets, 24-hour window) The only substantive new item within the window was a SingNova security-expo press release, which has been included with a note on its duplication relationship with the September 30 press release. Excluded items: multiple November 16 follow-ups on Purism Librem 5 v2 (BigGo, Notebookcheck, Mezha, Linuxiac, OMG! Ubuntu) are a continuation of the November 14 event, which was already covered as a top item in the November 16 report; KuCoin’s November 16 re-publication of “Alibaba T-Head’s Hong Kong IPO” is in fact a November 10 event already covered in the November 10 and November 11 reports; the Lingrui Zhixin IC Fengyun List candidate company report (November 16) is a company profile, and the P100’s SPEC 20+/GHz, 8-decode 10-issue, and SMT4 figures are all old data already cited in the November 13 funding report, so it is not re-included; the FURSOY Hackster follow-up is a November 15 event; all 30 Milk-V query results were dairy and food-market articles
Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International Within the HN window there was only one item, “RISC-V Matrix Extensions,” already covered in the November 15 report; the only RISC-V-related Phoronix item in the window was Librem 5 v2, already covered in the November 16 report; all four CNX Software articles on November 16 were unrelated to RISC-V (AMD mini PC, Jetson Thor edge box, Home Assistant controller, KAMRUI promotion), and its latest RISC-V-related piece was the November 14 ESP32-P4 desktop system; LinuxGizmos had no new articles on November 16, with its latest RISC-V-related item being Librem 5 v2 on November 15; the riscv.org blog and news have had no additions since October 20
GitHub organization and key repository checks (30+ organizations and repositories) Active within the window: XiangShan main repo 4 commits, difftest 1 commit, openRuyi-Project main repo 6 commits, revyos/runner 1 commit, spacemit-com 5 repos (mpp, model-zoo-vision, ai-sdk, docs-ros, llama.cpp), spacemit-robotics 1 commit, sophgo 14 repos (including 2 commits to bootloader-riscv), Nuclei-Software 2 commits, openhwfoundation 7 repos (including CVW vector branch PR #2070), radxa-linkr 3 repos (documentation submodule updates), sipeed 5 repos. DAMO XuanTie (XUANTIE-RV) and ESWIN (eswincomputing) had zero pushes within the window. The anonymous API hit the 60-requests-per-hour rate limit during this issue, so organization-level activity was verified instead via each repository’s Atom commit feed and branch probing; pushes to non-default branches may be underestimated
Exclusions: Sophgo edge product line and Sipeed NanoKVM-Pro Sophgo’s sophon-tools had 13 commits within the window, concentrated on the CV84X6 edge LLM service (0.6.37 / 0.6.38), WeMM zero-vector and device-copy completion-order fixes, and the vl_embed acceptance tool, with libsophav bumped to 2.4.0 and libsophon to 0.4.16; per the repository README, the cv84x6-llm-server artifact is an arm64 deb, belonging to the ARM edge platform, so it is not listed as a RISC-V item. Sipeed NanoKVM-Pro 1.2.16 adds MCP remote control, but its main controller is the AX630C (dual-core Arm A53), not RISC-V, so it is not included
riscv100.top inspection No changes: the homepage What’s New, Upcoming, and committee members, as well as the six SPEC CPU2006 / 2017 / 2026 leaderboards under the cross-architecture and RISC-V channels, all show no new records or score changes, so no space is devoted to them in the body
RISC-V member website inspection (391 sites) Took 207 seconds; 268 unchanged, 53 reported content changes, 70 fetch failures. Each change item was checked individually and all were homepage marketing rotations (NVIDIA livestreams and reference architectures, AMD branding, Axelera benchmark pages, Codasip, Espressif, Huawei, etc.), with no new products, funding, or architecture announcements, so none were included
openRuyi official news page and main repo The news page’s latest items remain the two from October 2026 (ecosystem infrastructure, Ruyi Firmware) and the 2026.09 release, with no new version within the window; the 6 commits to the main repo have been logged as recorded

Full Source List

Limitations: This issue covers a 24-hour window (November 16, 05:00 to November 17, 05:00). On the media side there was only one vendor press release; the main content comes from GitHub commit logs, so technical details are based on commit titles and repository descriptions and were not verified line by line. The GitHub anonymous API hit the 60-requests-per-hour rate limit during this issue, so organization-level activity was verified instead via each repository’s Atom commit feed and branch probing; pushes to non-default branches may be underestimated. CVW’s vector work and spacemit-robotics’ video module are both on feature branches, and whether they will enter the mainline is not yet determined. SingNova’s 256 TOPS is the company’s press-release figure; the September 30 press release for the same product line stated 200 TOPS INT8 and 32 TFLOPS FP16, and the discrepancy between the two versions has not been explained. The SoC supplier, pricing, and shipping date were all undisclosed. The architecture determination for Sophgo’s CV84X6 edge products is based on the “arm64 deb” artifact description in the repository README, and it cannot be ruled out that this series has a separate RISC-V model. The 53 “content changes” in the member website inspection were determined by page-text hash; after reading each one, all were marketing updates, though there is a possibility that an individual announcement was misclassified as marketing.