RISC-V Daily Intelligence Report (2026-11-19)
Research window: 2026-11-18 05:00 ~ 2026-11-19 05:00 (Beijing time, 24 hours) Sources: Google News aggregation (24 query sets in Chinese and English, plus 38 additional Chinese team-term and vendor English queries, via proxy), Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International official website and blog, openRuyi official website and main repo, RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep, and commit and release searches across GitHub organizations and key repositories including XiangShan/NEMU/difftest/XSCache/GEM5, CORE-V CVW/CVA6/core-v-verif, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, Radxa Linkr, and RuyiSDK.
This Issue’s Index
- Today’s Highlights: Zijing M100 automotive-grade chip enters volume delivery; cumulative automotive-grade chip shipments surpass 10 million units (11-18).
- I. Open-Source Community and Software Ecosystem.
- Four XiangShan repos advance on the same day: NEMU releases v2026.11.r9, ICache fixes two issues, difftest switches to downstream Chisel (11-18).
- CORE-V: CVW merges AHB compliance and ACT granule zeroing (11-18).
- openRuyi: twenty commits — curl upgraded to 8.23.0, nginx security fixes, fourteen Go ecosystem packages added in one batch (11-18).
- II. Product and Ecosystem Updates.
- SpacemiT showcases K1/K3 and V100 servers at the 5G+ Industrial Internet Conference (11-18).
- III. Deep Dive on Domestic Teams.
- SpacemiT llama.cpp backend: parallel repack of MoE expert weights, gated delta net enabled on K1 (11-18).
- Sophgo removes multi_video_qt subproject; Sipeed fixes MAC conflict in build artifacts (11-18).
- IV. Special Chapter: Embodied Intelligence and Robotics.
- SpacemiT robotics stack: operator layer opens read-only joint and IMU feedback, realtime-link adds video transmission (11-18).
- V. Summary and Trend Observations.
- Three-day trend: automotive-grade delivery numbers first, software version bumps, robotics interface additions (11-16 ~ 11-18).
- Supplementary Observation: No new entries on riscv100.top this issue; Upcoming matches the previous issue (11-19).
- Appendix: Full source list and source verification table.
Issue Summary: Today’s main thread is automotive-grade chip delivery numbers. Zijing Semiconductor’s M100 has entered volume shipment, with the company’s cumulative automotive-grade chip shipments surpassing 10 million units. The customer list and year-end order expectations were disclosed at the same time. On the open-source side, it was a day of cleanup and compliance. Four XiangShan repos advanced on the same day, with NEMU issuing a release and fixing two cache invalidation issues. CVW merged AHB compliance and ACT granule zeroing. openRuyi landed 20 commits in one day, adding fourteen Go packages alongside security updates. On the domestic team side, SpacemiT exhibited K1, K3, and V100 systems at the 5G+ Industrial Internet Conference in Wuhan. On the same day, it added MoE parallel repack and gated delta net on K1 to its RISC-V backend for llama.cpp.
Today’s Focus: Zijing M100 automotive-grade chip enters volume delivery, cumulative automotive-grade chip shipments surpass 10 million units (11-18)
From the evening of November 17 to November 18, several Hunan media outlets reported on Zijing Semiconductor’s RISC-V automotive-grade chip. Hunan Daily, Rednet, the Hunan Provincial Government Portal, and China News Service Hunan all published stories. There are three core facts. First, the Zijing M100 has entered mass production in vehicles and is being delivered stably in volume. Second, as of October this year, the company’s cumulative automotive-grade chip shipments exceeded 10 million units. Third, Zijing was incubated under the leadership of Great Wall Motor and established in Changsha Economic and Technological Development Zone in November 2024.
The certifications and customer figures given in the reports are as follows. The M100 fully passed AEC-Q100 automotive chip reliability testing and multiple other certifications in July this year, qualifying it to enter mainstream automakers and tier-one supply chains. Customers have expanded to more than 30, including Mande, Hive Precision, Steelmate, and United Automotive Electronic Systems. Orders are expected to reach 30 million yuan by the end of this year. The product’s entry point is body control, used in combination headlights, ambient lighting, and combination switches, with up to 17 chips per vehicle. Chief Strategy Officer Xu Xingkui said this type of chip occupies a peripheral position but has large volume, making it the least resistant to substitution. The second product line is the embodied intelligence chip E100; the company says about 70% of the technology foundation for automotive-grade and embodied intelligence can be directly reused.
The significance of this news lies in the customers and volumes, not the launch. The M100 was launched on December 18, 2025, at the general assembly of the China Automotive Chip Industry Innovation Strategic Alliance. At that time, the disclosed figure was an initial order of 250,000 units, of which 100,000 had been delivered. Performance metrics given were CoreMark 2.42, a 38% improvement over comparable products, and compliance with ASIL-B and national cryptographic requirements. This cycle adds three new items. Cumulative shipments have crossed 10 million units. Customers have moved from “cooperation intentions” to more than 30 that can be named. Year-end orders now have a monetary figure.
Several points warrant discounting. The 10 million units is a cumulative figure for “automotive-grade chips,” not broken down between the M100 and other models, and no monthly shipment figure was given. The 30 million yuan is an order expectation, not confirmed revenue. Extrapolating from automotive-grade MCU unit prices, this scale is still small-batch. The 11.9% domestic substitution rate cited by multiple media outlets refers to the first quarter of this year and comes from the China Association of Automobile Manufacturers. It measures automotive-grade chips overall, not the RISC-V architecture alone. The real observation point is next year. The first vehicle model to carry the M100 is scheduled to launch in 2026; whether it can move from body control to domain control and powertrain will determine how far this path can go.
The policy and exhibition context also overlays this news. The 2026 World Computing Conference will be held from November 19 to 21 at the Changsha International Conference Center, with the opening ceremony on the morning of the 20th. The theme is “Computing for All Things, Meeting the Future in Hunan.” Zijing will speak on RISC-V and China’s automotive chip localization under the “Architecture Innovation, Chip-Chain Synergy” topic at the computing chip forum. Hunan is treating this company as a showcase for local integrated circuits. The narratives of “patient capital” and “relocating headquarters” in the reports are inseparable from the investment-promotion context.
Date: 2026-11-18 Sources: Hunan Provincial Government Portal: Hunan-made RISC-V automotive-grade chip enters mass production in vehicles, Hunan Today: RISC-V automotive-grade chip enters mass production in vehicles this year, Rednet: World’s first to enter mass production in vehicles, Hunan-made automotive-grade chip builds a solid foundation for self-reliance, China News Service Hunan: Hunan-made RISC-V automotive-grade chip enters mass production in vehicles
I. Open-Source Community and Software Ecosystem
XiangShan’s four repos advance on the same day: NEMU releases v2026.11.r9, ICache gets two fixes, difftest switches to downstream Chisel (11-18)
On November 18, four XiangShan project repositories saw activity on the same day. The main XiangShan repo landed two ICache fixes on the kunminghu-v3 branch, at 16:50 and 16:52 respectively, authored by vincent-163. The first switched the ITLB lookup to fb(0) and carried over its gpaddr. The change touches three files—Bundles, ICachePrefetchPipe, and ICacheWayLookup—adding 71 lines and removing 60. The second stops reporting I_ECC for a certain class of fetch request slots in ICacheMainPipe, changing only 3 lines.
The reference model NEMU released v2026.11.r9 at 20:30 that evening. Two fixes were submitted by poemonsense the same day: clearing the tcache after address translation invalidation (#1300), and clearing the tcache after a directly trapping hostcall (#1304). Both point to the same class of problem: the translation result changed, but the stale mapping in the cache was not invalidated along with it.
The differential testing repo difftest landed four commits. First, a performance change that scans the ROB when mrelease completes (#1047). Second, a wholesale switch to the downstream-maintained Chisel 7.16.0-xiangshan-1 (#1045). Third, CI switching to xs-env to pin the mill version (#1046). Fourth, the FPGA side draining FORK checks in the exceed state (#1037). The xs-dev branch of the performance model GEM5 aligns KMHV3’s IQ capacity with the RTL (#1312).
Taken together, what this day accomplished was aligning models with hardware and cleaning up state. Switching Chisel to a downstream version means XiangShan no longer fully follows the upstream Scala build chain and maintains its own usable version. Pinning mill in CI is about making build results reproducible. Both ICache fixes are on the fetch path. The ITLB and gpaddr handling concerns boundary conditions in address translation. Narrowing I_ECC avoids misclassifying normal fetch slots as check errors. Changes like these won’t improve performance metrics, but they determine whether differential tests run cleanly.
Date: 2026-11-18 Source: OpenXiangShan/XiangShan (kunminghu-v3), OpenXiangShan/NEMU v2026.11.r9, OpenXiangShan/difftest, OpenXiangShan/GEM5 (xs-dev)
CORE-V: CVW merges AHB compliance and ACT granule zeroing (11-18)
openhwfoundation’s CORE-V Wally (CVW) merged two PRs on the evening of November 18. At 20:41 came davidharrishmc’s dh/act-granule-zero (#2128), and at 20:44 dh/ahb-compliance (#2120). The merger was rosethompson. Both are compliance wrap-ups: one changes granule zeroing in the architectural compliance test (ACT), the other adds AHB bus compliance.
The context is worth noting too. The CVA6 main branch had no commits in the window; the most recent was the November 17 tree sync merging master into main. core-v-verif also had no new commits in the window. The CORE-V family’s progress that day centered on verifying consistency rather than adding features. AHB compliance maps to SoC integration scenarios. For an open-source core to go into someone else’s chip, its bus behavior must first pass consistency checks. ACT granule zeroing, meanwhile, bears on whether the compliance test’s own results can be trusted.
Date: 2026-11-18 Source: cvw PR #2120: AHB compliance, cvw PR #2128: ACT granule zeroing, openhwfoundation/cva6
openRuyi’s twenty commits: curl bumped to 8.23.0, nginx gets security fixes, fourteen Go ecosystem packages in one go (11-18)
The openRuyi main repo landed 20 commits over the course of November 18. Two security updates come first: curl bumped to 8.23.0 and nginx to 1.31.9. Both commit titles carry security-fix markers. On the base component side, tpm2-tss was updated to 4.2.2, dbus enabled systemd and selinux support, and bash-completion dropped %{?_isa} and reverted to noarch.
The batch at 14:57 that day was the main event. Fourteen Go ecosystem packages were added at once, plus the network probing tool nexttrace bumped to 1.7.6 and granted capabilities. Among the fourteen, quic-go entered the repo together with qpack and ossfuzz seeds. The rest are dnscrypt-v2, adguardteam/golibs, ameshkov/dnsstamps, jedisct1/go-dnsstamps, sthorne/odoh-go, and cisco/go-hpke. Also cisco/go-tls-syntax, natesales/q, natesales/bgptools-go, robfig/cron-v3, and c2h5oh/datasize.
The common thread is clear: an encrypted DNS plus QUIC/HTTP3 stack, layered with HPKE and TLS syntax libraries. Encrypted DNS covers DoH, ODoH, and dnscrypt. Pulling the dependency chain into the distro means users don’t have to go install on a RISC-V machine themselves, and build-time versions stay controllable. Granting nexttrace capabilities is the other side of the same idea. Network probing needs raw sockets; configuring the permission bits at packaging time saves users from sudo.
Date: 2026-11-18 Source: openRuyi-Project/openRuyi, openRuyi official news page
II. Product and Ecosystem Developments
SpacemiT shows K1/K3 and V100 servers at the 5G+ Industrial Internet Conference (11-18)
On November 18, the 2026 China 5G+ Industrial Internet Conference opened at the Wuhan Optics Valley Science and Technology Exhibition Center. SpacemiT’s booth brought the K1 and K3 edge chips, along with V100-based server systems and cluster solutions. The on-site framing came from Gao Zhiyu, the company’s senior director of government and enterprise affairs. K1 is the edge chip released in 2024, capable of dual-domestic adaptation with OpenHarmony. K3, launched in 2026, is described as the world’s first mass-produced RISC-V chip compliant with the RVA23 standard, with 60TOPS of AI compute, and has been adapted for Ubuntu and OpenHarmony. V100 was announced this October at the RISC-V China Summit and is described as the industry’s first compute chip fully compliant with the RISC-V server specification, paired with three classes of systems: compute, intelligent compute, and storage.
The show highlighted two use cases. One is a dev board serving as a robot’s control cerebellum, handling limb movement. The other is servers plus compute cards forming a small local inference center, with a single rack targeting the storage, compute, and AI inference needs of SMEs and government/enterprise customers. The company also mentioned working with the Institute of Software at the Chinese Academy of Sciences to refine domestic OS adaptation. The OpenHarmony solution has already been used in education terminals, hospital bedside screens, campus displays, and facial-recognition access control.
It’s worth being clear about the nature of this news. It’s a trade-show appearance, not a product launch. K3’s RVA23 and 60TOPS, and V100’s server specification, are all claims already made public at the October summit. What’s new this issue is the outward display of system and cluster form factors, and explicitly positioning the edge chips as robot controllers. K3’s RVA23 compliance remains a vendor claim for now, with no corresponding result on third-party leaderboards. On riscv100.top’s Upcoming list, the K3-with-XSCC evaluation preview has been posted for some time, with no data yet.
Date: 2026-11-18 Source: Jimu News: Chips as robot “cerebellums,” domestic RISC-V chips appear at 5G+ Industrial Internet Conference, Phoenix News reprint of the same article
III. Deep Dive into Domestic Teams
SpacemiT llama.cpp Backend: Parallel Repack of MoE Expert Weights, K1 Enables gated delta net (11-18)
The mtmd-backend branch of spacemit-com/llama.cpp landed two commits on the morning of November 18. At 10:49, #71 “enable gated delta net on A60 (K1)” was merged. The change is a single conditional. When the GATED_DELTA_NET operator selects a tensor strategy, use_ime1 (the A60 core on K1) is now placed alongside use_ime2. The author’s email is from the Institute of Software, Chinese Academy of Sciences. At 12:00, #55 “optimize repack mode, switch to parallel repack” was merged. This commit adds a CMake option GGML_SPACEMIT_REPACK_MT to ggml-spacemit, off by default. When enabled, MoE expert weights are repacked in parallel using X100 cores at load time, adding 5 lines on the CMake side and 66 lines in the implementation.
gated delta net is a core operator of linear-attention-style architectures. Previously this backend only took the generic path on IME2. Including IME1 means older platforms like K1 can run this class of models, rather than only newer chips. Parallel repack addresses load time. The number of expert weights in MoE models grows linearly with scale, and single-threaded repack can drag startup to the minute level. This step affects only loading, not inference, and is well suited to being spread across multiple cores.
One background note: neither the xslim nor the onnxruntime repository had any commits in this window. SpacemiT’s software activity on this day was concentrated on the inference backend and the robotics side.
Date: 2026-11-18 Source: spacemit-com/llama.cpp PR #71, spacemit-com/llama.cpp PR #55
Sophgo Removes multi_video_qt Subproject, Sipeed Fixes MAC Conflicts in Build Artifacts (11-18)
Sophgo’s sophon-tools merged PR #249 at 10:51 on November 18, removing the multi_video_qt subproject, ticket number MYS-2941. At 16:24 the same afternoon, Sipeed’s LicheeRV-Nano-Build submitted a commit “reduce MAC address collisions,” addressing duplicate NIC MAC addresses in build artifacts.
Both are maintenance changes with little technical depth, but they point in the same direction. The repositories of domestic RISC-V board and system vendors have entered a phase of deleting old modules and fixing build details. Deleting a subproject means a product line or demo path is no longer maintained and needs to be cleared out of the repo. MAC address collisions cause multiple boards on the same LAN to knock each other off, a problem that must be resolved before mass production. These are worth recording because such changes say more about where a product stands than a launch event does.
Date: 2026-11-18 Source: sophgo/sophon-tools PR #249, sipeed/LicheeRV-Nano-Build
IV. Embodied Intelligence and Robotics Special Section
SpacemiT Robotics Stack: Operator Layer Exposes Read-Only Joint and IMU Feedback, realtime-link Adds Video Transmission (11-18)
Two repositories under spacemit-robotics each landed a commit on November 18. humanoid_common submitted “feat(operator): expose read-only joint and IMU feedback” at 14:44, merged as PR #34 at 19:55. Author beililcj. realtime-link merged PR #3 “feat(v1.0.0): add video transmission module” at 14:07.
The read-only qualifier deserves separate mention. The operator layer handles teleoperation and task orchestration and needs to see joint angles and IMU data to display posture, record trajectories, and supervise policy output. But if this layer could write back, teleoperation commands and autonomous policies would overwrite each other, and when something goes wrong it would be impossible to tell who sent the command. Exposing it as a read-only interface amounts to separating what can be seen from what can be changed. This is a standard boundary design in multi-control-source systems, and it must be done deliberately in code. realtime-link adding video transmission fills in the other essential input for teleoperation. Together, the two commits open up both the video and the body-state links.
The timing also lines up. The previous day (November 17), humanoid_common had just merged adaptation for the Hiking perception-based locomotion policy (PR #33), transmitting depth maps over ZMQ and retaining the old policy on handoff failure. Today’s feedback interface and video transmission let the upper layer observe what this policy is doing. Bringing the policy in first and then completing observability within the same week is the right order.
Date: 2026-11-18 Source: humanoid_common PR #34, realtime-link PR #3
V. Summary and Trend Observations
The most concrete item today is the automotive-grade chip delivery numbers. Zijin M100 has gone from its December 2025 launch to a named list of more than 30 customers, cumulative shipments of 10 million automotive-grade chips, and an expected 30 million yuan in orders by year-end. The path chosen is pragmatic. Entering through edge but high-volume scenarios like body control, with up to 17 chips per vehicle, keeps validation costs below those of powertrain and chassis. A caveat on the figures: the 10 million chips are not broken down by model, the 30 million yuan is an expectation rather than revenue, and the 11.9% refers to automotive-grade chips overall, not the RISC-V architecture.
The second judgment is that the open-source side today is doing cleanup and compliance rather than chasing metrics. Four XiangShan repositories moved on the same day: ICache fixed an address translation boundary, NEMU issued a release and cleaned up tcache, difftest switched to a downstream Chisel and pinned the build tool version, and GEM5 aligned IQ capacity. CVW merged AHB compliance and ACT granule zeroing. None of these changes produce performance numbers. They determine whether differential testing runs cleanly, and whether an open-source core can be dropped into someone else’s SoC.
The third judgment is that the robotics side is filling in interfaces, not algorithms. humanoid_common exposes joint and IMU feedback as read-only, and realtime-link adds video transmission; neither touches the policy itself. A third-party perception policy was merged just the day before, and today observability is filled in. This order shows the team is preparing for multiple policy sources to coexist. A chip vendor cannot build every policy; what it can do is standardize the four stages of input, inference, handoff, and observation.
Three-Day Trend: Automotive Ships Numbers First, Software Catches Up on Versions, Robotics Fills In Interfaces (11-16 ~ 11-18)
Looking at November 16 through 18 together, the three threads land in different places. On the automotive and system side, StarFive announced mass-production delivery of its server BMC chip JH-B100 on November 17. On November 18, Zijin laid out cumulative automotive-grade shipments and its customer list. Both answer the same question: where does RISC-V sit in the BOM of a whole vehicle and of a server. One is a small chip in the management plane, the other a class of small chips in body control. What they share is avoiding a head-on collision over main compute.
On the open-source core and verification side, XiangShan was closing memory-access queue defects and CVW was filling in vector memory access on November 16. On November 17, core-v-verif brought CSR state into the RVFI log, and XiangShan did a CSR naming refactor. On November 18, ICache fixed a boundary, CVW added AHB compliance, and NEMU cleaned caches. All three days of activity landed on verifiability and maintainability; not one was performance tuning.
On the software and toolchain side, openRuyi did version follow-up for three consecutive days. On November 17, 19 commits pushed OpenSBI to 1.9 and introduced OpenSSL 3.6.5. On November 18, 20 commits completed security updates for curl and nginx and pulled fourteen Go packages into the repository. SpacemiT, meanwhile, filled in on both the inference and robotics ends. The llama.cpp backend lets K1 run linear attention operators, and the robotics stack separates out the feedback interface. Taken together, this week’s progress is all in engineering delivery; the architecture narrative was comparatively quiet.
Supplementary Observations
riscv100.top: No New Entries This Issue, Upcoming Unchanged from Last Issue (11-19)
This round’s inspection of riscv100.top was judged “no major changes (minor updates only).” Across the cross-architecture and RISC-V channels, the six SPEC CPU2006/2017/2026 leaderboards show no new records, no score changes, and no additions or removals of committee members. Upcoming remains a single entry, LineShine (CPU-Only TOP500 Leader), identical to what the November 18 report recorded. This issue does not count it again.
Worth noting is that the K3 with XSCC Vectorization/IME benchmark preview has still not materialized. That preview is one of the few public windows into the actual gains from vector and matrix extensions. Once results appear, they will be handled as a leaderboard entry, with a note on comparability to results from the RTL 3GHz platform.
Date: 2026-11-19 Source: riscv100.top
Appendix
Source Verification Table
| Source | Verification Result |
|---|---|
| Google News aggregation (24 query sets in Chinese and English, plus 38 additional team-term and vendor English queries) | Two substantive new items within the window: the delivery figures for the Cercis M100 (first published on the evening of November 17, reposted by the Hunan Provincial Government Portal and China News Service Hunan on November 18, recorded under 11-18) and SpacemiT’s exhibit at the 5G+ Industrial Internet Conference, both included. Excluded items: the mass-production delivery of the StarFive JH-B100 is a November 17 event already covered in the November 17 and 18 reports; Andes Technology’s FY2026 Q3 earnings call was held on November 13, and the November 18 aggregation piece is a retrospective, outside the window; the MIPS Actus robotic joint platform was launched at electronica on November 9, and the November 18 India edition report is a follow-up, outside the window; Ingenic’s Q&A on the XBurst and Victory dual cores is an interactive-platform reply from 15:00 on November 17, outside the window, and was already excluded as a company profile item in the November 18 report; Alibaba’s spin-off of T-Head for a Hong Kong listing and Xingyun Technology’s supply-chain statements concern PPU and the Yitian line, not RISC-V; the OpenAI–AMD chip design collaboration and Siemens’ acquisition of Breker contain no RISC-V-specific information; several Espressif- and Sophgo-related pieces are stock-market reports, excluded as noise |
| Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International | No new posts on HN within the window; the most recent is still the November 17 Librem 5 v2 (100 points, 115 comments), already included; the Phoronix RSS feed contains only one old Purism item within the window; CNX Software and LinuxGizmos have no new RISC-V pieces within the window; the latest items on the riscv.org blog and news pages remain the November 17 five-question interview (Doug Norton and Nick Brown), with nothing new within the window |
| GitHub organization and key repository checks (40+ repositories) | Active within the window: XiangShan 2 commits, NEMU 2 commits plus the v2026.11.r9 release, difftest 4 commits, GEM5 1 commit, cvw 2 commits, openRuyi 20 commits, spacemit-com/llama.cpp 2 commits, humanoid_common 2 commits, realtime-link 1 commit, sophon-tools 2 commits, LicheeRV-Nano-Build 1 commit. Zero pushes: XSCache, llvm-project, cva6, core-v-verif, revyos (revyos, buildroot), xslim, onnxruntime, spacemit-com’s linux and u-boot, opensbi, humanoid_unitree_g1, manipulation, manifest, sophgo’s edk2-platforms and build, NanoKVM, radxa-linkr (linkr, plugins), esp-idf, RuyiSDK (ruyi, packages-index), riscv-software-src (opensbi, riscv-isa-sim), XUANTIE-RV (opensbi, zero_stage_boot, linux), Nuclei-Software, eswincomputing. The GitHub anonymous API quota was exhausted during the window this issue (60 requests/hour, with the egress address resetting at 06:31), so organization-level pushed_at could not be enumerated; instead, each repository’s commits and releases Atom feeds were checked individually, so pushes to non-default branches and non-key repositories may be undercounted. The 20 commits in revyos/linux are the same patch set recorded in the November 18 report, with author dates reaching back to April 3, 2026; this time they landed again at 20:49 under new commit hashes (commit timestamps refreshed after rebase) and are not counted as a separate item |
| openRuyi dual-source | The official news page still shows the two October 19, 2026 posts and the 2026.09 release as the latest, with no new version within the window; the 20 commits in the main repository are recorded per the log |
| riscv100.top inspection | No changes: Upcoming is the same as on November 18 (LineShine), none of the six leaderboards show additions or score changes, and there are no committee additions or removals; noted in the supplementary observations as unchanged from the previous issue |
| RISC-V member website inspection (391 sites) | 129 changed, 192 unchanged, 70 failed to fetch. Sampled changed items are all homepage marketing rotations (NVIDIA’s data center and Vera CPU promotion, the AMD brand page, Axelera AI, Espressif, IBM, Huawei, etc.), with no new products, funding, or architecture announcements, so they are not included |
| Exclusion checks | Sipeed’s LicheeRV-Nano-Build MAC conflict fix and Sophgo’s removal of the multi_video_qt subproject are both maintenance changes, included as short items without in-depth analysis; esp-idf had no pushes within the window, and there are no new chips this issue; Radxa Linkr had no commits or releases within the window |
Full Source List
- [1] Hunan Provincial Government Portal: Hunan-made RISC-V automotive-grade chip enters mass production and hits the road — https://hunan.gov.cn/hnszf/hnyw/sy/hnyw1/202611/t20261118_34104125.html
- [2] Hunan Today: This company relocated to Changsha after the previous World Computing Conference; its RISC-V automotive-grade chip enters mass production and hits the road this year — https://www.hunantoday.cn/news/xhn/202611/34281823.html
- [3] Rednet: World’s first to enter mass production and hit the road; Hunan-made automotive-grade chip lays a solid foundation for self-reliance — https://hn.rednet.cn/content/646055/51/16480375.html
- [4] China News Service Hunan: Hunan-made RISC-V automotive-grade chip enters mass production and hits the road — https://www.hn.chinanews.com.cn/news/cjxx/2026/1118/536557.html
- [5] Sina Finance repost of Rednet: World’s first to enter mass production and hit the road; Hunan-made automotive-grade chip lays a solid foundation for self-reliance — https://finance.sina.com.cn/wm/2026-11-17/doc-inizaitt1565830.shtml
- [6] Cercis Semiconductor official website: Cercis M100 launch announcement — http://www.cercis-semi.com.cn/index.php/2025/12/19/%e7%b4%ab%e8%8d%86m100-%e5%85%a8%e7%90%83%e9%a6%96%e6%ac%be%e9%87%8f%e4%ba%a7%e4%b8%8a%e8%bd%a6%e7%9a%84risc-v%e8%bd%a6%e8%a7%84%e7%ba%a7%e8%8a%af%e7%89%87%e6%ad%a3%e5%bc%8f%e5%8f%91%e5%b8%83/
- [7] Hunan Provincial Government Portal: 2026 World Computing Conference press conference (November 19–21) — https://www.hunan.gov.cn/hnszf/hdjl/xwfbhhd/wqhg/202611/t20261112_34100273.html
- [8] OpenXiangShan/XiangShan (kunminghu-v3 branch) — https://github.com/OpenXiangShan/XiangShan/tree/kunminghu-v3
- [9] OpenXiangShan/NEMU v2026.11.r9 — https://github.com/OpenXiangShan/NEMU/releases/tag/v2026.11.r9
- [10] OpenXiangShan/difftest — https://github.com/OpenXiangShan/difftest
- [11] OpenXiangShan/GEM5 (xs-dev branch) — https://github.com/OpenXiangShan/GEM5/tree/xs-dev
- [12] cvw PR #2120: AHB compliance — https://github.com/openhwfoundation/cvw/pull/2120
- [13] cvw PR #2128: ACT granule reset to zero — https://github.com/openhwfoundation/cvw/pull/2128
- [14] openRuyi-Project/openRuyi — https://github.com/openRuyi-Project/openRuyi
- [15] openRuyi official news page — https://openruyi.cn/zh-Hans/news/
- [16] Jimu News: Chip serves as the robot’s “cerebellum”; domestic RISC-V chip showcased at the 5G+ Industrial Internet Conference — https://www.ctdsb.net/c1666_202611/2908179.html
- [17] Phoenix News repost of the same article — https://news.ifeng.com/c/8yA3n8ba6VY
- [18] spacemit-com/llama.cpp PR #71: gated delta net on K1 — https://github.com/spacemit-com/llama.cpp/pull/71
- [19] spacemit-com/llama.cpp PR #55: parallel repack — https://github.com/spacemit-com/llama.cpp/pull/55
- [20] humanoid_common PR #34: read-only joint and IMU feedback — https://github.com/spacemit-robotics/humanoid_common/pull/34
- [21] realtime-link PR #3: video transmission module — https://github.com/spacemit-robotics/realtime-link/pull/3
- [22] sophgo/sophon-tools PR #249 — https://github.com/sophgo/sophon-tools/pull/249
- [23] sipeed/LicheeRV-Nano-Build — https://github.com/sipeed/LicheeRV-Nano-Build
- [24] riscv100.top leaderboards — https://riscv100.top/
Limitations: This issue covers a 24-hour window (05:00 on November 18 to 05:00 on November 19). There are only two substantive news items on the media side; the main content comes from GitHub commit logs, release notes, and on-site exhibition reporting. Technical details are based on commit titles, release notes, and repository documentation, without line-by-line code verification. Cercis’s 10 million units is a cumulative figure for automotive-grade chips, not broken down by model or year. The RMB 30 million figure is a year-end order expectation, not confirmed revenue. The customer list comes from media reports, not company announcements. The 11.9% localization rate is overall automotive-grade chip data for Q1 2026. “World’s first RISC-V automotive-grade chip to enter mass production and hit the road” is the vendor’s and the Beijing Institute of Open Source Chip’s characterization, faithfully relayed here without independent verification. SpacemiT’s K3 and V100 figures are vendor statements made at the exhibition; RVA23 compliance and 60TOPS have no third-party benchmark corroboration. The two spacemit-com/llama.cpp commits are on the mtmd-backend branch, and whether they will land in the mainline is undetermined. The actual performance gain from the gated delta net adaptation is not given in the commits. The GitHub anonymous API quota was exhausted during the collection period this issue, so organization-level push statistics were replaced with per-repository Atom feeds, and non-default branches and non-key repositories may be undercounted. The 20 revyos/linux commits are treated as the same patch set landing again; whether they carry new content cannot be determined from commit titles. The 129 “content changes” in the member website inspection are determined by page text hash. The sampled items are marketing updates, but individual announcements may still be misclassified.