RISC-V Weekly Report (2026-11-02 ~ 2026-11-06)
Reporting period: Monday, November 2, 2026 through Friday, November 6, 2026 — 5 working days, 5 daily-brief source sets. The 11-02 issue used a 72-hour window (covering 10-30 05:00 to 11-02 05:00); the other four used 24-hour windows. The body therefore contains a few items dated 10-30 to 11-01, with the original date noted in each item’s title.
Sources: The daily RISC-V Global Briefs (five issues: 11-02, 11-03, 11-04, 11-05, 11-06); plus a 168-hour window re-check of followed GitHub organizations and key repositories (anonymous API rate limits were exhausted for most of the week, so each repo’s commits.atom was checked individually), dual-source verification of openRuyi (main-repo commits and the official news page), a day-by-day archived summary from riscv100.top, and supplementary checks of RISC-V International member company websites.
Editorial note: This week’s throughline is “the layer that turns an open instruction set into deliverable products.” In automotive, four layers advanced simultaneously within one week: Infineon moved its flagship AURIX product line to RISC-V (silicon); Quintauris released EM-Altair, a binary-compatible execution environment (platform specification); C*Core completed AUTOSAR and IAR adaptation (domestic software stack); and Zijing M100 was selected as a “China Chip” ecosystem promotion demonstration case (a large-scale adoption example). On the AI matrix side, SpacemiT open-sourced the RVISA virtual ISA working draft, attempting to insert a “programmer-expressible” mapping contract between PTX and cuTile. On the distribution side, openRuyi set the default ISA level for riscv64-openruyi-linux to RVA23U64, making RVA23 the default baseline rather than an option. All three threads answer the same question: which layer absorbs the fragmentation that comes with an open instruction set.
I. Weekly Highlights
-
Infineon launches its first RISC-V automotive-grade MCU platform, AURIX RC1, with up to 12 cores, ASIL D, and samples in H1 2027 (11-04): This is Infineon’s first MCU platform based on RISC-V and the first time the AURIX product line adopts an open instruction set. Officially positioned as a complement to, not a replacement for, the existing portfolio, current TriCore parts remain in supply. On specifications, a single chip carries up to 12 32-bit processing cores, a parallel processing unit for edge AI, and up to 40MB of embedded memory, with plans for same-family variants adding 64-bit microprocessor capability; on the safety side it reaches ASIL D, information security meets ISO/SAE 21434 and the Cyber Resilience Act, and post-quantum cryptography support is reserved; on communications it supports CAN XL, 10BASE-T1S and multi-gigabit Ethernet, with some variants integrating BRIGHTLANE switching and hardware data routing. The key to the product logic is merging the previously separate AURIX and TRAVEO T2G lines into a single generation, using one RISC-V base to cover zone control, body, chassis, battery management, powertrain, ADAS safety and cockpit, backed by the DRIVECORE software stack providing AUTOSAR, virtual prototypes and AI-assisted migration tools. Timeline: the family concept will be shown at electronica 2026 in Munich from November 10 to 13 (Hall C3, Booth 502), with samples targeted for H1 2027. Previously, automotive-grade RISC-V was mostly driven by new entrants and domestic vendors; a first-tier vendor switching its mainline products over means the AUTOSAR stack, functional safety certification, toolchains and Tier 1 adaptation resources will follow this line. Source: Infineon press release infpr202611-026
-
Quintauris releases EM-Altair, codifying binary compatibility for MCU-class RISC-V as four layers of compliance requirements (released 11-03, reported 11-05): It does not define a new core but a set of hardware, firmware and runtime requirements that “a compliant implementation must exhibit consistently”: four layers — ISA profile, privileged architecture, boot and runtime services, and security services — where vendors can each implement their own microarchitecture while software faces a single target. The baseline is RV32IM, with a clearly bounded set of mandatory and optional extensions fixing interrupt, atomic operation, CSR behavior and memory ordering; on the interrupt side it specifies a bounded high-throughput architecture for hard real-time, isolation uses PMP covering machine mode and optional user mode, and security practice aligns with ASIL-B equivalent requirements. The accompanying QNTAF verification and certification framework supports cloud, simulator and FPGA targets for benchmark, conformance and security testing, and integrates with CI/CD to generate traceable artifacts. Quintauris was formed in 2023 as a joint venture by Bosch, Infineon, Nordic, NXP, Qualcomm and STMicroelectronics. Read alongside Infineon’s RC1, chip vendors do the implementation and the joint venture does consistency, so RTOSes and compilers need to adapt to only one target once — domestic teams face “lower barriers for base software, higher barriers for compliance” happening at the same time. Sources: Quintauris release, New Electronics report
-
DAMO XuanTie C950 wins the “China Chip” Annual Major Innovation Breakthrough Product Award, and ten customer chips carrying XuanTie cores win awards on the same list (conference 11-02 / reported 11-03): On November 2, at the results ceremony of the 21st “China Chip” outstanding product selection held in Hengqin, Zhuhai, Alibaba DAMO Academy’s XuanTie C950 took the highest-threshold “Annual Major Innovation Breakthrough Product Award”; at the previous edition XuanTie received “Outstanding Supporting Service Enterprise” and “RISC-V Ecosystem Promotion Demonstration Case.” Moving from the ecosystem-promotion category to the product-innovation category says something in itself about this IP vendor’s position in the industry. Parameters are unchanged from the March launch: 8-instruction decode, 16-stage pipeline, out-of-order window of over 1000 instructions, up to 3.2GHz, and SPECint2006 single-core breaking 70 points for the first time; what is new this time is deployment data — cumulative shipments across the series exceed 6 billion units, with over 400 customers and over 1000 licenses, and C950 supports all mandatory and optional extensions of RVA23.1 with native CoVE confidential computing. More noteworthy than the main award is the customer list on the same roster: Dongfang Suanxin DF1000 (C908 plus C906) also won the Major Innovation Breakthrough Product award; Allwinner T536 (E907), Jiangyuan D1 (C908 plus E907), Wuqi WQ9301 (E907) and Hongsi HSRU3 (E906) won Outstanding Technological Innovation Product; Qingwei TX81 (C908) and Taixin TXW82X won Outstanding Market Performance Product; Jiangxinchuang D12X (C906), Jiuchen TD1601 (E906) and Maitexin MEGA188 (E902) won the Xinhuo Rising Star Product award. The ten chips span six markets — near-memory computing, industrial control, cloud inference, wireless connectivity, reconfigurable AI and display drivers — yet use four tiers of cores from the same IP family, which says more about the product line’s coverage density than a single flagship winning an award. Source: Jiwei report (reposted by Sohu)
-
RISC-V embodied intelligence open solution RV-EAI released at the Wuzhen “Light of the Internet” expo, open-sourced by six organizations (11-04): The solution was released by the openKylin Embodied Intelligence Working Group together with Milk-V, Chaorui Technology, Shanghai OpenLoong, SpacemiT and Kuya Technology — six organizations in total — under the working codename “Qiyuan,” with all details evolving continuously as open-source projects in GitHub repositories and on the official website. The proposition is to “release interfaces, manifests and reference implementations rather than a fixed complete machine,” splitting a robot by real-time requirements into a body layer (microsecond-level current loop), a cerebellum layer (millisecond-level control loop, with a real-time core running an RTOS and connecting to the joint bus via CAN-FD and EtherCAT), and a cerebrum layer (ten-millisecond-level inference handling perception, planning and VLA), all three built on the same RISC-V instruction set (RVA23, with RVV, IME, AME and VME). The verifiable definition of “open” is five pluggable points: on the body side, joint drive and MCU offerings from Zhongke Benyuan, HPMicro, GigaDevice, WCH and Espressif; on the cerebellum side, K3/SoM260, Zhihe A210 and UR-DP1000/Titan; on the cerebrum compute side, K3, Axera AX650N and Rockchip RK1828; at the operating system layer, five distributions — openKylin, Debian, Ubuntu, Fedora and openEuler — plus the RT-Thread real-time domain; at the framework and model layer, FlagOS, llama.cpp, TileLang and vendor runtimes. The acceptance criterion is stated explicitly: swap out one point and the other four need no changes, nor does the upper layer need re-adaptation, and it must pass conformance tests. The site also publishes self-tested data on a real SpacemiT K3 machine (Bianbu 4.0.7, kernel 6.18.3, collected October 21): local inference of a 1B vision-language model at 24.09 t/s, a 24B MoE model at 12.00 t/s, and IME2 at 970 GMAC/s per core (5632 GMAC/s across 8 threads), with a clear distinction between vendor-claimed peaks and usable throughput. The weight of this material lies not in the launch event but in how many real machine models the compatibility matrix ultimately fills in. Sources: Jiachen Plan WeChat account, RV-EAI website, kubuds/RV-EAI repository
-
SpacemiT open-sources the RVISA virtual ISA specification working draft, with the first round adding IME/VME/AME mappings and blocked MMA (repo created 11-04, commits 11-05): The new repository
spacemit-com/rvisaproposes a RISC-V Virtual ISA, defining a layer of virtual instruction notation for scalar, vector and matrix computation; the current specification version is 0.2.0. Its positioning is stated clearly: it deliberately sits above NVIDIA PTX and below cuTile — it binds to no front-end language (source-language front-end implementations are explicitly out of scope for this version) and does not directly specify any chip’s instruction encoding, instead making “target mappings” versioned profiles that declare item by item which register classes, register groups and instruction sequences they map to, with upstream revision numbers and verification evidence attached. The notation follows RVV Intrinsics in expressing typed operations, operand roles and register-group constraints; tiles support one to N dimensions, reshape and axis permutation must preserve element positions, and layouts are derived from linear layouts. What best explains the problem it aims to solve is the “expressibility requirement”: any optimization of an RVISA tile computation must be writable by the programmer at the notation level — targeting precisely the failure mode where, as matrix extensions land, compilers silently vectorize and code must be rewritten for every new chip. Maturity needs to be stated plainly: the specification status file itself says this is a working draft, not a RISC-V International-approved specification; target mapping profiles are registered only as drafts with no implementation conformance evidence yet established; the MLIR dialect and executable conformance tools are both marked TBD; and the reference workload exists only as.rvisanotation source form and cannot be compiled or executed. The repository also contains three draft RFCs, 0001 through 0003, covering the instruction and register mapping contract, the core type and operation contract, and programmer-expressible tile layouts. It is the same logic as EM-Altair on the MCU side in two manifestations: fragmentation must ultimately converge on some stable layer of “software target,” with the difference in governance structure — one led by a joint venture of six major vendors, the other a single vendor’s working draft. Sources: spacemit-com/rvisa repository, specification status page -
Linux 7.4 turns “fast unaligned access” into a runtime probe, letting distribution kernels gain up to 11% bandwidth (10-30): This change came in two steps. The 7.3 change submitted by Tenstorrent let a “kernel compiled for RISC-V hardware supporting fast unaligned memory copies” skip software alignment, with self-tested bandwidth gains of up to 11%, but it required enabling at compile time — and a distribution kernel must serve a range of hardware, always including chips with slow unaligned access, so the benefit could only remain in custom kernels. 7.4 moves the determination onto
HAVE_EFFICIENT_UNALIGNED_ACCESSas a runtime probe: only when the hardware is probed to actually support it do user-space copies skip software alignment, so generic builds and distribution kernels can carry it by default. Other RISC-V changes merged into 7.4 in the same batch include crash hotplug, some page-table functions taking optimized paths, context-switch optimizations, the B extension reported in hwcap and hwprobe, and Zic64b reporting support. Together these are a step toward “the kernel ascertains platform capabilities and schedules on them itself,” and reporting on the hwprobe side is precisely the prerequisite for optimizations like runtime probing. One judgment from community discussion is worth recording: almost every vendor except SiFive supports fast unaligned access, and SiFive supports it only on its newer-generation performance cores — so once distributions enable this optimization, the actual benefit depends on whether chips report it truthfully in hwprobe. Sources: Phoronix, Hacker News discussion -
openRuyi sees about 58 main-repo commits over five working days, RVA23U64 becomes the default baseline, and RuyiAI components enter the distribution (10-30 to 11-06): The daily pace was 6, 7, 15, 10 and 20 commits, plus 4 more during Friday daytime and commits to supporting repos such as clang-wrap. Three threads are most worth noting: first, llvm23 23.1.4 was added (a 1220-line spec plus two patches), adding an
openruyivendor to LLVM and setting the default ISA level for theriscv64-openruyi-linuxtriple directly to RVA23U64 — without passing an extra-march, the compiler generates code for RVA23, making RVA23 the default baseline rather than an optional optimization tier. Second, on 11-03 python-tilelang-ruyiai, python-triton-ruyiai and buddy-mlir were brought into the distribution as SPECs, meaning this RISC-V AI software stack is starting to be distributed as installable packages rather than via source-build scripts; kernel debuginfo support was completed the same day, making stack traces for online performance issues usable for the first time. Third, on 11-05 a single timestamp pushed in 20 document-format import libraries (libvisio, libwpd, libwps, libzmf, libabw, libcdr, libetonyek, libmspub, libodfgen, liborcus, libstaroffice and others), all from the same family of libraries that LibreOffice’s format import filters depend on — closing a dependency closure rather than adding scattered packages. The rest are engineering updates: Firefox 159, an expat 2.10.0 security fix, ollama 0.40.7, and CI switching to a scoped OBS signing key. The official news page’s latest items are still the two from October 2026 (10-19), and the November release has not shipped, so this week there is only repository-side progress and no distribution release. Sources: openRuyi main repository, openRuyi official news page, llvm23 spec -
Andes slows its high-end RISC-V CPU Cuzco, redirecting engineering resources to an AI business that already accounts for half of revenue (11-02): In a November 2 Taipei report by EE Times, Andes CTO Charlie Su confirmed the company has slowed development of its high-end core Cuzco. Cuzco was originally a project of Condor, the company spun out for Andes’s data-center-class high-end core, publicly disclosed at Hot Chips 2025 and brought back in-house earlier this year; it is not cancelled, will still become a product and may be renamed, but the timeline has been stretched. On the numbers, AI now accounts for half of Andes’s revenue, having grown 30% last year with 50% growth expected this year; its in-house deep learning accelerator is modest in scale (about 4 TOPS at 1GHz) and comes with its own compiler, and the bulk of AI revenue comes from AI chip companies using Andes CPU cores for control, orchestration and data movement, with Meta, d-Matrix, Fractile and Axelera named among customers. Near-term priorities are controller cores for the storage market (10 customers already in volume production with small controller cores), automotive and robotics. This runs counter to the narrative that “RISC-V is charging into high-end data center” but is not contradictory — high-end core development has not stopped; the commercial cadence is now set by market momentum rather than the roadmap, and what actually pays is AI edge and control planes. Source: EE Times
II. Instruction Set and Architecture Ecosystem
XiangShan: 18 main-repo commits in one week, the heaviest being a CBO decode bug
The XiangShan main repo (kunminghu-v3) merged 18 commits this week, plus GEM5 6, NEMU 4, and difftest 11 (including nine from 10-30 to 11-01). The direction is highly consistent: finding bugs and adding observability, not adding features.
- CBO opcode mistakenly OR-combined into cbo.zero (11-03, #6754): In cache block operation instructions, the operation type should select the complete opcode as a whole, but the code performed a bitwise OR of the “full CBO opcode” with the “base opcode.” When a CBIE request converts inval into flush, the OR operation combined cbo.inval and cbo.flush into cbo.zero—changing the semantics from “invalidate this block” to “zero this block.” Inval merely discards; zero writes data. A data-bearing zero landing on a line in use by another hart or about to be written back is silent data corruption, and it triggers only under specific operation sequences—undetectable by benchmarks or routine regression.
- Four memory-access fixes concentrated troubleshooting (11-03): #6762 makes flush wait for CBO zero to finish enqueueing first; #6718 fixes the forwarding mask for unaligned accesses crossing 16 bytes; #6714 fixes the MdpHit determination when waitSqidx has already retired; #6678 fixes LoadUnit rollback info being overwritten by the next instruction under the same RobIdx. All four share the trait of failing only under specific interleavings: unaligned access combined with forwarding, retirement and hit determination in the same cycle, two consecutive loads with the same RobIdx triggering replay.
- TAGE stores alt info into meta (11-03, #6800): Previously training reads and prediction reads contended for the same storage; the change stores alt information additionally into meta, and the training side issues a training read only when TAGE mispredicts or the actual branch outcome differs from the prediction—taking a high-frequency read path off the critical path, the most typical tradeoff in a branch predictor.
- Frontend and prefetch: The frontend introduces FtqPrefetchQueue (10-30), the first step of the iprefetch refactor; in the same period, prefetch.i/r/w are decoded into LSU prefetch operations, and MBTB allocation suppression for hidden VBTB branches is fixed. On 11-05, another fix addresses MissQueue still merging stores while a prefetch miss is speculative (#6842)—merging a store with an undetermined prefetch means that if the prefetch is cancelled, the writeback data’s ownership goes wrong.
- Observability and engineering process: Added BPU performance counters related to s2_override (#6777), and decode signals for former/latter remaining writeback counts in CROB (#6749); refreshed the README and added an AGENTS.md at the repo root aimed at AI coding agents (#6849), specifying build methods, code style, and commit conventions—an RTL project actively maintaining such a file indicates agents are already assumed participants in the daily development workflow.
- Verification chain: NEMU fixes vscause retaining the injected interrupt number when hvictl injects an interrupt (#1279), marks zero-length vector loads as dirty (#1278), and routes matrix accesses through the shared TLB optimization path (#1275); difftest splits matrix AMU events into “writeback” and “completion” categories (#1025), and preserves asynchronous AME checker state during squash replay (#987)—long-latency instructions like MMA inherently have completion order differing from writeback order, and expressing two things with one event必然 misjudges in multi-core or replay scenarios. Among the nine difftest commits from 10-30 to 11-01, the MMA backend packing FP8 products into cache words is a step on the precision path.
- All six GEM5 commits concern testing methodology: google trace and cbp trace wired into CI, Fetch Block Size statistics PMU added, SPEC06 RVV checkpoint profile updated, optional switch added to the weekly performance task, a hardware prefetch algorithm survey document added, and cpu-o3 filtering resolve inputs by “oldest same-cycle squash.”
- Sources: XiangShan, #6754, NEMU, difftest #1025, GEM5
CORE-V: 29 commits in three batches, half on “installable, buildable, benchmark-runnable”
CORE-V Wally (cvw) merged 29 commits in three batches this week (three from 11-02 to 11-03, fifteen from 11-04 to 11-05, eleven from 11-05 to 11-06), plus one cvfpu commit. Structurally they fall into three groups:
- Spec details: #1974 stops caching invalid PTEs per Svvptc (an RVA23 optional extension)—previously HPTW would write walk results landing on V=0 leaf or non-leaf PTEs into ITLB/DTLB, whereas Svvptc allows the handler to set V to 1 and retry directly without issuing SFENCE.VMA; an implementation caching invalid PTEs would fault permanently. The fix routes the walker through the existing FAULT path and writes nothing to the TLB. This is a textbook case of “the spec gives you an optional extension, and the implementation must leave a path for its legal use.”
- Microarchitecture progress: #2050 changes the register file to a latch-based implementation (trading area and power for timing), #2046 routes HPTW AMO accesses through bus atomic transactions, #2052 fixes eBuFsmArb last-beat handling, #2048 introduces synthesis auto-ungroup, #2045 advances the RV32IMCB configuration into the regression flow.
- Delivery engineering quality: Fixed the custom install script under Ubuntu 26.04 (#2061), rebuilt buildroot (#2039), adjusted Embench speed test and CoreMark compile options (#2031, #2024), changed code coverage to per-configuration statistics and brought rv64gc to 100%, and fixed stale values in wallyTracer’s RVVI trace fields (pte/page type/ppn/virtual and physical addresses and access type often carried values from the trap handler’s code page on TLB hits, causing ACT functional coverage to be recorded incorrectly).
- Assessment: CVW is a teaching- and research-oriented core, but it has recently treated Ubuntu 26.04 compatibility, buildroot rebuilds, and CoreMark/Embench compile options as serious work—half its commits land on “installable, buildable, benchmark-runnable.” That cadence is closer to the engineering quality of tapeout preparation and external delivery than to RTL meant only for papers.
- Sources: cvw commit log, CVW #1974
Secure microarchitecture: SecDMP delivers the first open-source DMP implementation on a RISC-V core
A team including KU Leuven (NDSS 2027 submission) first implemented the first open-source DMP (Data Memory-dependent Prefetcher) on a RISC-V processor, empirically demonstrating that the only existing defense aiming to preserve DMP performance cannot stop the entire class of leakage, then proposed SecDMP: it requires marking secrets in memory, ensures the DMP never encodes secrets into microarchitectural state, and formally proves its leakage does not exceed a “baseline without DMP,” all without sacrificing performance on public data; the hardware cost is no more than 0.5% area on the research core. The value to the open-source core community is that optimizations like DMP are moving from big-vendor proprietary designs to configurable cores, so security design must enter the RTL early rather than being patched after tapeout. In the same week, a joint team from Ruhr University Bochum, the Max Planck Institute for Security and Privacy, and Georgia Tech implemented all four randomized caches—CEASER-S, ScatterCache, SassCache, and MIRAGE—as the last-level cache of a single multicore out-of-order RISC-V SoC, built a complete FPGA prototype, and ran the full SPEC CPU 2017 rate and GAPBS workloads; by their account, this is the first horizontal evaluation of these four designs on the same full-system hardware platform. The paper focuses on implementation difficulties and tradeoffs; the public abstract gives no quantitative performance-overhead comparison. Both lines of research point to the same thing: RISC-V microarchitectural innovation has reached the stage where security costs must be quantified alongside it. Sources: arXiv:2610.37454 SecDMP, IndexBox coverage of the randomized-cache paper
Compilers and toolchains: RVA23U64 becomes the distro default, Nuclei goes VS Code, Espressif splits trace decoding into its own repo
- openRuyi llvm23 23.1.4 (10-30): Adds an
openruyivendor to LLVM, sets the default ISA level of theriscv64-openruyi-linuxtriple to RVA23U64, and wires that triple into both the x86_64 and riscv64 Triple handling paths to keep cross-builds working. This is the clearest statement on the toolchain side this week: RVA23 is no longer an optional optimization tier. - Nuclei releases Nuclei Studio VS Code IDE preview (10-30, Shenzhen): The official development tool now comes in both Eclipse and VS Code forms; the current release is a Windows developer preview 0.1.0 with Chinese and English UIs, installed as a VSIX into an existing VS Code (requires 1.134.0 or later), and still relies on Nuclei Studio IDE 2025.10 underneath. The compatibility strategy is “don’t rewrite, copy as much as possible”: Eclipse-version projects can be imported directly, cproject compatibility handling is built in, the new-project wizard is identical, and SDK and BSP are still distributed as NPK packages interchangeable between the two versions; debugging supports OpenOCD, J-Link, QEMU, and NucleiModel, all through CDT gdbtarget, with Live Watch added; on the AI side there is no walled garden—Kimi, GLM, DeepSeek, and Qwen are all supported. The roadmap lists Linux and macOS support, profiling, code coverage, ETrace, and flash programming. For the domestic RISC-V developer ecosystem, the real significance of this step is the migration of the entry point—Eclipse’s slow startup and heaviness have long been complaints, while the VS Code side can conveniently connect to local or cloud models.
- Espressif splits trace decoding into its own repo (11-05/11-06):
esp-riscv-tracepushed 4 commits, adding GitLab CI, pre-commit configuration, updating GitHub workflows, and finally mergingci/pre_commit_and_pipelinesinto master. The repo was created on August 26; at the top level is a Python package namedesp_riscv_trace, which does trace data decoding and tool-side support rather than RTL. Chip trace support is only the first step; being able to decode it into something human-readable is what makes it usable—delivering it as an independent component (dual CI plus pre-commit checks) is how that judgment is implemented. In the same period, Nuclei’s tool guide added rvtrace documentation and Hummingbird debugger troubleshooting, as RISC-V instruction tracing moves from high-end cores and dedicated probes into the standard documentation system of MCU vendors. - Sources: openRuyi llvm23 spec, Design & Reuse release, esp-riscv-trace, nuclei-tool-guide
III. Operating Systems and Distributions
openRuyi: ~58 commits across five periods, moving from “having packages” to “being operable”
Beyond the three main threads described in the first chapter (RVA23U64 default baseline, RuyiAI components entering the repository, office dependency closure), two engineering threads this week are worth noting separately.
- Security fixes are becoming a process: A batch of packages received security-fix-tagged upgrades, including Firefox 159.0, expat 2.10.0, python-aiobotocore 3.10.0, python-nltk 3.10.4, python-paste 3.11.0, python-multidict 7.2.1, python-hydra-core 1.3.8, and others; CI replaced apt-key with a scoped OBS signing key, and rolled back a change that “ran the update workflow using the openruyi container.” Security updates going through the process in tagged batches indicates they have entered a process-driven channel.
- Operability gaps filled: openruyi-linux-build now supports generating debuginfo packages; kernel modules are installed without stripping or compression and handed off to find-debuginfo, while CONFIG_DEBUG_INFO_SPLIT is marked unsupported to ensure debuginfo packages remain valid; simdjson was upgraded to 5.0.8 and switched to running upstream tests offline via Ninja, avoiding network access during builds. Previously kernel modules were stripped, so stack traces could only show addresses with missing symbol names.
- Friday daytime fill-in: After 11-06 05:00, the main repository had another 4 commits (clonezilla, llvm-defaults plus python-clang, tmux 3.8, qt6-qtwebengine build fix), still within the Friday workday and counted in this week. The official news page scrapes normally; the latest announcements remain the two from October 2026 (Ruyi Firmware, RISC-V ecosystem public infrastructure), and the November release has not shipped—this week is rolling maintenance rather than a release sprint.
- Sources: openRuyi main repository commit log, openRuyi official news page
revyos/linux: Twenty config commits merged at once, K3’s platform surface widens noticeably
revyos’s 7.3.y branch merged 20 kernel config commits at once on 11-03, all stamped with the same timestamp—a single concentrated sync. On the hardware side, newly added or enabled: SpacemiT K3’s UFS and display controller, K3’s USB3 PHY and a full group of Type-C/TCPM/UCSI/FUSB301/DP alternate mode, ESWIN EIC7700’s ARCH plus DWMAC Ethernet and pinctrl, AST display, LM63 temperature sensor, and Banana Pi F3’s K1 USB2. On the platform capability side, two are more critical: a new linuxboot config (for deployment scenarios that load the kernel as firmware), and two CI pipelines—cross and native builds—added to solve the long-standing problem that kernel config changes could not be validated at commit time; also a PicoHeart errata switch and ONIE TLV NVMEM layout enabled (an asset-tag format for batch deployment of switches and network devices). Assessment: this distribution is moving from “the board boots” toward “complete system peripherals and verifiable changes.” Source: revyos/linux commit log (7.3.y)
OpenBSD 8.0 and felix86: hwcap, AIA, Sstc, and the long tail of the x86 compatibility layer
- OpenBSD 8.0 (released 10-20, reported 11-03): This release’s riscv64 increments concentrate on architectural features rather than drivers—the kernel now probes CPU extensions and reports them to userland via AT_HWCAP/AT_HWCAP2 in
elf_aux_info, so userland programs can dispatch code paths by hardware capability; the kernel adds V vector extension support, implements bounce buffers, supports the RISC-V advanced interrupt architecture with two new drivers, imsic and aplic, and adds Sstc timer extension support; the same release retires the Loongson port. For RISC-V distributions, hwcap reporting is more fundamental infrastructure than any single optimization—without it, userland can only guess branches at compile time. - felix86 fourteen commits (11-04/11-05): The recompiler that brings x86/x86-64 games to RISC-V Linux released version 26.11 on November 2; this window is the post-release fix wave, focused on Linux compatibility layer correctness: signal delivery for PTRACE_SYSCALL, PTRACE_SETSIGMASK restoration, 32-bit setrlimit, returning EINVAL when statfs64 length is invalid, the tid pointer for fork-cloned child threads, root directory escape checks for creat and statfs, the extended mask for 32-bit signal frames, and an infinite loop on slash-only paths; on the recompiler side, fixing the FSW C1 bit on x87 stack exceptions, moving the bad-address thunk before the code cache, and implementing TSO optimization for Volatile Metadata. Dense commits of this kind usually mean the main flow already runs, and what remains is making games and runtimes stop crashing intermittently.
- Sources: OpenBSD 8.0 release notes, felix86 commit log
IV. Chips and Hardware
Automotive-grade: Within one week, chips, platform specifications, software stacks, and ecosystem cases advanced simultaneously across four layers
Infineon AURIX RC1 (11-04) and Quintauris EM-Altair (11-03) were already covered in the first chapter; the other two layers are added here:
- ChipON CCRC4XXX completes AUTOSAR and IAR adaptation (reported 11-04): The automotive-grade AI MCU based on a RISC-V multi-core architecture has completed adaptation with Neusoft Reach’s NeuSAR CP platform and IAR Embedded Workbench for RISC-V, with the three parties showcasing a joint demo at the 6th RISC-V China Summit. The chip is China’s first RISC-V multi-core post-quantum-resistant automotive-grade AI MCU, built on a 22nm RRAM process, integrating a 0.3 TOPS NPU subsystem, with functional safety up to ASIL D and information security up to Evita-Full, compatible with both post-quantum and conventional cryptographic algorithms and supporting NIST FIPS 203/204, relying on CGDMA for in-vehicle network data routing. Two parts have already been sampled: CCRC4088S with 8+4 cores, 500MHz, 14000DMIPS, targeting central domain control and intelligent driving; CCRC4046S with 4+4 cores, 500MHz, 7000DMIPS, targeting body, BMS, and zonal domain control. NeuSAR cCore complies with AUTOSAR R21-11/R23-11 and has obtained ASIL D certification, and the functional safety edition of the IAR toolchain is certified by TÜV SÜD.
- Zijing M100 selected as a “China Chip” RISC-V ecosystem promotion demonstration case (reported 11-05): The automotive-grade RISC-V MCU from Hunan Zijing Semiconductor was selected for the newly established “RISC-V Ecosystem Promotion Demonstration Case” category, which focuses on selecting projects with demonstration value in technology deployment, ecosystem adaptation, upstream-downstream joint promotion, and large-scale adoption. The M100 is built on China’s first automotive-grade MCU designed on an open-source RISC-V core, developed by a team incubated at Great Wall Motor’s Technology Center, entering mass production and shipping in vehicles on December 18, 2025, used in body control scenarios such as combination headlights, ambient lighting, and combination switches; as of October 9 this year, cumulative shipments exceeded 10 million units; the company states it has developed more than 15 core IPs in-house, with a complete MCAL, low-level driver SDK, RTOS, and development/debugging environment, and has passed ISO 26262 ASIL-D process certification, ASIL-B product certification, ASPICE, and AEC-Q100. The product line extends to M100/M200/M300 and companion power driver chips, with M200 planned for tape-out in Q4 this year (40nm), targeting combination instrument clusters, T-BOX, gateways, and BCM; the E100 cerebellum control chip for embodied intelligence is in development. The highlight is not the chip specifications but the award criteria: listing “ecosystem promotion” as a separate category means the review process has begun to value “upstream-downstream joint adaptation and large-scale adoption.”
- Sources: Sohu Tech (ChipON), Changsha Evening News (Sohu reprint)
IP and security chips: Andes slows down voluntarily, Viettel delivers its first self-designed security chip
- Vietnam’s Viettel releases its first self-designed RISC-V security chip VGC v1.0 (released 11-03, reported 11-04): The Vietnamese Government Cipher Committee and military electronics and telecommunications group Viettel unveiled the country’s first self-designed security semiconductor platform and first chip in Hanoi. The chip uses the RISC-V architecture and a 28nm CMOS process, packaged in a 10mm square, integrating approximately 300 million transistors, hardwiring Vietnam’s self-developed encryption algorithms into hardware, primarily targeting government and defense; official tests claim encryption/decryption speeds approximately 15 times that of pure software solutions for equivalent tasks. The chip has been used in secure USB storage devices for government and defense agencies, with the next step being entry into the national digital signature certification system (over 2 million digital certificates already issued); the two parties also signed a cooperation agreement to jointly develop a multi-core security chip platform integrating dedicated cryptographic components, with long-term expansion into AI, 6G, smart cameras, drones, and robots. In related news, Reuters reported that Viettel’s first wafer fab, with a $1.3 billion investment, may not be able to proceed with its 2027 trial production target.
- The other side of the IP layer: Andes slowing down Cuzco (see first chapter) and onsemi acquiring Codasip’s low-end RISC-V processor business for $14.5 million (11-03 supplement with price and personnel details, involving 27 of approximately 260 employees, about 90% of the price already paid, with Studio tool licensing attached) point to the same conclusion—general-purpose core licensing is becoming increasingly difficult amid oversupply and price wars; those who survive are either buyers who use cores in their own chips or sellers who have found a differentiated security narrative (Codasip pivoting to CHERI). The price, headcount, and the judgment on Codasip’s pivot to CHERI for this transaction were already covered in the previous weekly report (10-26 to 10-30); this week only adds the “transaction completed” statement and does not re-count it as a headline item.
- Sources: Guancha (Viettel), Jon Peddie Research (Codasip)
Edge devices and dev boards: A $21 dual-chip board and a launcher that treats an MCU like a PC
- Wireless Tag WT99P461-S1 dev board (11-04): Based on a 25×25mm module, with the main controller being ESP32-P4X (dual-core RISC-V at 360/400MHz, with AI instruction extensions and single-precision FPU), plus a low-power RISC-V core at up to 40MHz, with 32MB PSRAM and 16MB NOR; wireless connectivity is provided by ESP32-C61 with 2.4GHz Wi-Fi 6 and Bluetooth 5.4. The board features 100M Ethernet, RS-485 and CAN terminals, MIPI DSI/CSI, microphone and 3W speaker interfaces, microSD, and dual USB OTG, preloaded with a phone-like interface based on ESP-Brookesia and AI vision examples. The board is $21, and a kit with a 7-inch screen is $42.50.
- P4 Launcher (11-05): Developer Gil Tal released an open-source application platform for ESP32-P4 dev boards with touchscreens, booting directly into its own interface, with applications stored as
.pappfiles on the microSD card and loaded into PSRAM at startup; adding or removing apps only requires copying files to the card without reflashing firmware. There are currently 21 PAPP applications, mainly emulators and games (Neo Geo, Atari 2600, Doom, Quake, Duke Nukem 3D, OpenTyrian), plus a clock and file manager; the interface is based on LVGL, the audio engine supports up to 8 channels at 4 to 192kHz, supports up to four input devices for multiplayer gaming, firmware can be updated from microSD or GitHub with automatic rollback if a new version fails to boot; an SDK and two compatibility layers are also provided (reimplementing the odroid_* interface and implementing the legacy PAPP v1 service table). The value of such projects is not in “being able to play games” but in running the entire set of conventions for app packaging, loading, updating, and rollback on an MCU. - Nuclei opens dual SDK repos for T0 Chiplet (11-02): Newly created
nuclei-t0-sdk(bare-metal/RTOS) andnuclei-t0-linuxsdk(Linux), both currently empty repos with zero commits. Empty repos carry limited information, but the simultaneous launch of two SDKs both named after T0 Chiplet indicates that the next phase of the product line will treat the chiplet-form T0 as an independent platform for building out the software stack, rather than attaching it to the SDK of an existing core. - Sources: CNX Software (WT99P461-S1), CNX Software (P4 Launcher), nuclei-t0-sdk
The software side of domestic SoCs: Sophgo provides stress-test numbers, SpacemiT provides preprocessing contracts
- Sophgo’s 2-core/3-core variants for Qwen3.5-0.8B (11-05): sophon-demo added two variants with
--num_core 2/3on the CV84X6 platform for the 0.8B model; the author provides 20-round serial stress-test averages: 2-core first token 0.868 seconds, TPS 51.89; 3-core first token 0.681 seconds, TPS 55.20. Going from 2 to 3 cores reduces first-token time by 21.5% while throughput increases only 6.4%—the speedup ratio is far below the core-count ratio, indicating the bottleneck is mainly not in matrix compute but in weight movement and scheduling; for those deploying small models on edge devices, this directly determines whether to add cores or change the data layout. In the same week, Sophgo also released the 2026.11.03 weekly build of cvi_rtt, disabled PCIe ECRC checking by default in SG2044 EDK2, and hardened I2C broadcast and bus recovery in the mcu repo. - SpacemiT model-zoo-vision tightens result contracts (11-03/11-04): #113 changes failed result responses to direct rejection rather than continuing with an error, allows “successful empty results” as a legitimate state, unifies output vector reads to finite-value checks, tightens the Python-side type extractor, and changes embedding normalization; #115 unifies image scaling geometry calculations in one place to avoid result deviations from different models implementing their own; #114 changes instance mask cropping from truncation to rounding up to half-open interval pixel boundaries, preventing an edge row or column from being clipped. Such changes do not appear in accuracy or throughput numbers, but they determine whether upper-layer business logic receives partial data under abnormal inputs.
- Edge agent stack takes shape (10-30): SpacemiT created the
bianbu-agenticumbrella repo and completed bootstrap; it does not vendor code itself but uses submodules to pin a commit for each component, making the full device image reproducible from a single SHA. The components in the repo form a complete edge agent operating system: agent runtime (YAML-driven session event streams, lifecycle, tool and plugin execution, security boundaries), agent-security (tool approval, per-agent synchronized capabilities,securitydsigned manifests), agent-memory (Rust, three-tier drawers of L1 session/L2 working/L3 long-term), bibit-intent (intent routing from deterministic rules to local inference to cloud escalation), BiBit GUI (Tauri desktop frontend based on Jan), bibit-voice (KWS, VAD, ASR), bibit-filesearch (local semantic search, data never leaves the machine), and lxqt-schedule. Making “memory” an independent subsystem with three tiers and making “security” a resident daemon with signed manifests are the kinds of components that only appear when turning an agent from a demo into a deliverable system. - Sources: sophgo/sophon-demo, model-zoo-vision #115, bianbu-agentic
V. Open Source Community and Industry Activities
Robotics Software Stack: Manipulator Framework, Full Dice Pipeline, and Humanoid Navigation — Three Layers Completed in One Week
- Unified manipulator control framework (10-30):
spacemit-robotics/manipulationlanded a single commit of 85 files, +8605 lines, with directories split by capability:calibrationcontains ChArUco calibration, a hand-eye calibration CLI, and sampling tools;diagnosticshas arm_check and an operator CLI;kinematicsprovides URDF-based forward kinematics;motionis the device service and CLI;bindings/pythonprovides the client and operator support layer. The vendor treats the manipulator as a system component with “calibration, diagnostics, kinematics, and a service interface” rather than a snippet of example code. - Device sessions and safety boundaries (11-05): The architectural trade-offs in manipulation are stated explicitly — C++ applications link the component library directly and use DeviceSession in-process; the command line and Python go through DeviceClient, which spawns a local process and dispatches requests over JSONL; YAML waypoint tasks are executed by
manipulation-device run. Device sessions uniformly handle device exclusivity, arrival determination, cancellation, timeouts, and fault locking; the driver layer only handles model protocols and unit conversion, while kinematics and hand-eye calibration load models from configuration — adding a new device just means implementing Arm and EndEffector and registering them. Most noteworthy is the passage stating that “components do not auto-enable, do not auto-clear errors, do not auto-home, do not auto-retry” — once middleware does these four things automatically, on-site personnel lose the basis for judgment when problems arise; placing safety boundaries at the session layer rather than the driver layer is the restraint that robotics middleware ought to have. - dice_arena full pipeline (11-05): SpacemiT K3 serves as the main controller; an AgileX Nero seven-axis manipulator connects via a USB-CAN adapter (SocketCAN, 1Mbps); the end effector is a Revo2 right-hand dexterous hand (six control channels); an Intel RealSense D435i is used for cup grasping and calibration; dice recognition uses a top-mounted USB camera; a ChArUco calibration board, speaker, and green cups round out the setup. On the software side,
dice-arena servestarts video forwarding, local voice, and the game backend; the backend runs a state machine (rule explanation, cup grasping, cup shaking, cup placement, visual judgment, result gesture, return to HOME). Cup grasping is handled by CupApplication, which uses RealSense images to recognize the cup body, solves inverse kinematics, and checks the path; recognition is performed by vision_yolov8_objdetect inferring on the SpacemiT EP, with the feed pushed via RTSP to MediaMTX and then played in the browser via WebRTC; voice is synthesized by tts_moss_nano calling the local MOSS-TTS. What is truly valuable is the failure handling: the task policy library checks stage ordering via C ABI; once an action is issued and fails, it locks immediately; states requiring manual intervention are written totask_state.json, and after a program restart it still refuses to move until on-site personnel resolve the issue and execute reset. - Humanoid and SDK consolidation (11-04/11-05): humanoid_linglong adds LingLong’s SuperMimic standing policy configuration, gripper models, and manipulation scenarios; ros2_humanoid adds two complete configurations of RGBD SLAM and Nav2 for G1; manifest pushed three commits in a row to formally add manipulation and dice_arena to the SDK repository manifest, while docs-ros was updated with the Orbbec camera guide and K3 manipulator operation and dice game guides. The repository manifest determines which components belong to the SDK, the documentation determines whether developers can follow along, and the components determine whether hardware actions can be reused — all three advancing on the same day shows this line is moving from point demos to deliverables with versions, manifests, and manuals.
- Sources: manipulation, dice_arena, humanoid_linglong, manifest
Conferences and Initiatives: HPMicro Ecosystem Conference Scheduled, ERICI Defines Industrial Computers, India Pushes SHAKTI and VEGA Toward Commercialization
- Second HPMicro Ecosystem Conference set for November 21 in Shenzhen (11-04): The second HPMicro Semiconductor Ecosystem Day will be held on November 21 (Saturday) from 9:30 to 17:00 at the Westin Shenzhen Nanshan Yitian Hotel, themed “AI Empowers Chips, Developers Build the Ecosystem Together,” by invitation only. Of the 14 agenda items, only 3 feature HPMicro itself presenting products; the rest are toolchain vendors (SEGGER), open-source RTOS (RT-Thread, presenting HPM6750 hardware-software co-design), manufacturing services (Huaqiu DFM), and community developers. The morning starts with products and toolchains (including “Embracing AI Agentic/Vibe Coding”), while the afternoon shifts to drone flight control and ESCs, AI voice plus vision plus motion control, the cherry suite, PocketJS, AI-assisted SDK development, and a HelixLink deployment case study. This ratio is itself a signal of ecosystem maturity — it turns “ecosystem” from the company’s own SDK documentation into reports of others delivering work on its chips. This aligns with last week’s moves: HPMicro pushed 12 RT-Thread BSP repositories to v1.13.0 in a unified manner within 13 minutes in late October to early November, and opened a new BSP repository for HPM5100EVK, while HPM5100/HPM6E8Y have been listed in RV-EAI’s body layer and cerebellum layer selections.
- Europe’s ERICI initiative aims to turn RVA23 into an “industrial computer” (11-02): The initiator is Germany’s Yuri Zaporozhets, positioned as non-commercial industry collaboration. The problem statement is direct: RVA23 silicon has arrived, but the “industrial computer” has not — board vendors wait for software, software waits for boards, customers wait for both, and this cycle must be broken by an open reference stack running on real hardware. The reference stack is split by layer: firmware performs initialization and POST before handing off to the UEFI bootloader; gRouvr is loaded like a kernel, packages each partition’s kernel and initramfs with CPIO, then allocates CPU, memory, devices, and their respective device trees to real-time partitions and Linux partitions, which run in parallel and isolated on the same SoC; licensing states that HFI BIOS is GPL-2.0 (based on U-Boot), mr-bml is GPL-3.0, and gRouvr and QSOE are Apache 2.0. The partitioned stack has been demonstrated on the SpacemiT K3 PicoITX board. It also honestly lists the gaps: off-the-shelf RVA23 machines (Firefly AIBOX-K3, Milk-V Jupiter2) are mostly “not specified” on wide-voltage input, −40 to 85°C, fieldbus and real-time I/O, DIN-rail mounting, industrial EMC and safety certification, and 10–15 year long-term supply.
- India’s MeitY sets a 24-month homegrown GPU goal while pushing SHAKTI and VEGA commercialization (11-05 report): Under the “India Semiconductor Mission 2.0” framework (approved by the cabinet in July 2026, roughly $15 billion, 12-year period), India plans to produce a domestic GPU chip in about 24 months; directly relevant to RISC-V is another statement — simultaneously pushing the commercialization of the two homegrown processors SHAKTI and VEGA. Verifiable progress includes C-DAC’s VEGA ET301 (32-bit, up to 100MHz, 26KB SRAM), used in the Aries 3.0 development board launched in 2024, positioned for sensing, wearables, and IoT; at SEMICOM India on September 26 this year, the government showcased a SHAKTI-based water surface cleaning robot, a VEGA-based digital telescope, and a forest environment monitoring project; the Design Linked Incentive (DLI) scheme has supported 24 companies. The need for restraint is also clear: the GPU goal has not disclosed manufacturing partners, chip suppliers, or architecture licensing details, making it a policy declaration and roadmap, while the verifiable existing SHAKTI and VEGA products (100MHz-class 32-bit MCUs) remain far from “AI acceleration.”
- Sources: WeChat public account (HPMicro), ERICI official site, Hacker News discussion, Tom’s Hardware (India)
RISC-V 100 Leaderboard Weekly Update
No new measurement results appeared on riscv100.top this week; the leaderboard saw only backfilled entries and new site features, placed later in this report per convention.
- Four backfilled records: The 11-02 sweep surfaced a 10-17 entry, “XSCC 2.0 on XiangShan KMHv3 RTL 3GHz,” int 68.41, fp 76.95 (speed/ref, Completed), up int +0.8% and fp +0.5% versus the previous record from the same compiler (10-15); the 11-03 sweep surfaced a 07-30 entry, “XSCC 1.1 on SpacemiT K3 X100 2.4GHz,” int 0.9, fp 1.07 (rate/ref), up int +2.3% and fp +9.2% versus the platform’s previous record (05-31, LLVM 22.1.0); the 11-04 sweep surfaced a 10-10 entry, “XSCC 2.0 on SpacemiT K3 X100,” int 1.01, fp 1.13, up int +12.2% and fp +5.6% versus the platform’s previous record (07-30, XSCC 1.1). All three predate their sweep dates and are backfilled entries; they should not be read as new results from this week, nor used to infer architectural improvements. The 07-30 and 10-10 entries form the only paired before/after comparison across major versions on the same platform.
- Two site features: Hardware Comparison (a cross-hardware comparison page) launched 11-03, and Run Comparison launched 11-05, organizing the three cross-architecture SPEC leaderboards into an item-by-item comparison view (normalized per GHz). The comparison data is more informative than total scores: XiangShan KMHv3 leads Graviton5 (Neoverse V3) on libquantum and xalancbmk, but trails by 18% to 31% on compilation and graph-search workloads such as gcc, astar, and mcf; DAMO XuanTie Yitian 710 beats Zen 5 on per-GHz values for llvm, gcc, and sqlite, but trails by 75% on stockfish.
- One artifact worth noting: The 11-06 sweep script reported 39 “leaderboard entries delisted,” which was actually caused by the comparison-view reference rows launched 11-05 disappearing from the six leaderboard pages that day; these are not real delistings and are not counted as score changes in this report. Three Upcoming items remain in Preparing: AWS Graviton5 NeoverseV3 (SPEC2006/2017/2026), SpaceMiT K3 X100 (Vectorization/IME), and GCC 16 SPEC CPU Testing; committee membership was unchanged this week.
- Sources: riscv100.top cross-arch SPEC CPU2006, cross-arch SPEC CPU2026, homepage
VI. Trend Watch
- The automotive battleground has shifted from “does it have a RISC-V core” to “which profile does it conform to”: This week it appeared at all four layers simultaneously—at the chip layer, Infineon moved its AURIX mainline to RISC-V; at the platform-spec layer, Quintauris used EM-Altair to pin down interrupts and memory ordering; at the software-stack layer, Guoxin completed AUTOSAR and IAR adaptation; and at the ecosystem-case layer, Zijin M100 was selected as a demonstration case on shipments of tens of millions of units. For domestic teams this is both a dividend (AUTOSAR, RTOS, and toolchains need not be rebuilt by every vendor) and pressure (customers will ask “does it conform to a given profile”).
- Convergence layers for fragmentation are emerging at different heights, each in a different form: RVISA offers a virtual instruction notation on the AI-matrix side, turning “optimizations a programmer can express” into hard constraints; EM-Altair provides a binary-compatible execution environment on the MCU side; openRuyi makes RVA23U64 the default triple; Linux 7.4 uses hwprobe to make runtime probing possible. What these four share is that, rather than leaving each chip vendor to interpret the spec on its own, they insert a stable software target between spec and implementation.
- Open-source core investment is shifting toward verification and observability: Of XiangShan’s 18 commits this week, the weightiest was a decode bug that would turn inval into zero; the same period added BPU counters, CROB debug signals, and an AGENTS.md aimed at AI coding agents. CVW brought rv64gc coverage to 100% and treated Ubuntu 26.04 installation and buildroot rebuilds as serious work; all six GEM5 commits concerned checkpoints, traces, PMU, and weekly performance tasks. Scores are only worth comparing when built on reproducible measurement—this has become the consensus among open-source cores.
- Domestic suppliers are moving from “publishing specs” to “delivering usable products”: SpacemiT advanced on specs (RVISA), SDK manifests, robot-arm manuals, and safety boundaries (no auto-enable/clear faults/return to home/retry) all in one week; Sophgo offered end-to-end stress tests on 2-core and 3-core configurations rather than peak compute; Zijin turned “ecosystem promotion” into a case that can be formally submitted; Nuclei moved its development entry point from Eclipse to VS Code. All four are shoring up the “can others actually use it” side.
- Commercialization pace is set by market momentum, not roadmaps: Andes slowing its high-end Cuzco core and shifting resources to its AI business, which already accounts for half of revenue, and Infineon moving its mainline to RISC-V are two signals pointing in opposite directions but not contradictory—Cuzco is not cancelled, it is waiting for a signal worth accelerating on; what is actually paying today is AI edge, control plane, and automotive. For teams building high-end cores, this is a reality check on “who pays first.”
- Benchmark sites are in a consolidation phase of “locking down methodology first”: This week’s leaderboard had only backfills, comparison views, and a version-progress page, with no new measurements; the 39 “delistings” on 11-06 were an artifact of comparison rows disappearing. This is the same kind of move as XiangShan adding checkpoints, traces, and PMU on GEM5—the comparability of numbers matters more than the numbers themselves.
Appendix: Sources and Verification Notes
Source Composition
- Five daily briefs: 2026-11-02 (72-hour window, covering 10-30 05:00 to 11-02 05:00), 11-03, 11-04, 11-05, 11-06 (each a 24-hour window, continuing from the prior cutoff). The main body of this report therefore covers events across the eight days from 10-30 to 11-06, of which 10-30 to 11-01 falls within the normal coverage of the Monday brief; affected items are marked in their titles with the original date.
- Friday backfill: the final cutoff of the five daily briefs was 11-06 05:00. This report re-checks that period separately and counts four openRuyi main-repo commits into the body (clonezilla, llvm-defaults plus python-clang, tmux 3.8, qt6-qtwebengine build fix), verified via commits.atom.
- Post-period additions (not counted in the body): from the early hours of 11-07 to now, NEMU #1280 and difftest #1028/#1029 replaced the matrix completion fingerprint with CRC32C and compressed the DiffTest event payload; CVW merged a batch (PWM center-aligned sticky and APB read-strobe fixes, UART tests, a core comparison table giving rv32gc CoreMark/MHz 4.33, counter mask shifts, README psyn); spacemit-robotics/manifest #32 registered spacemit vslam. These are developments after the early hours of Saturday and are left for the next issue.
Supplementary Verification Results
| Verification item | Result |
|---|---|
| GitHub watched organizations (168-hour window) | The organization-level anonymous API quota was exhausted this week, returning rate limit per organization; verification was switched to per-repo checks via each key repository’s commits.atom. Push counts within the window (from the inspection sweep): spacemit-robotics 14 repos, sipeed 8 repos, revyos 6 repos, openhwfoundation 9 repos, openRuyi-Project 13 repos, radxa-linkr 0 repos; XuanTie and ESWIN had only sporadic pushes to documentation and test repos, and not all non-key repos were expanded |
| Per-repo review of the XiangShan family | Main repo 18 commits (four on 11-03, nine on 11-04, four on 11-05, one on 11-06); NEMU 4 commits; difftest 11 commits (including nine from 10-30 to 11-01); GEM5 6 commits; XSCache 1 commit (submodule pointer updated from the XiangShan side). The XiangShan API side returned quota limits, so verification was done against the PR numbers recorded commit-by-commit in the daily briefs |
| CORE-V and OpenHW | CVW had three batches totaling 29 commits this week (3 + 15 + 11); cvfpu 1 commit (FFLARNC naming of the clear signal); the CVA6 main branch has had no new commits since going read-only on 10-23, and pushes within the window came from non-default branches |
| openRuyi dual-source | Main repo 6 / 7 / 15 / 10 / 20 commits across the five briefs plus 4 during the Friday daytime, roughly 62 in total; clang-wrap had another 2 commits (ninja link command, LLVM IR handling); the official news page scraped normally, with the latest announcements still the two from October 2026 (10-19) and no November release yet |
| revyos | The 7.3.y branch merged twenty configuration commits at once on 11-03, including K3’s UFS and display controller, a full USB3/Type-C group, ESWIN EIC7700, linuxboot, and CI cross-compilation; the main branch also had pushes at the end of the window, not expanded commit-by-commit |
| DAMO XuanTie, ESWIN, Nuclei, Espressif | XuanTie: damo-rv-priv-ats two commits (capability gating changed to declarative configuration, fixing PMP not being cleared after an S-state to HS-state switch), no pushes in other key repos; ESWIN: zero pushes all week, latest still the 10-19 linux-stable; Nuclei: nuclei-tool-guide multiple commits (2026.10 changelog, rvtrace and Hummingbird troubleshooting), new empty T0 Chiplet dual-SDK repos; Espressif: mainly routine ESP-IDF ecosystem maintenance, with RISC-V-related work in esp32s31-bt-lib, esp-vision, and the standalone esp-riscv-trace |
| SpacemiT spacemit-com and spacemit-robotics | spacemit-com: new bianbu-agentic umbrella repo and rvisa specification repo, multiple commits in model-zoo-vision, docs-ros, docs-chip, docs-product; spacemit-robotics: manifest three commits, manipulation and dice_arena 16 / 18 commits respectively, with pushes in humanoid_linglong, ros2_humanoid, humanoid_common, and control_whole_body |
| Sophgo | sophon-demo (2/3-core variants of Qwen3.5-0.8B, Qwen3.5-0.8B CV84X6 adaptation), LLM-TPU #184, weekly cvi_rtt release, SG2044 EDK2 disabling ECRC, mcu I2C hardening; the same-minute batch push across 10 repos including libsophon on 10-31 was a full SDK sync |
| riscv100.top inspection | Daily archived summaries: one backfill record each on 11-02, 11-03, 11-04; Run Comparison went live on 11-05 with comparison entries generated in bulk; 11-06 saw only a What’s New reordering and a spurious delisting from 39 comparison rows disappearing; no new measurements this week, the three Upcoming items still Preparing, no committee changes |
| Member company website inspection (391 sites) | Results by period: no substantive updates on 11-02 and 11-03; on 11-04, 391 sites inspected, 100 changed (mostly university and alliance event pages, homepage redesigns), 68 fetch failures; on 11-06, 391 sites inspected, 92 with page changes, involving Google, National University of Singapore, StarFive, Infineon, Renesas, etc., all reviewed individually and found to be homepage marketing content updates; on 11-07, re-inspection of 391 sites (223 unchanged, 99 content changes, 69 fetch failures), with changes concentrated in university homepages, association event pages, and Cloudflare block pages, and no new products, funding, partnerships, or architecture announcements |
| riscv.org official blog | No new articles within the window |
| Pending-inclusion queue | No pending items on this topic during the period |
Limitations and Methodology Notes
- Backfills are not new results: the three backfill records on riscv100.top this week (10-17, 07-30, 10-10) all predate the inspection date; this report marks each individually and does not count them as “new performance this week,” and the 0.8% to 0.5% differences under the same compiler are within noise. The 39 “delistings” on 11-06 resulted from comparison rows disappearing in the site’s comparison view and are not treated as score changes.
- Vendor claims are not third-party verified: XuanTie C950’s 6 billion units shipped, 400 customers, 1,000 licenses, and 30%+ cloud workload performance improvement are official figures; Zijing M100’s 10 million units shipped and ASIL-D/ASIL-B certification are company figures; Vietnam’s Viettel “encryption/decryption roughly 15x software solutions” is an official test figure; the Indian GPU’s 24-month target discloses no details on manufacturer or architecture licensing and is a policy statement.
- Specifications whose maturity must be distinguished: RVISA is a working draft (as stated in the specification status file), not a RISC-V International-approved specification; the target mapping profile is registered only as a draft with no implementation-conformance evidence established, and the MLIR dialect and conformance tooling are TBD; EM-Altair is a joint venture’s platform specification, and its ASIL-B wording is “equivalent requirements,” which does not mean certification has been completed.
- Commit count methodology: the commit counts for XiangShan, CVW, and openRuyi are aggregated from the PR/commit merge counts recorded day by day in the daily briefs; across day-boundary windows the same batch of commits may be counted once in each of two adjacent briefs (e.g., CVW’s 11-04 to 11-05 and 11-05 to 11-06 batches), and this report lists them separately per the daily-brief methodology without forcing deduplication and summation; organization-level push counts come from the inspection script and use a different methodology from per-commit counts.
- Duplicates and deduplication: the price, headcount, and pivot-to-CHERI judgment of onsemi taking over Codasip’s low-end business were covered in the previous weekly report; this week only adds the “transaction completed” wording and does not count it again as a headline item; XuanTie’s 6 billion units and the C950 summit content are secondary传播 of the 10-19 summit, counted as one retrospective item in each brief from 11-02 to 11-04, and this report expands only on the new fact of the “China Chip” award; reprints this week of items already reported in late October, such as onsemi, OpenBSD 8.0, and Qwenputer, are not counted again.
- Time window: there are usually no daily briefs on weekends, so this weekly report covers Monday through Friday; developments after the early hours of 11-07 are listed under “Post-period additions” and not counted in the body’s judgments.
- Sources: all items are taken from the daily RISC-V Global Developments Briefs (2026-11-02 to 2026-11-06); key data were verified against this period’s GitHub commit stream, the openRuyi official news page, the daily riscv100.top archives, and the member website inspection.