RISC-V Daily Intelligence Report (2026-11-10)
Research window: 2026-11-09 05:00 ~ 2026-11-10 05:00 (Beijing time, 24 hours) Sources: Google News aggregation (30+ query sets in Chinese and English, via proxy), Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International official site and weekly newsletter, openRuyi official site and main repo, RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep, and commit searches across GitHub organizations and key repositories including DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, Radxa Linkr, CORE-V, and XiangShan.
In This Issue
- Today’s Highlights: Nuclei releases NI1000, the first RISC-V general-purpose AI compute core supporting the VME matrix extension (11-09)
- I. Open-Source Community and Software Ecosystem
- openRuyi main repo: 20 commits in one day — wlroots 0.20.2, wget2 2.3.0, and Python binding fixes for LLVM 22 through 24 (11-09/11-10)
- CORE-V CVW merges 11 commits: 16550 UART LSR error bits, SPI empty-RX zeroing, and lockstep misalignment exemption (11-09/11-10)
- M9B builds a RISC-V computer on Tang Nano 9K, running five DOS-era games on a 36MHz soft core (11-09)
- RISC-V Day Tokyo 2026 Autumn opens tomorrow, with Tenstorrent and AIST attending (11-11)
- II. Product and Ecosystem News
- M5Stack Core Matrix ships with ESP32-C61 plus a 16×16 LED matrix, on sale at $29.99 (11-09)
- Infineon AURIX RC1’s RISC-V core reportedly derived from MIPS IP, fabricated by GlobalFoundries (11-09)
- III. Deep Dive on Chinese Teams
- T-Head plans Hong Kong IPO in 2027, targeting $2–3 billion raise (11-09)
- SpacemiT: four commits in one day to ggml-spacemit; gated delta network rewritten with RVV and TCM (11-09)
- SpacemiT: digital human pipeline renamed bianbu-avatar, aligned with K3’s EP core allocation (11-09)
- Sophgo: LLM-TPU supports MTP for Qwen3.5 dense models; bootloader switched to SG2044 toolchain to build SG2042 (11-09)
- Nuclei: 12 commits to toolchain docs; 2026.10 release upgrade notes and QEMU release notes land (11-09)
- DAMO XuanTie and ESWIN: zero organization-level pushes during the window (11-10)
- IV. Special Chapter: Embodied Intelligence and Robotics
- MIPS Actus uses dual Atlas M8500 real-time RISC-V cores for robot joints, with TSN and post-quantum security (11-09)
- spacemit-robotics: unitree-go2 supports smallcore motor communication; perception packages the inference runtime into las2 (11-09)
- V. Summary and Trend Observations
- Appendix: Full Source List and Source Verification Table
Today’s Focus: Nuclei Releases NI1000, the First General-Purpose RISC-V AI Compute Core Supporting the VME Matrix Extension (11-09)
Nuclei Technology released the NI1000 RISC-V CPU IP on November 9. The company says it is its first high-performance general-purpose AI processor IP supporting the Vector Matrix Extension (VME). The NI series is Nuclei’s dedicated product line for AI applications and can be configured as the 900 or 1000 series; the VLEN of the RVV 1.0 vector unit (VPU) can be set to 512, 1024, 2048, or 4096 bits, and each core in a cluster can independently configure its own VPU. Cores are tightly coupled with NPU accelerators via IOCP (IO Coherent Port) while maintaining cache coherence, and the NICE/VNICE interfaces allow users to define custom scalar and vector instructions; NICE can also be used for tightly coupled interaction with the NPU.
The technical highlight lies in how VME is implemented. VME is tightly coupled with the CPU and RVV within the same core; source operands directly reuse RVV’s vector register file, and the compute units also reuse RVV’s execution units, adding only the area of the Matrix Tiles. The new native matrix instructions accumulate outer-product results into dedicated N×N accumulators, four independent Matrix Tiles run in ping-pong parallel to boost throughput, and native support for the FP8 data type is added. Nuclei’s figures are: running a 32×32 FP32 matrix multiplication on the NI1020 (VLEN 1024) increases overall area by only 15% (mainly the Matrix Tiles) while improving matrix multiplication performance by 2.5×.
The positioning is stated with some restraint. The company explicitly says the NI series does not intend to replace accelerator IP such as NPUs and GPUs, but rather to combine with them and better serve accelerators on AI SoCs. This is consistent with VME’s trade-offs: compared with IME, which has independent matrix accumulation hardware and higher matrix compute energy efficiency, and with AME, which eliminates the full independent matrix register file and hardware matrix accelerator, VME sits in the middle, trading controllable hardware overhead for matrix compute throughput at the cost of lower peak energy efficiency than dedicated matrix units. Nuclei also stresses that VME is built entirely on the RVV foundation, so the software ecosystem can be reused, and judges that VME will become the earliest-standardized RISC-V matrix extension.
Placed in this month’s context, this fills in the “general-purpose compute core” layer. The previous week’s news centered on automotive-grade MCUs (Infineon AURIX RC1), edge vision SoCs (VeriSilicon VS815V), and security cores (AMD’s RISC-V TEE), all of which put RISC-V into specific functional units; NI1000 addresses the general-purpose core in an AI chip that “has to do a bit of everything”—running scalar control, running vector operators, and also handling matrix multiplication of ordinary scale, then dividing work with the NPU. Domestic AI chip companies have long faced a binary choice at this layer: either buy high-performance application cores or develop their own and bear the software stack burden. VME’s approach of reusing RVV essentially lowers the cost of that choice; whether customers accept it depends on whether the toolchain and operator libraries can keep up.
Date: 2026-11-09 Source: Nuclei Technology press release (Chinese), Design And Reuse report
I. Open Source Community and Software Ecosystem
openRuyi Main Repo: 20 Commits in a Single Day — wlroots 0.20.2, wget2 2.3.0, and Python Binding Fixes for LLVM 22 through 24 (11-09/11-10)
The openRuyi main repo landed 20 SPECS commits within the window, with the rhythm of a typical “rolling-release daily” rather than a concentrated release: the desktop stack bumped wlroots to 0.20.2, networking tools bumped wget2 to 2.3.0, and on the Python side httpx2 and httpcore2 were updated to 2.13.0, idna to 3.20, and langgraph-sdk to 0.4.11 in one go, while python-pymongo now bundles BSON into the main package. Several other commits are packaging-correctness fixes: the llvm22, llvm23, and llvm24 packages use sed to change clang’s Python bindings to link against libclang, and openruyi-linux-build fixes devel file conflicts and installs versioned documentation.
Worth noting are the three llvm commits. LLVM’s Python bindings across versions have long relied on dynamic library path inference; if a distro’s packaging points clang’s Python module at a static library or the wrong soname, users installing dependencies like llvmlite or triton on RISC-V will fail at the import stage. Changing all three versions to explicitly link libclang is a root-cause fix for the “installs but doesn’t work” class of problems — more meaningful to teams building a RISC-V AI stack than the version numbers themselves.
Date: 2026-11-09 to 2026-11-10 Source: openRuyi-Project/openRuyi
CORE-V CVW Merges 11 Commits: 16550 UART LSR Error Bits, SPI Zero-on-Empty RX, and Lockstep Misaligned-Access Exemption (11-09/11-10)
CORE-V CVW merged 11 commits within the window, with the main thread still “making peripherals and the verification environment solid” rather than changing the microarchitecture: the 16550 UART gains LSR (Line Status Register) error-bit behavior, SPI reads back zero when the receive FIFO is empty, branch-prediction redirect coverage statistics gain exclusions, lockstep mode gets an exemption rule for misaligned accesses, and a batch of examples and scripts are cleaned up — removing SHA’s Questa scripts, removing Docker build and regression scripts, and unifying ruff and uv script entry points.
This set of commits and last week’s “remove two local branch predictors, publish RV32GC CoreMark/MHz 4.33” are two halves of the same round of wrap-up. Last week touched performance and area; this week touches verifiability: UART error bits and SPI empty-read semantics are both details that “the protocol specifies but implementations easily miss” — missing them won’t stop the core from running, but will hang drivers on error paths; lockstep exemptions and coverage exclusions directly determine whether verification reports can serve as delivery evidence. For teams using CVW for automotive or industrial evaluation, this kind of commit is worth more than a few hundred extra MHz.
Date: 2026-11-09 to 2026-11-10 Source: openhwfoundation/cvw
M9B Builds a RISC-V Computer on Tang Nano 9K, Running Five DOS-Era Games on a 36MHz Soft Core (11-09)
Developer Mika Tammi’s m9b project appeared on Hacker News: on a Sipeed Tang Nano 9K (Gowin GW1NR-9) costing about twenty dollars, a complete small computer is built with the VexiiRiscv soft core — 36MHz, with cache, using the board’s onboard 8MB PSRAM as memory, 720×480 HDMI video, 48kHz audio, two USB ports for keyboard and mouse, and an SD card for storing programs. It boots into an interface listing programs on the SD card and can run five DOS-era games, all with sound effects and music: Doom (shareware, timedemo about 22 fps), Descent (about 10 fps), Tyrian 2.1 (full speed 35 fps), System Shock (about 7 fps), and Ultima VII (two engines, 7 to 13 fps).
The value of such a project is not “it can play games” but drawing the line of “what a computer minimally needs.” The 9K’s logic resources can’t fit an out-of-order core, nor Linux-level memory management, so the trade-offs become very concrete: want HDMI and you must cut cache; want sound effects and you must pre-render music as a stream read from the card. The project splits the SoC, runtime, five game ports, and SD card image into separate repositories, all open source, each documenting where upstream was modified — for teams working on RISC-V teaching boards and FPGA onboarding, this is a directly reusable sample of engineering layering.
Date: 2026-11-09 Source: m9b-system repo, Hacker News discussion
RISC-V Day Tokyo 2026 Autumn Opens Tomorrow, with Tenstorrent and AIST Present (11-11)
The RISC-V Association (Japan) will hold RISC-V Day Tokyo 2026 Autumn on November 11 at the University of Tokyo’s Hongo Campus, with about 400 attendees expected. The morning agenda is a chip design tutorial; the afternoon features technical talks and booths: Tenstorrent speaks on physical AI and the Open Chiplet Atlas, the National Institute of Advanced Industrial Science and Technology (AIST) speaks on 2nm edge AI, Andes Technology has a live demo, and the venue also hosts a special Chip Museum exhibit showing the hardware specs and software overview of Google TPU v3.
Placing this event alongside last week’s RISC-V China Summit, one can see regional events forming a fixed division of labor: the China event leans toward industrial deployment and domestic substitution, the Japan event toward design and education (tutorials, Chip Museum, university venue). For domestic teams, the two talks worth watching are Tenstorrent and AIST — the former’s stance on chiplets and open interconnects will shape the narrative of the RISC-V server路线, and the latter represents Japan’s investment direction in advanced-process edge AI.
Date: 2026-11-11 (event) Source: RISC-V Day Tokyo 2026 Autumn
II. Product and Ecosystem Updates
M5Stack Core Matrix launches with ESP32-C61 and 16×16 LED matrix at $29.99 (11-09)
M5Stack has launched the Core Matrix, an IoT controller built on the ESP32-C61: the SoC is the ESP32-C61HR8, a single-core 32-bit RISC-V running at up to 160MHz, with 320KB SRAM plus 8MB PSRAM and 16MB QSPI flash, 2.4GHz Wi-Fi 6 and Bluetooth LE 5.0; the front features a 16×16 white LED matrix (256 LEDs, driven by the TM1680), paired with a Bosch BMI270 six-axis IMU, microSD, a passive buzzer, three user buttons and a power button, with expansion via a 30-pin M5-Bus and Grove I2C, supporting USB or 3.7V external battery power, measuring 54×54×11.6mm, priced at $29.99. On the software side it supports Arduino IDE and UIFlow2, with driver libraries for M5Unified, M5GFX and the power management chip.
The significance of this board is that it replaces the “display” with a dot matrix. The ESP32-C61 is Espressif’s low-cost Wi-Fi 6 RISC-V model; its compute and memory are insufficient to drive an LCD, but sufficient to drive 256 monochrome LEDs and run state machines and networking logic; using the dot matrix for icons, animations and notifications effectively fills a tier between “no screen” and “has a screen.” For teams building desktop status lights, server-room indicators and low-cost interactive devices, this is a ready-made reference form factor, and the price falls in a range that can be deployed at scale.
Date: 2026-11-09 Source: CNX Software report
Infineon AURIX RC1’s RISC-V core reported to come from MIPS IP, fabricated by GlobalFoundries (11-09)
On November 9, Design And Reuse republished an analysis from jonpeddie.com stating that Infineon’s AURIX RC1 uses MIPS processor IP, fabricated by GlobalFoundries, and calling it a major design win for MIPS and GlobalFoundries in RISC-V. The article also raises a question: will RISC-V eventually replace Infineon’s own TriCore architecture? When Infineon released the RC1 on November 4, the official line was “the first RISC-V-based automotive-grade MCU platform, with the TriCore family continuing to ship,” without elaborating on the IP source.
This fills a gap at the supply-chain level. Automotive-grade chip announcements typically cover only the platform and ecosystem, not where the core comes from; for teams building domestic automotive-grade RISC-V, the source of the core directly determines two things—who provides the functional safety evidence chain (IP-level evidence for ISO 26262 ASIL D cannot be supplied by the OEM alone), and which foundry the capacity and process node land at. After being acquired by GlobalFoundries, MIPS shifted its entire roadmap to RISC-V; if it is indeed the core supplier for the RC1, that means the “RISC-V automotive-grade core” market has gained a player with a foundry background; for domestic IP vendors, this is a competitor requiring direct comparison, not a distant concept.
Date: 2026-11-09 Source: Design And Reuse repost
III. In-depth Look at Domestic Teams
T-Head Plans Hong Kong IPO in 2027, Targeting $2–3 Billion Raise (11-09)
Citing a Reuters report on November 9, multiple media outlets reported that T-Head Semiconductor, Alibaba’s chip subsidiary, plans to list on the Hong Kong Stock Exchange in 2027, targeting a raise of $2–3 billion (approximately HK$15.6–23.4 billion). Public information shows that T-Head was formed in October 2018 through the integration of C-SKY Microsystems and the DAMO Academy chip team, with products spanning AI accelerator chips (Zhenwu series), Arm server CPUs (Yitian series), RISC-V processor IP (XuanTie series), and storage controllers. In March 2026, Alibaba CEO Eddie Wu disclosed on an earnings call that T-Head chips had shipped more than 470,000 units cumulatively with annualized revenue exceeding RMB 10 billion; in August, the company disclosed that the Zhenwu series had served more than 650 customers, with external paying customers accounting for over 60%. The Zhenwu V900 announced at the September Apsara Conference claims 216GB of memory and 1,200GB/s inter-die bandwidth, with mass production planned for Q1 2027.
The reporting needs to be parsed carefully: the report counts the XuanTie RISC-V processor IP line as part of T-Head’s product portfolio, whereas the XuanTie team moved to DAMO Academy in 2024 to focus exclusively on RISC-V processors (such as the C950), while T-Head’s own product focus is PPU (AI accelerator cards) and Yitian (Arm CPU). The figure of “6 billion cumulative XuanTie shipments” is a group-level historical cumulative metric and does not equal what T-Head currently operates. How the listing entity will ultimately carve up assets, and whether XuanTie IP ownership will be adjusted accordingly, has not been disclosed—this is the part of the news most worth tracking.
For the RISC-V industry, the IPO itself does not change the technology roadmap, but it will change how capital prices this sector. Over the past two years, the valuation anchor for domestic RISC-V companies has been “substitution space”; if T-Head lists with RMB 10 billion-level annualized revenue and 60% external customers, the anchor shifts to “already-commercialized compute supplier.” This has a two-way impact on RISC-V server, AI chip, and IP companies currently raising funds—comparable companies become clearer, but capital markets will also hold them to the same revenue and gross margin standards.
Date: 2026-11-09 Source: iyiou.com, Sina Finance reprint
SpacemiT: Four ggml-spacemit Commits in One Day, Gated Delta Network Rewritten with RVV and TCM (11-09)
spacemit-com’s llama.cpp fork (mtmd-backend branch) merged four ggml-spacemit backend commits within the window: optimizing the recurrent gated delta network with RVV and TCM (#54), unrolling and specializing the gated delta network by head size (#59), optimizing MoE matrix multiplication and adding dense Q4_0 and F16 GEMM (#53), and fixing an implementation error where softplus was missing the log1p term (#57).
Together, the four commits point to the same class of model architecture. The gated delta network (GDN) is a recurrent layer in the linear attention family; at inference each step performs only small-scale matrix operations, and the bottleneck is not compute but memory access and state residency—exactly where TCM (tightly coupled memory) should be used. Unrolling with head size as a specialization dimension eliminates branch and indexing overhead in the inner loop. MoE matrix multiplication and Q4_0/F16 GEMM are another line: the K3’s vector units must serve both sparse expert routing and low-bit dense layers, and missing either would drop whole-machine throughput to the next tier down. The softplus commit is the least conspicuous, but the missing log1p is a numerical stability issue that accumulates into visible drift in long-sequence inference.
Date: 2026-11-09 Source: spacemit-com/llama.cpp
SpacemiT: Digital Human Pipeline Renamed bianbu-avatar, Aligned with K3’s EP Core Allocation (11-09)
The spacemit-com/model-zoo-avatar repository saw nine commits within the window, centered on a digital human (avatar) inference pipeline: adding two inference components, FeatherHuBERT and UNet; unifying resource download paths to model_zoo/avatar; removing generic test audio from the repository in favor of per-persona downloads; fixing face replacement issues in the native paste-back path; and renaming the project from FeatherTalk to bianbu-avatar. One commit explicitly states “match the EP core allocation of bianbu-avatar on K3.”
This set of commits reveals the deployment form of on-device digital humans: the speech side uses a HuBERT-class model for features, the vision side uses UNet for generation, and both stages must land on different compute units of the same SoC—making “EP core allocation” (which core runs which subgraph) a parameter that must be explicitly aligned. The rename is also informative—moving from a third-party model name back to an in-house distribution brand (Bianbu) indicates that this pipeline has progressed from “getting an open-source project running” to “maintained as a distribution component.”
Date: 2026-11-09 Source: spacemit-com/model-zoo-avatar
Sophgo: LLM-TPU Adds MTP Support for Qwen3.5 Dense Models; Bootloader Switches to SG2044 Toolchain to Build SG2042 (11-09)
The Sophgo LLM-TPU repository merged a commit titled “Qwen3.5 dense model support mtp,” adding MTP (multi-token prediction) support for Qwen3.5 dense models. bootloader-riscv saw three commits: switching to the SG2044 toolchain to compile SG2042, changing to build from the devel-sophgo branch, and unifying the BMC firmware packaging script to sophgo-edk2’s gen-tar-for-bmc.sh. On the sophon-tools side, work continued to wrap up the bmssm component, switching the default model to the DeepSeek-V4.1-Flash series, making migration and rollback land in the same transaction, and aligning RELEASE.md with version 2.4.x.
The MTP commit matters more for inference throughput than the version number suggests. Multi-token prediction lets the model produce multiple candidate tokens per forward pass, which are then accepted by a verification stage; the measured benefit depends heavily on whether the backend can efficiently handle verification steps that are “small batch, short sequence.” Sophgo integrating this into LLM-TPU indicates its TPU backend has already accounted for speculative decoding scheduling. The bootloader switch to the SG2044 toolchain for building SG2042 converges the build chains of two platform generations into one, reducing issues like “the same firmware behaving differently on two generations of boards.”
Date: 2026-11-09 Source: sophgo/LLM-TPU, sophgo/bootloader-riscv, sophgo/sophon-tools
Nuclei: 12 Toolchain Documentation Commits, 2026.10 Upgrade Notes and QEMU Release Notes Landed (11-09)
The Nuclei-Software/nuclei-tool-guide repository saw 12 commits within the window, all documentation and release notes: adding 2026.10 upgrade notes and unifying the “known issues” anchor, supplementing the QEMU 2026.10 release notes and known-issues sections for OpenOCD and the Nuclei Studio IDE, splitting libncrt fix entries by version into their respective Known Issues, and bumping the VS Code extension version to 0.3.0.
Viewed alongside the NI1000 released the same day, this clarifies Nuclei’s product cadence: an IP release is only the beginning; customers can only truly deploy once the toolchain (compiler, libncrt, QEMU, OpenOCD, IDE) is complete under the same version number. Twelve commits to the documentation repository in a single day, many of them “known issues,” indicates that the 2026.10 release has entered the customer delivery stage—the known-issues list is part of the deliverable, not an internal memo.
Date: 2026-11-09 Source: Nuclei-Software/nuclei-tool-guide
DAMO XuanTie and ESWIN: Zero Organization-Level Pushes Within the Window (11-10)
The XUANTIE-RV organization’s most recent push remains November 4 (damo-rv-priv-ats and torq-tile), with llvm-project and bcc stopped in late October; ESWIN’s eswincomputing last pushed on October 19. Neither had verifiable new development output within the 24-hour window.
This does not mean the product lines are quiet. XuanTie’s high-performance cores (C950 series) and automotive-grade E908A progress were covered in the October reports, and ESWIN’s RISAA ecosystem platform and RDI Alliance activities were concentrated during the October RISC-V China Summit; public information within the window mainly comes from the T-Head IPO report’s citation of XuanTie shipment figures, which is group-level historical data and does not constitute new technical progress. Not repeated in this issue.
Date: 2026-11-10 Source: XUANTIE-RV organization page, eswincomputing organization page
IV. Special Chapter on Embodied Intelligence and Robotics
MIPS Actus Uses Dual Atlas M8500 Real-Time RISC-V Cores for Robot Joints, with TSN and Post-Quantum Security (11-09)
MIPS (acquired by GlobalFoundries) released progress on the Actus reference platform on November 9 ahead of electronica 2026: a chip for compact robot joints that puts deterministic motor control, intra-robot networking, and edge inference on the same die, based on GlobalFoundries’ FDX FD-SOI process. At its core are two MIPS Atlas M8500 real-time RISC-V processors, using hardware multithreading to eliminate context-switching overhead during interrupt servicing; a dedicated trigonometric accelerator is more than 16x faster than standard trigonometric libraries on field-oriented control (FOC) algorithms; it pairs 100-picosecond high-resolution PWM and sigma-delta ADCs to drive high-current motors, and additionally uses an ARC VPX 512-bit vector unit to run sensor fusion and predictive maintenance models on the same chip.
The system side is equally well provisioned: three-port Time-Sensitive Networking (TSN) gigabit Ethernet plus CAN-FD and serial ports, allowing joint controllers and the central controller to exchange position and torque data; an on-chip hardware security module supports post-quantum cryptography and provides support for functional safety. Actus, together with Acies (an inference-native platform) and Aegis (a security-native CPU platform), forms MIPS’s physical AI product line.
The significance of this item lies in the “joint” position. The compute narrative for humanoid robots has long centered on the brain (whose NPU, how many TOPS), while what actually bottlenecks mass production is the segment from elbow to fingertip: space, power, and heat dissipation are all extremely tight, yet smooth humanoid motion and deterministic timing are required. Packing dual real-time cores, motor control peripherals, TSN, and a security module into a single joint chip amounts to turning the distributed architecture of “the brain issues commands, the joint closes the loop itself” into a purchasable component. For domestic teams working on robot joints and real-time control, this is a solution that directly contrasts with the “MCU plus external driver” approach; the key difference lies in hard real-time performance and network determinism, not in compute.
Date: 2026-11-09 Source: MIPS press release
spacemit-robotics: unitree-go2 Adds smallcore Motor Communication Support, perception Packages Inference Runtime into las2 (11-09)
The spacemit-robotics unitree-go2 repository merged #6, adding support for smallcore motor communication (feat/smallcore-motor-tra… branch); the perception repository has two fixes around las2: packaging the matching inference runtime into the release artifact (#14), and before that removing the model files packaged into the source repository (#13).
Both items point to “delivery form.” unitree-go2 is a port of the Unitree Go2 quadruped robot to the K3 platform, which had already completed IMU snapshot reading, terrain walking policies, and model downloading; adding smallcore motor communication means the control chain must be able to interface with joint driver boards from different batches — a hurdle that must be cleared to turn a demo into a reproducible project. perception’s #13 and #14 come as a pair: first strip the large model files out of the source repository (avoiding repository bloat), then package the inference runtime together with the release artifact (so users don’t have to hunt for dependencies themselves) — one subtraction and one addition, indicating that this perception stack is beginning to be organized as “an installable software package” rather than “a pile of scripts.”
Date: 2026-11-09 Source: spacemit-robotics/unitree-go2, spacemit-robotics/perception
V. Summary and Trend Observations
1. A dedicated product has emerged at the “general-purpose core” layer of AI chips. Nuclei’s NI1000 uses VME to graft matrix capability onto RVV, adding only 15% area while making matrix multiplication 2.5x faster; it is positioned to divide labor with the NPU rather than replace it. This differs from the narrative of the past two years, which centered on “whose NPU has the higher peak”: the general-purpose core in an AI SoC has long been addressed by purchasing application cores or developing a custom big core, and now a third path has appeared—vector as the foundation, with matrix added on demand. The real barrier is not the IP, but whether the operator libraries and compilers can actually put VME to use.
2. Robot joints have become a new battleground for RISC-V real-time cores, and the approaches are beginning to diverge. MIPS Actus uses dual Atlas M8500 cores plus TSN and post-quantum security to build a joint-level single chip, while domestically spacemit-robotics continues to work on the K3 for quadruped porting and control-chain adaptation. The former sells devices and real-time guarantees; the latter accumulates reproducible engineering and ecosystem. Both paths point to one thing: the compute competition in embodied intelligence is sinking from the “brain” down to the “cerebellum and joints,” and the evaluation criteria at this layer are determinism and power consumption, not TOPS.
3. Capital-market moves on the domestic side are beginning to shape the technology narrative. T-Head plans to pursue a Hong Kong IPO in 2027, raising $2 billion to $3 billion; if it goes through, RISC-V and domestic AI chip companies will for the first time have a comparable benchmark that is “already commercialized, with external customers and revenue.” A measure of caution is warranted: reports count the XuanTie IP within T-Head’s product line, yet the XuanTie team was transferred to DAMO in 2024, and the asset boundaries of the listing entity have not been disclosed. For industry observers, what matters next is not the valuation, but how revenue from external customers and related-party transactions are separated in the prospectus.
4. Progress on the software side is concentrated in details of the “installs and works” variety. Among openRuyi’s 20 commits in a single day, the most representative is the change to explicit linking against libclang for the Python bindings of llvm22 through 24—it improves no metric, yet it determines whether llvmlite and triton can be installed normally on RISC-V; the UART LSR error bit, SPI empty-read zeroing, and lockstep exemption merged this week in CORE-V CVW fall into the same “doesn’t affect benchmarks, affects whether it can ship” category; Sophtgo’s convergence of the bootloader build chains for two platform generations and SpacemiT’s folding of the digital-human pipeline into the Bianbu brand are likewise moves from demo toward maintenance.
5. A quiet news surface and a restless repository surface. In this issue, the only substantive new items on the Google News side are four: Nuclei NI1000, MIPS Actus, M5Stack Core Matrix, and the IP source for Infineon AURIX; three others—VeriSilicon VS815V, AMD RISC-V TEE, and Openchip BER10—are second-hand circulation of old news already covered in previous reports and are not repeated here. This is consistent with last week’s rhythm: major vendors’ announcements are sparse, but commit density on the GitHub side is rising—SpacemiT, Sophgo, Nuclei, openRuyi, and CORE-V together had more than 60 verifiable commits in the same 24 hours.
Limitations note: Anonymous access to the GitHub API from this machine hit the rate limit of 60 requests per hour during the collection window, so organization and repository activity was instead obtained from data embedded in GitHub organization pages and from each repository’s Atom feed, which does not consume API quota; therefore, the determination of “had a push within the window” is based on the commit time on the default branch, and pushes to non-default branches (such as SpacemiT’s wip branch) may be underestimated. The “zero push” conclusions for DAMO XuanTie and ESWIN are likewise subject to this methodological constraint and only mean that the default branch had no commits within the window. Chinese-language queries on Google News still contain noise from gambling and stock-market pieces, which have been removed based on title semantics; the 108 “changes” from the RISC-V member website inspection were all homepage marketing rotations or security interstitial pages, with no substantive product or architecture announcements found.
Appendix: Complete Source List and Source Verification Table
Complete Source List
- [1] Nuclei NI1000 press release (Chinese) — https://www.nucleisys.com/newsdetail.php?id=475
- [2] Nuclei Releases the NI1000 Series (Design And Reuse) — https://www.design-reuse.com/news/202531411-nuclei-releases-the-ni1000-series-general-purpose-compute-core-with-support-for-the-vme-matrix-extension
- [3] openRuyi-Project/openRuyi main repository — https://github.com/openRuyi-Project/openRuyi
- [4] openhwfoundation/cvw repository — https://github.com/openhwfoundation/cvw
- [5] m9b-system repository — https://github.com/the-parts-bin/m9b-system
- [6] M9B discussion (Hacker News) — https://news.ycombinator.com/item?id=50428584
- [7] RISC-V Day Tokyo 2026 Autumn — https://riscv.or.jp/en/risc-v-day-tokyo-2026-autumn-2
- [8] M5Stack Core Matrix coverage (CNX Software) — https://www.cnx-software.com/2026/11/09/m5stack-core-matrix-esp32-c61-iot-controller-features-16x16-led-matrix-wifi-6-and-6-axis-imu/
- [9] Analysis of the IP origin of Infineon AURIX RC1 (republished by Design And Reuse) — https://www.design-reuse.com/news/202531415-infineon-s-risc-v-aurix-platform-is-a-major-win-for-mips-and-globalfoundries
- [10] T-Head reportedly plans Hong Kong IPO in 2027 (iyiou news brief) — https://www.iyiou.com/briefing/202611091945019
- [11] T-Head IPO coverage (republished by Sina Finance) — https://t.cj.sina.cn/articles/view/5105460195/1304f23e3001024aqo
- [12] spacemit-com/llama.cpp — https://github.com/spacemit-com/llama.cpp
- [13] spacemit-com/model-zoo-avatar — https://github.com/spacemit-com/model-zoo-avatar
- [14] sophgo/LLM-TPU — https://github.com/sophgo/LLM-TPU
- [15] sophgo/bootloader-riscv — https://github.com/sophgo/bootloader-riscv
- [16] sophgo/sophon-tools — https://github.com/sophgo/sophon-tools
- [17] Nuclei-Software/nuclei-tool-guide — https://github.com/Nuclei-Software/nuclei-tool-guide
- [18] DAMO XuanTie organization page — https://github.com/XUANTIE-RV
- [19] ESWIN Computing organization page — https://github.com/eswincomputing
- [20] MIPS Actus press release — https://mips.com/press-releases/mips-unveils-actus-platform-for-precise-secure-robotic-motion-control
- [21] spacemit-robotics/unitree-go2 — https://github.com/spacemit-robotics/unitree-go2
- [22] spacemit-robotics/perception — https://github.com/spacemit-robotics/perception
Source Verification Table
| Source | Verification Result |
|---|---|
| Google News aggregation (30+ Chinese and English query sets, via proxy) | 4 valid new items within the window: Nuclei NI1000, MIPS Actus, M5Stack Core Matrix, and the analysis of the IP origin of Infineon AURIX; VeriSilicon VS815V, AMD RISC-V TEE, and Openchip BER10 (released in July) were second-hand recirculation of older news and were not re-included; queries for SiFive, Tenstorrent, XiangShan, SpacemiT, Sophgo, and others returned zero hits within the window |
| Hacker News | 1 valid item: the M9B RISC-V computer on a Tang Nano 9K; the rest were unrelated technical discussions |
| Phoronix | No new RISC-V items within the window (the most recent was the 10-21 Linux 7.4 crypto library optimization) |
| CNX Software | 1 item: M5Stack Core Matrix (ESP32-C61) |
| LinuxGizmos / Design And Reuse / EEWorld | Design And Reuse: 2 items (NI1000, IP origin of AURIX); LinuxGizmos: no RISC-V items within the window; EEWorld’s “RISC-V China Summit Recap” with IAR is a vendor blog and was not included as an item |
| RISC-V International official website and weekly newsletter | No updates on the official blog within the window; the 11-06 weekly newsletter provided leads on M9B and riski5, which were used for background verification and not listed as separate items |
| openRuyi dual sources (official news page + main repository) | The latest items on the official news page remain the two from October 2026 (Ruyi Firmware, ecosystem infrastructure); the November release has not been published; 20 SPECS commits in the main repository within the window, included |
| GitHub organization and key repository checks (24 hours) | Active within the window: Sophgo 7 repos, SpacemiT 10 repos, openRuyi 5 repos, CORE-V 5 repos, Sipeed 4 repos, spacemit-robotics 2 repos, Nuclei 1 repo, Espressif (toolchain and ESP-IDF ecosystem) multiple repos, RevyOS 2 repos; DAMO XuanTie, ESWIN, Radxa Linkr, and riscv-software-src had zero pushes |
| XiangShan mainline | No commits within the window; llvm-project last updated 11-03; recent commits to NEMU and difftest were already included in the 11-09 report |
| RISC-V member company website sweep | 391 companies: 214 unchanged, 108 changed, 69 failed; all changes were homepage marketing rotations or security block pages, with no substantive product or architecture announcements |
| riscv100.top ranking sweep | No changes |
| WeChat official accounts and vendor websites | Checked the official websites and GitHub organization pages of Nuclei, Sophgo, and SpacemiT; no other substantive releases were found within the window |