Research window: 2026-11-19 05:00 ~ 2026-11-20 05:00 (Beijing time, 24 hours). Sources: Google News aggregation (24 query sets in Chinese and English, plus 10 additional sets of Chinese team keywords and vendor English queries, via proxy). Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International and RISE blogs, openRuyi official website. RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep. Commit and release searches on GitHub covering organizations and key repositories including XiangShan, CORE-V, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT, revyos, and openRuyi-Project.


Index

  • Today’s Highlights: StarFive’s Dubhe-70 ships in volume with JH-B100; domestic out-of-order RISC-V CPU IP enters scaled commercial deployment in data centers for the first time (11-19).
  • I. Open Source Community and Software Ecosystem.
    • RISE: Native riscv64 Python wheels pass 1,500, with 519 still missing (11-18).
    • openRuyi twenty commits: gRPC upgraded to 1.85.3 and relinked against system OpenSSL 4, protobuf to 36.2 (11-19).
    • revyos/linux twenty commits: kernel moves to Debian packaging layer, K3 adds WRS.STO errata (11-20).
    • XiangShan three repos ten commits: matrix and AME verification become the main line, GEM5 adds SMT and divider (11-19).
    • Hypnos: TU Delft’s dual-phase continual learning platform, RISC-V MCU whole system at 33.3 mW (11-19).
  • II. Product and Ecosystem Developments.
    • 2026 RISC-V Industry Development Conference Macau session: Bay Area Open Source Ecosystem Center launched, open computing platform signed (11-19).
    • Infineon moves flagship AURIX family to RISC-V, riscv.org publishes five-question interview (11-19).
    • LILYGO T-CAN485 C5: ESP32-C5 enters industrial control boards, CAN and RS-485 on the same board (11-18).
  • III. Domestic Team Deep Dive.
    • SpacemiT llama.cpp syncs with upstream, precision enum changes follow upstream (11-19).
    • ESWIN StarPro64: 1.6GHz operating point, fan and watchdog added to device tree (11-20).
  • IV. Embodied Intelligence and Robotics Special Chapter.
    • SpacemiT robotics stack three commits: clock anchor, depth hole filling, reference motion planner (11-19).
  • V. Summary and Trend Observations.
    • Three-day trend: management layer ships numbers, packaging layer fills in foundations, platform side signs agreements (11-17 ~ 11-19).
  • Appendix: Complete source list and source verification table.

Issue Summary: The most concrete item today is StarFive putting an out-of-order RISC-V core into volume data-center shipments. The Dubhe-70 is integrated into the server BMC chip JH-B100, and StarFive claims this is the first scaled commercial deployment of domestic out-of-order RISC-V CPU IP in server scenarios, with annual shipments “expected” to reach the ten-million level. The second item is Infineon. This vendor, which accounted for more than one-third of global automotive MCUs in 2025, has brought out the first generation of RISC-V products in its flagship AURIX family, and riscv.org simultaneously published a five-question interview. All three open-source items land on foundations: RISE’s Python wheels pass 1,500, openRuyi switches gRPC to system OpenSSL 4, and revyos moves kernel configuration into the Debian packaging layer. On the domestic team side, SpacemiT’s robotics stack filled in three areas in one day: time baseline, depth map, and planner interface.


Today’s Focus: StarFive Dubhe-70 Enters Mass Production with JH-B100, Marking the First Large-Scale Commercial Deployment of a Domestic Out-of-Order RISC-V CPU IP in Data Centers (11-19)

On November 19, StarFive announced that its self-developed RISC-V CPU IP “Dubhe-70” has achieved mass production and delivery, integrated into the server BMC chip JH-B100. The company’s statement: this is the first time a domestic out-of-order RISC-V CPU IP has achieved large-scale mass production commercialization in a data center server scenario, with annual shipments expected to reach the tens of millions. Previously, on November 17, StarFive had officially announced JH-B100 mass production and customer delivery, without disclosing the core model or shipment expectations at that time.

Dubhe-70 specifications per StarFive’s public materials: 9-stage or deeper pipeline, three-issue superscalar, deep out-of-order execution, support for multi-core cache coherency and I/O cache coherency, instruction set RV64GCBH. This core was announced in August 2024. The metrics given at that time were SPECint2006 7.2 points per GHz, with performance benchmarked against Arm Cortex-A72/A510. Relative to Cortex-A55, StarFive claimed 80% higher performance, 32% higher energy efficiency, and 90% higher area efficiency. The wording in this mass production announcement has changed to “approaching Cortex-A73.” The two statements are not entirely consistent; this record follows the new announcement while retaining the old figures.

In addition to Dubhe-70, JH-B100 also integrates StarFive’s self-developed Dubhe-Star-500 coherent NoC. The chip is a quad-core configuration at 19200 DMIPS, compatible with DDR4-3200 and DDR5-4800, integrates PFR and PROT, and supports Chinese national cryptographic algorithms SM2/SM3/SM4. Interfaces include dual PCIe 4.0 endpoints, full-featured eSPI Slave, PECI, and two LTPI channels. In its product line, StarFive also has two higher-performance CPU IPs, Dubhe-83 and Dubhe-100, as well as the Star-700 NoC, which the company says have already been selected by customers.

The significance of this news lies in the choice of scenario, not the announcement itself. BMC is the management chip on a server motherboard, responsible for status monitoring, power management, fault alerting, and remote operations. There is one per board, and its reliability and fault-tolerance requirements approach those of the main compute, though it does not carry business workloads. Using it as the first mass production landing point for an out-of-order core follows the same logic as automotive-grade chips entering through body control: positioned at the edge, with a high validation bar, but once you’re in, it’s long-term supply. StarFive’s own phrasing is “passing end customers’ stringent qualification validation.”

The figure “annual shipments expected to reach the tens of millions” should be discounted. Global server annual shipments are on the order of tens of millions of units, with typically one BMC per board. If ten million units were truly achieved, that would mean capturing most of the incremental market. This is “expected,” not an order, nor delivered volume. StarFive did not provide customer names, production scheduling cadence, or single-quarter shipments. The earlier official statement on November 17 only said it had “passed the product introduction and review processes of multiple server manufacturers and internet companies,” and that customer systems were “about to enter mass production.”

The position in the supply chain also needs to be clarified. Domestic RISC-V mass production cases in data centers are currently concentrated in “secondary core” scenarios such as BMC, edge gateways, and AI inference cards; there is no public case of dual-socket or quad-socket main CPUs replacing x86 at scale. What Dubhe-70 proves is that out-of-order cores can pass server-level validation processes, not that general-purpose compute has already been replaced. There are two tracking points: whether JH-B100’s quarterly shipments can be delivered, and whether Dubhe-83/100 have customer tape-outs.

Date: 2026-11-19 Sources: Mydrivers: Domestic out-of-order RISC-V CPU IP lands in data centers, News18a: Dubhe-70 first mass production, StarFive official site: Server BMC chip mass production and delivery, IT之家: Dubhe-70 announcement in August 2024.

I. Open Source Community and Software Ecosystem

RISE: Over 1,500 native riscv64 Python wheels, 519 still missing (11-18)

On November 18, the RISE project (RISC-V Software Ecosystem) published a progress report titled “Python on RISC-V: 1500 Packages and Counting.” The author is Ludovic Henry of Qualcomm. The report’s scope: RISE’s wheel dashboard tracks the 15,000 most-downloaded packages on PyPI. Of these, 2,076 contain native code and must be built separately for riscv64. As of November 18, 1,506 have riscv64 wheels.

The numbers need to be broken down. Of the 1,506, 117 are published on PyPI by the upstream projects themselves, and 1,389 are provided by RISE’s own index. Broken down by download ranking: of the top 100, 92 are available and 5 are missing; of the top 500, 438 are available and 50 are missing; of the top 1,000, 842 are available and 133 are missing; of all 2,076, 519 are still missing—a quarter of the total. In other words, for the vast majority of users to be able to install them, they must add --extra-index-url https://pypi.riseproject.dev/simple on the command line.

Among the packages already on the list are NumPy, SciPy, pandas, cryptography, PyArrow, PyTorch, and vLLM. PyTorch’s 2.13.0 wheels cover CPython 3.12, 3.13, 3.14, and the free-threaded 3.14 build, and were built on real RISC-V hardware rather than in QEMU or via cross-compilation, with all 212,038 test cases run to completion. vLLM’s 0.29.0+cpu was merged on September 21, with 0.30.0 following on September 24; the first release produced CPython 3.12 ABI3 manylinux_2_39_riscv64 wheels.

The most noteworthy item in the report is a comparison. On a StarFive VisionFive 2, building NumPy from source takes about 15 minutes, while installing a prebuilt wheel takes about 25 seconds. When a project’s dependency tree contains a dozen or so packages with extensions, this gap runs into hours. The report also gives an exit mechanism: once upstream starts publishing its own riscv64 wheels, the automation will propose retiring RISE’s build. The real metric for this progress is not how large the index is, but how many packages can be deleted because upstream has taken over.

Two caveats. First, this is an independent supply chain. The 1,389 packages depend on RISE’s build and release process; despite immutable releases and SHA-256 digests, responsibility is separate from that of each package’s upstream maintainers. Second, the target platform is glibc Linux and does not cover MCU-class RISC-V running MicroPython or bare metal. Being installable does not mean being performant: scalar implementations will run, but vector optimization still has to be done separately.

Date: 2026-11-18 Source: RISE: Python on RISC-V: 1500 Packages and Counting, Adafruit reprint.

openRuyi’s twenty commits: gRPC upgraded to 1.85.3 and relinked against system OpenSSL 4, protobuf to 36.2 (11-19)

The openRuyi main repository landed 20 commits between the evening of November 19 and the early hours of November 20. A group of five commits at 01:17 is the main event. gRPC was upgraded to 1.85.3 and relinked against system OpenSSL 4, and protobuf was upgraded to 36.2. protobuf-c was changed to build against the new Protobuf field API, re2 was upgraded to 2025.11.05, and abseil-cpp to 20260817.0. These five packages form a single dependency chain and were changed in one pass.

At 00:06 there were two GRUB changes: providing kernel-install integration, and packaging the default GRUB configuration file. The batch at 23:28 consists of 13 Go ecosystem packages, including the golang.org/x family of peripheral toolchains, gofumpt, misspell, go-billy, paho.golang, and several test helper libraries.

The switch of gRPC to system OpenSSL 4 deserves separate mention. gRPC ships with BoringSSL by default, and distributions often change it to link against system OpenSSL in order to unify TLS patches and FIPS policy. This change presupposes that OpenSSL 4 is already in the openRuyi repository; otherwise it would not link. In the same batch, protobuf and protobuf-c must be upgraded together, or the serialization layer will have an ABI mismatch. The ordering of these three commits shows that the packager understands the dependency direction.

The two GRUB commits point in another direction: making kernel installation follow the standard process. kernel-install is a hook provided by systemd; distributions use it to install the kernel, initramfs, and boot entry in one go. Packaging the default GRUB configuration file is so that users have a file to edit when changing boot parameters, rather than hand-editing grub.cfg. Changes like this do not appear in release notes, but they determine whether a distribution can be used as a daily-driver system.

Date: 2026-11-19 Source: openRuyi-Project/openRuyi, openRuyi official news page.

revyos/linux’s twenty commits: kernel moves to the Debian packaging layer, K3 gains a WRS.STO erratum (11-20)

The revyos/7.3.y branch of revyos/linux landed 20 commits in one burst between 01:29 and 01:33 on November 20, with timestamps concentrated within four minutes. Judging from the titles, this batch does two things: move the kernel build onto Debian’s packaging layer, and expand platform support to six vendors.

The packaging-layer changes include: adding debian_defconfig, a temporary scheme using the major version number as the package name, and having linux-image provide wireguard-modules. There is also a linuxboot configuration that switches the EFI framebuffer to efidrm and enables DRM panic support, and CI gains both cross and native builds. Six platform switches are turned on at once: ESWIN EIC7700, UltraRISC DP1000, SpacemiT K3 and K1, DAMO XuanTie TH1520, Sophgo SG2042/SG2044, plus a Picoheart erratum.

There are three substantive chip-related changes. One is riscv: errata: spacemit: Use WRS.STO instead of WRS.NTO on K3, replacing WRS.NTO with WRS.STO for the wait instruction on K3. Errata like this usually correspond to a specific behavioral difference in silicon; the difference between NTO (no timeout) and STO (short timeout) lies in the wait-timeout semantics. The second adds a 1.6GHz CPU operating point for the ESWIN StarPro64. The third enables fan control and the watchdog on the same board.

Landing 20 commits within four minutes matches the pattern of previous days: a batch push after cleanup, not 20 patches written that day. The direction readable from it is clear: revyos is pushing the kernel from “changing configuration per platform” to “maintaining it as a Debian kernel package.” If this step succeeds, following upstream stable releases will be much easier; the cost is maintaining a set of packaging patches. The K3 WRS.STO erratum is worth tracking separately—it is one of the few pieces of evidence that K3 is running a kernel on real hardware.

Date: 2026-11-20 Source: revyos/linux (revyos/7.3.y).

XiangShan’s three repos, ten commits: matrix and AME verification become the main line, GEM5 gains SMT and divider support (11-19)

On November 19, the XiangShan project landed ten commits across three repositories. The main XiangShan repository had two on the kunminghu-v3 branch. #6987 at 14:35 makes the DCache issue DiffRefillEvent separately per memory channel instead of queuing. #6947 at 09:31 sets the size of itlb/pmp requests in the ICache to 1.U.

All three commits in the difftest repository point to the matrix extension. #1050 at 17:12 parallelizes asynchronous CPU MMA verification; #1049 at 15:00 skips ROB scanning when there are no pending REF controls; #1044 at 14:45 tracks TR store requests and responses. The authors are yu-yake2002 and ecall73 respectively. Together the three are the same thing: the volume of verification data for the matrix unit is growing, and scanning item by item can no longer keep up, so it is being changed to parallel and conditional skipping.

The xs-dev branch of the performance model GEM5 had five. #1346 at 15:12 uses local write snooping to detect SMT read-after-write conflicts. #1355 at 14:58 aligns performance CI logs with perf triggers. #1319 at 12:50 adds teacher-student BOP optimization. #1327 at 12:45 implements wakeup for the integer and floating-point dividers, and #1324 at 10:26 fixes the calculation of integer division CLZ latency. The latter two are a pair: first fix the latency algorithm, then hook up the wakeup logic.

There are no performance numbers this day; it is all verifiability work. The ICache’s req.size and the DCache’s refill events are both details where “a mismatch between model and RTL causes false reports.” SMT RAR conflict detection is a new direction; Kunminghu’s SMT support previously had no dedicated check in the model. The matrix side is the most concentrated: the AME and matrix prefixes account for all three difftest commits, indicating that verification of the vector and matrix units is becoming the bottleneck.

Date: 2026-11-19 Source: OpenXiangShan/XiangShan (kunminghu-v3), OpenXiangShan/difftest, OpenXiangShan/GEM5 (xs-dev).

Hypnos: TU Delft’s two-phase continual learning platform, 33.3 mW for a complete RISC-V MCU (11-19)

On November 19, IoT Tech News reported on the Hypnos platform from Delft University of Technology (TU Delft), with a paper concurrently posted on arXiv (2611.24637). The authors are Guilherme Guedes, Marco P. E. Apolinario, and Charlotte Frenkel. The problem this platform addresses: letting edge devices continue learning new samples after deployment without sending raw data to a server.

Hypnos splits learning into two phases, corresponding to the brain’s wake and sleep. The wake phase runs on-device, centered on an X-HEEP RISC-V microcontroller plus a Quadrilatero matrix coprocessor, with a compact accelerator handling depthwise separable convolution, per-channel requantization, and prototype operations. This phase performs only gradient-free prototype updates: each incoming sample updates the class prototype once, while compressed features are stored. Backpropagation does not run on-device. The sleep phase is delegated to a smart charger: the charger retrieves product-quantization codes and weights from the device, decodes the hidden-layer activations, performs gradient updates via replay, and writes the consolidated parameters back.

Metrics per the paper: on the CORe50 NICv2-391 protocol, the INT8-activation configuration reaches 78.6% accuracy with 7.79 s per-sample on-device latency; the INT16-activation configuration trades up to 81.6% at 10.54 s. Compared with quantized hidden-layer replay, it needs 3.8× fewer learning steps and 22× less learning compute per sample, while achieving higher accuracy. The 28nm synthesis result is 9.43 mm², 33.3 mW at 375MHz, with the wake-phase accelerator occupying only 10.7% of the area. The 5.48MB replay buffer sits on the charger side and does not consume device memory. The entire design has been open-sourced.

What is worth noting is this resource-scheduling approach. It does not try to squeeze training into a battery-powered device; instead it moves the “consolidation” step to moments when wall power and large memory are available. The cost is that the device must be plugged in to complete long-term memory; the benefit is that on-device power and area are both pushed into the usable range. The limitations are also clear: 7.79 s per sample is only usable for slow-changing sensor streams, and camera-grade continuous learning is far off. The paper uses a frozen MobileNetV3-Large backbone plus an online prototype classifier, not end-to-end training.

Date: 2026-11-19 Source: IoT Tech News: Hypnos brings continual learning to RISC-V edge devices, arXiv:2611.24637, cogsys-tudelft/hypnos.

II. Product and Ecosystem Updates

2026 RISC-V Industry Development Conference Macau Session: Greater Bay Area Open Source Ecosystem Center Launched, Open Computing Platform Signed (11-19)

On November 19, the 2026 RISC-V Industry Development Conference Global Cooperation Special Forum and RDSA International Forum were held at the Galaxy International Convention Center in Macau under the theme “Standards Interoperability · Global Coexistence.” The event was guided by the Macau Economic and Technological Development Bureau and jointly hosted by the RISC-V Working Committee of the China Electronics Standardization Association, the RDSA Industry Alliance, the Guangdong-Hong Kong-Macao Greater Bay Area RISC-V Open Source Ecosystem Development Center, and the Macau Institute of Industrial Technology. The RDSA International Forum has been held for three consecutive years since 2024.

Three concrete outcomes emerged from the event. First, the Guangdong-Hong Kong-Macao Greater Bay Area RISC-V Open Source Ecosystem Development Center was officially launched, co-built by Zhuhai, Hengqin, and Macau, with three sub-units: a basic research platform, an application innovation platform, and an open source community. Second, the RISC-V Internationalization Supply-Demand Matching Platform was launched; according to Duo Jing, Secretary-General of the China Electronics Standardization Association, the supply side can publish products and solutions, while the demand side can publish application requirements. Third, the RISC-V International Open Computing Platform completed a strategic intent signing. Signatories include ESWIN Computing, Beijing Institute of Open Source Chip, Suzhou C*Core, BlueCore Computing, SpacemiT, as well as overseas institutions such as EverPure, ExaLeapSem, and 2V Systems.

The conference also released the first batch of RISC-V overseas expansion flagship products and solutions — nine in total — from Deepin Digital, SimComputing, Sophgo, LeapFive, Aowei Lingxin Technology, and Haoli Intelligent Technology. The 2026 DSA International Innovation and Entrepreneurship Competition was launched the same day, with a team track and an enterprise track covering three categories: DSA algorithms and applications, hardware acceleration, and DSA chip design. The prize pool is 1.15 million Macau patacas, and for the first time the judging criteria include “willingness and feasibility to establish operations in Macau.”

The industry data cited at the event: global RISC-V SoC shipments in 2025 were approximately 6.9 billion units, projected to reach 36 billion by 2031, a compound annual growth rate of 31.7%. These figures come from conference materials, with no third-party attribution, and are recorded as cited. Tang Dan, President of the Beijing Institute of Open Source Chip, said in an interview that the institute can provide technical support to Macau and jointly participate in international standards development and compatibility testing. The weight of such conferences should be judged by what is delivered: this time the deliverables were a center, a platform, and signing intents — organizational-level items — with no specific project list or funding arrangements yet visible.

Date: 2026-11-19 Source: Macau Convention & Exhibition Economy News: 2026 RISC-V Industry Development Conference Global Cooperation Special Forum Held in Macau, Macao Daily: RISC-V Forum Held in Macau to Deepen Global Industry Collaboration, Semiconductor World reprint.

Infineon Moves Its Flagship AURIX Family to RISC-V; riscv.org Publishes Five-Question Interview (11-19)

On November 19, the RISC-V International blog published a five-question interview with Thomas Schneid, Vice President of Automotive Microcontroller Software and Ecosystem Management at Infineon, on the launch of AURIX RC1. AURIX RC1 was released on November 4 and is Infineon’s first microcontroller family based on the RISC-V instruction set.

The background figures in the interview are worth noting: according to TechInsights, Infineon supplied more than one-third of the world’s automotive microcontrollers in 2025. AURIX is the family on which the company’s automotive business was built and the mainstay of its market position. RC1 merges the existing AURIX TriCore and TRAVEO T2G product lines into one generation, spanning from real-time controllers to MCU/MPU hybrid devices. Applications listed include zonal, body, chassis, battery management, powertrain, driver assistance safety management, and cockpit. Some models will include a parallel processing unit (PPU) for AI acceleration.

Infineon emphasizes migration cost. Schneid’s statement is that the company put substantial effort into minimizing both the engineering investment required for automakers to move to RISC-V and the impact on customer toolchains. The weight of this statement lies in the fact that the replacement cost of automotive MCUs is mainly not in the chip itself but in AUTOSAR configuration, debuggers, functional safety certification, and supply chain re-validation. That a supplier holding one-third of the market is willing to move its flagship family suggests it judges the ecosystem to be sufficient.

There is also considerable room for caution. Infineon explicitly said that specific model and product line details have not yet been announced, meaning there are no SKUs, no sample timelines, and no mass production milestones. Migration of existing AURIX customers will not happen overnight either; automotive-grade project design cycles typically run three to five years. Viewed alongside StarFive’s BMC mass production, the 2026 RISC-V industry signals are concentrated in the “management plane” and “control plane,” not yet in main compute.

Date: 2026-11-19 Source: RISC-V International: V Questions with… Infineon’s Thomas Schneid.

LILYGO T-CAN485 C5: ESP32-C5 Enters Industrial Control Boards, CAN and RS-485 on the Same Board (11-18)

On November 18, LinuxGizmos reported that LILYGO listed two development platforms. Among them, the T-CAN485 C5 is based on Espressif’s ESP32-C5, which is RISC-V architecture. The other, the T-Display CM0, uses the quad-core Arm Cortex-A53 Compute Module Zero and is not RISC-V; it is noted here only for distinction and not counted.

T-CAN485 C5 specifications: 16MB flash plus 8MB PSRAM; wireless support for 2.4/5GHz dual-band Wi-Fi 6, Bluetooth 5 LE, and IEEE 802.15.4 (Zigbee and Thread). The industrial interfaces are fairly complete. The SN65HVD251DR CAN transceiver reaches 1Mbps; the MAX13487E half-duplex RS-485 has automatic direction control; the W5500 Ethernet controller runs over SPI. There are also 0 to 10V analog inputs, 3.3 to 24V digital inputs, a microSD slot, and a Qwiic interface. The board measures 128×62×17 mm and is powered by 7 to 24V DC input or 5V USB.

The significance of this board is that it pushes the ESP32-C5 from “Wi-Fi 6 development board” into industrial field deployment. CAN and RS-485 are the two mainstays of production lines and building control; combined with 0 to 10V analog quantities and wide-voltage input, this indicates the target scenario is PLC peripherals or gateways rather than desktop experimentation. The ESP32-C5 is Espressif’s first RISC-V chip with dual-band Wi-Fi 6; as such boards proliferate, RISC-V’s shipment mix on the MCU side will shift from consumer to industrial.

Date: 2026-11-18 Source: LinuxGizmos: LILYGO CM0 terminal and ESP32-C5 board.

III. Deep Dive into Domestic Teams

SpacemiT’s llama.cpp Syncs with Upstream, Precision Enum Changed Along with Upstream (11-19)

On November 19, spacemit-com/llama.cpp’s mtmd-backend branch underwent a major upstream sync. At 14:36, #81 “Sync upstream llama.cpp master into mtmd-backend” was merged. At 13:17, a commit spacemit: use current precision enum after upstream sync landed, authored by timeuser4. The latter is a patch for the former: upstream renamed or changed the values of the precision enum, so the downstream backend had to follow suit.

Several other upstream changes came in the same batch. The OpenCL backend now allows adjusting KV padding in decomposed FA (#32522), and ggml added WASM Relaxed SIMD enhancements (#19590). There are also fixes for sticky NaN in loop state (#31913) and a switch of CI’s LLVM to 22.1.8 (#32548). These were not written by SpacemiT; they came in with the sync as upstream code.

This sync is worth noting because it exposes the structure of maintenance costs. The RISC-V backend modifies ggml’s tensor operators and memory layout, and upstream is rapidly refactoring this part. When a low-level type like the precision enum changes, all downstream backends must be recompiled and their semantics verified. SpacemiT’s approach is to maintain a long-lived branch and sync in bulk periodically, rather than rebasing commit by commit. The advantage is that merge conflicts are handled in a concentrated way; the downside is that each sync requires a batch of adaptation commits, and the distance from upstream periodically grows.

On the same day, spacemit-com had two other small changes: mpp fixed an issue where pkgconfig still pulled in C++ headers when the module was not enabled; ai-sdk and bianbu-agentic each updated their submodule pointers once. All three are maintenance work and do not constitute functional progress.

Date: 2026-11-19 Source: spacemit-com/llama.cpp (mtmd-backend), PR #81: Sync upstream master.

ESWIN StarPro64: 1.6GHz Operating Point, Fan, and Watchdog Added to Device Tree (11-20)

In the early hours of November 20, revyos/linux had three device tree changes targeting the ESWIN EIC7700 platform’s StarPro64: adding a 1.6GHz operating point to the CPU, enabling fan control, and enabling the watchdog. In the same batch, another commit enabled the master switch for EIC7700 platform support. The authors are Icenowy and RevySR.

The significance of these three changes is that they are all “only needed once on the board” changes. The CPU operating point means this chip has been verified to run stably at 1.6GHz, not just a number on the datasheet. Fan control indicates the board has active cooling and that the temperature curve needs to be described in the device tree. Enabling the watchdog relates to automatic recovery in unattended scenarios.

ESWIN Computing’s EIC7700 is its RISC-V edge computing SoC, and StarPro64 is the corresponding high-performance single-board computer. Previously, software support for this platform was mostly concentrated in the vendor’s own linux-stable and opensbi repositories, with the most recent push on October 19. This change landed in revyos rather than the vendor repository, indicating that third-party distributions are beginning to maintain this board as a formally supported platform. The practical meaning for users is: after installing revyos, the fan will spin, the frequency will rise, and a hang will trigger an automatic reboot.

Date: 2026-11-20 Source: revyos/linux (revyos/7.3.y), eswincomputing.

IV. Special Chapter on Embodied Intelligence and Robotics

SpacemiT Robotics Stack: Three Commits — Clock Anchor, Depth Hole Filling, Reference Motion Planner (11-19)

On November 19, three PRs were merged into humanoid_common in spacemit-robotics, addressing time baseline, depth map quality, and planner integration respectively.

At 03:15, #36 “feat(perception): add robot clock anchor service” was merged, authored by chenzhaoqiJM. It adds a read-only RPC service in the hardware driver that exposes IMU sampling timestamps and host receive timestamps via ZMQ, allowing external depth camera publishers to map capture moments onto the robot clock. The implementation runs on a background thread without blocking the control loop; on clock rollback, it invalidates old anchors and switches sessions.

At 07:57, #37 “feat(perception): simulate and fill depth holes” was merged, by the same author. It adds configurable depth holes, noise, and simulated latency to simulation, defaulting to 30FPS capture without catch-up frame interpolation; hole filling uses raw valid neighbors within a two-pixel radius, preserves capture timestamps, and discards stale frames from the renderer warm-up phase. #36 and #37 form a pair: the former provides the time baseline, the latter creates in simulation the defects only real cameras have, so downstream algorithms can handle them during the simulation stage.

At 08:18, #38 “feat(planner): support optional reference motion planners” was merged, authored by beililcj. The change introduces optional reference motion planners with request ownership, runtime availability checks, and status reporting, while upgrading the internal transport protocol from v7 to v8; joint and IMU feedback still use separate messages. The commit message documents verification: local package build, lint, and 10 planner, transport, and operator regression tests passed.

On the same day, ros2_humanoid had two supporting commits. At 11:45, “bridge RealSense depth to Hiking control” was added; at 16:08, “align Hiking depth preview.” omni-agent had one commit adding provider override for ASR and LLM /health startup wait.

The intent behind the three commits is clear. The robotics stack is integrating third-party perception and third-party policies, and doing so requires resolving two things first: how data from different sources aligns to the same timeline, and who has authority over control commands from different sources. #36 addresses the former; #38’s “request ownership” addresses the latter. Keeping joint and IMU feedback as separate messages stems from the same consideration — state feedback and command dispatch must not be mixed on a single channel.

Date: 2026-11-19 Source: humanoid_common PR #36: robot clock anchor, PR #37: depth hole filling, PR #38: optional reference motion planner, ros2_humanoid, omni-agent.

V. Summary and Trend Observations

The first judgment is that the data center management plane has become the first mass-production landing point for domestic out-of-order RISC-V cores. SiFive’s Tianshu-70 was an IP released in August 2024 as a “high-performance, ultra-low-power” core, and two years later it ships inside the JH-B100. The path chosen is right: one chip per BMC board, high reliability requirements, no business workload — a suitable place to validate the engineering maturity of an out-of-order core. What should be discounted is the “tens of millions of units shipped annually” figure — that is an expectation, not an order, and global server shipments themselves are on the order of tens of millions of units per year. The real observation point is whether quarterly shipments can be delivered, and whether Tianshu-83/100 has any customer tape-outs.

The second judgment is that ecosystem-side actions have shifted from “can it be installed” to “who maintains it.” Of RISE’s 1,506 wheels, 1,389 are hosted on its own index, and the report itself treats “taking RISE builds offline after upstream takes over” as a success metric. openRuyi switched gRPC from bundled BoringSSL to system OpenSSL 4, and revyos moved the kernel into the Debian packaging layer. Both point in the same direction: fold patches into the distribution’s unified maintenance surface rather than keeping a copy in each repository. This kind of work does not appear in release notes, but it determines whether upstream can still be followed three years from now.

The third judgment is that the leading supplier of automotive MCUs has begun to move its flagship family. Infineon holds more than one-third of the global automotive MCU market, and it is merging the AURIX and TRAVEO lines into a single generation of RISC-V products, emphasizing minimized migration cost. This is one of the heaviest architectural signals of the year, but there is currently no SKU and no sample timetable. Automotive-grade projects have design cycles of three to five years, so RC1 actually appearing in production vehicles will not happen until around 2029.

Looking at November 17 to 19 together, each of the three lines takes one step. On the management plane, SiFive officially announced mass-production delivery of the JH-B100 on November 17, and on November 19 added the core model and shipment expectations, moving from “delivered to customers” to “large-scale commercialization.” Over the same period, Zijin’s M100 laid out cumulative shipments and a customer list for its automotive-grade chips. Both are answering where RISC-V sits in the BOM of a complete machine, and the answer in both cases is the marginal but essential chip.

On the foundational software side, all three days’ work was packaging and configuration. On November 17, openRuyi pushed OpenSBI to 1.9 and introduced OpenSSL 3.6.5. On November 18, it completed security updates for curl and nginx and pulled in 14 Go packages. On November 19, it switched the gRPC chain to OpenSSL 4. revyos moved from the “configuration baseline” of November 18 to the “Debian packaging layer” of November 20. RISE, meanwhile, turned Python wheel coverage into a trackable table. Not one item in three days was a performance breakthrough — all were about maintainability.

On the platform and organizational side, the Macau meeting on November 19 launched the Greater Bay Area Open Source Ecosystem Center and the International Supply-Demand Matching Platform, and signed a strategic letter of intent with the Open Computing Platform. On the same day, riscv.org released its interview with Infineon. One side is building organizational structure, while the other is turning a leading supplier’s statement into a public text. Together, the two show that the focus in the second half of 2026 is “organizing what can already run,” not launching new architectures.

Appendix

Source Verification Table

Source Verification Result
Google News aggregation (24 query sets in Chinese and English, plus 10 additional Chinese team-name queries) Three substantive new items within the window: SpacemiT TianShu-70 mass production (November 19, first reported by Kuaikeji, republished by News18a and others), the Macau RISC-V Industry Development Conference (November 19, with Macao Daily and others publishing between November 19 and 20), and RISE’s Python wheels report (released November 18, circulated via Adafruit and others on November 19) — all included. Excluded: Spanish and Portuguese follow-ups on the Purism Librem 5 v2 (the November 18 report already covered it per the official press release), the Changsha Evening News visit piece on Zijing M100 (already featured as today’s highlight in the November 18 report), Alibaba CEO remarks on Qwen being trained on T-Head chips (the subject is PPU and the Yitian line, not RISC-V), multiple pieces on the OpenAI–AMD chip design collaboration (no RISC-V-specific information), Microchip RT PolarFire space-grade qualification (the subject is FPGA; RISC-V is only the core in its SoC sub-series), and the WCH ch32fun open-source stack (not a new project; a repost)
Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International One HN item within the window: a Show HN post on a RISC-V emulator booting 256-hart SMP Linux on an RTX3060 (1 point, 1 comment), plus a November 18 discussion on “an AI loop for RISC-V CPU design predating OpenAI’s open-source release” (9 points) — both personal projects, not listed as entries. CNX Software and LinuxGizmos each had one item: the former on the Waveshare ESP32-S31 display kit (ESP32-S31 is a new Espressif series; the product page gives no architecture details, so not included), the latter on two LILYGO boards (T-CAN485 C5 included; T-Display CM0 is Arm, excluded). One riscv.org blog item within the window, the Infineon five-question interview, included
GitHub organization and key repository checks (40+ repositories) Active within the window: XiangShan 2 commits, difftest 3, GEM5 5, openRuyi 20, revyos/linux 20, cvw several (Atom feed unavailable; no new commits on the main branch within the window), spacemit-com/llama.cpp 14 (including one full upstream sync), spacemit-com/mpp 1, spacemit-com/ai-sdk and bianbu-agentic 1 each, humanoid_common 4, ros2_humanoid 2, omni-agent 1. Zero pushes: NEMU, XSCache, cva6, core-v-verif, XUANTIE-RV (74 repos), eswincomputing (17 repos), sophgo, Radxa Linkr, RuyiSDK. The pushed_at timestamps for two Nuclei-Software repositories fall within the window, but there are no commits on the main branch; the pushes went to the develop and develop_n100 branches at 16:55 Beijing time on November 18, outside the window, so not counted this issue
openRuyi dual-source The official news page still shows the two October 19, 2026 posts and the 2026.09 release as the latest; no new version within the window. The 20 commits to the main repo are included per the log
riscv100.top inspection No changes: no new records or score changes across the six leaderboards, no additions to What’s New or Upcoming, no additions or removals of committee members
RISC-V member website inspection (391 sites) Changes were mainly on university and foundation homepages (Brandenburg University of Technology, China University of Petroleum (Beijing), CNCF, CMC Microsystems, ETH Zurich, European Space Agency, etc.); spot checks found only homepage news rotation and admissions announcements, with no new products, funding, or architecture releases, so not included. 74 sites failed to fetch, including the Institute of Software at the Chinese Academy of Sciences, Beijing Institute of Open Source Chip, Synopsys, and Tencent
RVSC2026 summit website check No changes: the official sitemap, 11 key pages, and source repository all show no updates; source HEAD remains October 20
Exclusion checks Alibaba CEO remarks on Qwen training on T-Head chips are excluded by subject matter; T-Head’s PPU and Yitian line are not RISC-V. EDA-related pieces from companies such as X-Epic and X-Epic (Xinhua Zhang) contained no RISC-V-specific content within the window. Headlines such as “OpenAI model tops RISC-V core tests” are restatements of AI-assisted chip design; the original text contains no RISC-V-specific benchmark data, so not included

Full Source List

Limitations: This issue covers a 24-hour window (November 19, 05:00 to November 20, 05:00). Of the three substantive media stories, the SpacemiT and Macau conference items are Chinese-language sources, and the Infineon item is an official riscv.org interview; technical details are based on commit titles, release notes, and paper abstracts, without item-by-item verification of code or silicon. SpacemiT’s claim that “annual shipments are expected to reach the ten-million level” is a vendor projection, unsupported by customer lists, production schedules, or quarterly data. TianShu-70 performance claims are inconsistent between the 2024 launch release and the 2026 mass-production release: the former benchmarks against A72/A510 and gives specific percentages versus A55, while the latter says it is close to A73. This report records both without choosing between them. Infineon’s AURIX RC1 has only a family positioning disclosed, with no SKUs, sample timelines, or production milestones; the migration-cost remarks in the interview are the company’s own account. RISE’s coverage figures come from its own dashboard, with no third-party review; the 519 “gaps” are counted under its own classification criteria. The 20 revyos/linux commits landed within four minutes as a batch push, and whether they share provenance with the batches from the previous two days cannot be determined from commit titles. Most of the 14 spacemit-com/llama.cpp commits came in via upstream sync; the only SpacemiT-originated change is one precision-enum adaptation. Hypnos is an academic result; its accuracy and latency on CORe50 are self-reported in the paper and have not been reproduced on third-party hardware. “Content changes” in the member website inspection are determined by page text hashes; 74 sites failed to fetch. The GitHub anonymous API was partially available this issue; organization-level statistics combine Atom feeds, and non-default branches and non-priority repositories may be undercounted.