RISC-V Daily Intelligence Report (2026-11-13)
Research window: 2026-11-12 05:00 ~ 2026-11-13 05:00 (Beijing time, 24 hours) Sources: Google News aggregation (24 query sets in Chinese and English, via proxy), Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International official site, openRuyi official site and main repo, RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep, plus commit and release searches on GitHub across organizations and key repositories including XiangShan/NEMU/difftest/XSCache, CORE-V CVW/CVA6, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, and Radxa Linkr.
Today’s Index
- Today’s Highlights: DAMO XuanTie releases XL-300 throughput-class interconnect IP, 8-core asynchronous connection with 32MB configurable L3 (11-12)
- I. Open-Source Community and Software Ecosystem
- XiangShan Kunminghu V3, three commits: ROB split into functional and debug buffers, FP issue queue wired into topdown counters, BJU adds isRVC metadata (11-12)
- openRuyi, five commits: LLVM 23 bumped to 23.1.5, dual-kernel patches and config refresh, fwupd support for gcab and OVMF merged (11-12)
- difftest, two commits: image decoding separated from RAM loading, tee pipe removed from gsim build (11-12)
- Espressif batch-pushes 22 repos: module libraries, video and voice components, chip errata updated in the same batch (11-12)
- II. Product and Ecosystem News
- Start9 launches RISC-V router: SpacemiT K1 eight-core, StartWRT distro, $300 (11-12)
- SingNova showcases full-stack RISC-V AI product line at Singapore security expo, desktop model rated at 256 TOPS (11-12)
- Baochip-1x: Bunnie Huang’s 22nm “mostly open-source” SoC, RV32IMAC single core at 350MHz (11-12)
- MIPS’ three Physical AI reference designs detailed: real-time control, 50 TOPS inference, and secure domain, one tier each (11-12)
- III. Deep Dive on Domestic Teams
- XuanTie XT-Link three-generation line: positioning evolution and spec verification from XL100 to XL200 to XL300 (11-12)
- SpacemiT llama.cpp 0.2.4: lead-2 triple-buffered F16 GEMM scheduling and manual release channel (11-12)
- Nuclei tool guide adds directory ignore rules; Sipeed SLogicWeb closes out frontend hints with 20 commits in a single window (11-12)
- IV. Embodied Intelligence and Robotics Special Section
- SpacemiT spacemit-robotics, two repos: vslam initializes stereo localization and mapping app, humanoid aligns with G1 navigation constraints (11-12)
- Compute tiering for Physical AI as seen through three reference designs: real-time cores, vector cores, and NPUs each in their place (11-12)
- V. Summary and Trend Observations
- Appendix: Full source list and source verification table
Today’s Focus: DAMO XuanTie Releases XL-300 Throughput-Oriented Interconnect IP, 8-Core Asynchronous Connection with Configurable 32MB L3 (11-12)
IT之家 reported on the morning of November 12 that XuanTie, Alibaba DAMO Academy’s RISC-V brand, has released XL-300, a high-throughput cache-coherent interconnect IP design. The specifications given in the report are: support for asynchronous connection of up to 8 cores, 256-bit data width, L3 Cache configurable up to 32MB, 8-core concurrent linearity of over 90%, and a bandwidth ceiling roughly double that of the previous-generation XL-200. The newly added Direct Memory Transfer support was singled out, with the stated purpose of bypassing the traditional path to break through the throughput bottleneck caused by large-capacity L3 and high bus latency. 驱动之家, 凤凰网科技, 新浪科技 and 搜狐 republished the same item on the same day, with identical content, all following the IT之家 account.
The significance of this news lies not in any single parameter but in the position of “interconnect IP.” RISC-V’s progress in single-core performance, vector and matrix extensions over the past few years is verifiable, but stitching multiple cores into a usable SoC depends on cache coherence protocols, cross-core bandwidth and latency, and L3 capacity and partitioning strategy—the part that the Arm and x86 ecosystems built into a moat through years of iteration, and also the area where domestic RISC-V server chips are most likely to expose problems post-silicon. XuanTie puts XL-300’s three metrics front and center (8-core asynchronous connection, 32MB L3, 90% linearity), pointing precisely at the most basic question: “can multiple cores actually deliver multi-core performance?” 8-core concurrent linearity of 90% means the marginal gain from the eighth core is still nine-tenths that of the first core—a figure that says more about the actual state of the coherence protocol and bus arbitration than peak bandwidth does.
The timeline needs to be laid out clearly to avoid reading an old product as a new one. XL-300 is not making its first appearance: in a March in-depth analysis of the C930, XuanTie’s technical lead had already listed it as the most powerful interconnect IP paired with the flagship server core C930, with the stated figures at the time being up to 8 cores per Cluster, L3 up to 23MB, a 20% frequency increase and doubled bandwidth versus XL-200, and only a 5% area increase. Further back, at the February 2025 XuanTie ecosystem conference, XL200 was still only an R&D plan to “provide larger-scale, higher-performance multi-cluster coherent interconnect.” In other words, XL-300 is positioned as the companion interconnect for the C930 generation of server-class IP, and this report brings a more complete public disclosure of specifications.
Two discrepancies in the figures are worth noting. First, L3 capacity: the March account said a maximum of 23MB, while this report says 32MB. The two may be different configuration tiers (single-Cluster versus multi-Cluster, or SRAM capacity under different process libraries), or the specification may have been raised over iterations; no verifiable datasheet has been officially released, so this publication records the two accounts side by side without letting one override the other. Second, this report uses “release” rather than “mass production” or “licensing secured,” which does not contradict the March statement that it was “already paired with C930”—IP public release, customer licensing and on-silicon validation are three different stages, and the report itself does not disclose the number of licensed customers.
For industry observers, the more valuable question is where this line goes next. XuanTie’s XT-Link series went from XL100 in July 2024 (the first cache-coherent multi-Cluster interconnect, scaling from 8 to 32 cores, supporting big.LITTLE) to XL200 and then XL300, at a pace of roughly one generation per year, with each generation’s focus shifting from “can it connect” to “how fast can it connect” and then to “throughput efficiency under large L3.” This path is complementary to C930’s SPECint2006 single-core performance at the 15/GHz level: the stronger the single core, the higher the bandwidth and coherence demands on the interconnect—otherwise multi-core merely moves the bottleneck from the execution units to the bus. For domestic teams building servers and high-performance SoCs, if they adopt XuanTie cores, the range of available interconnect IP and its licensing terms effectively determine whether they can build machines with 32 cores or more.
Date: 2026-11-12 Sources: IT之家: XuanTie Releases High-Throughput Interconnect IP XL-300, 凤凰网科技 reprint, 新浪科技 reprint
I. Open Source Community and Software Ecosystem
XiangShan Kunminghu V3: Three Commits — ROB Split of Functional and Debug Buffers, FP Issue Queue Wired into topdown Counters, BJU Adds isRVC Metadata (11-12)
The XiangShan Kunminghu V3 branch landed three commits within the window, spanning three directions. The first is structural governance of the reorder buffer (ROB): previously the ROB unconditionally stored a full copy of each enqueued uop per slot in the debug_microOp field; despite its name, exception handling and VLS handling also read it, leaving normal RTL dependent on a structure meant only for debugging. The change introduces RobSlotInfo, retaining only the three fields normal RTL actually needs (commitType, fuType, fuOpType) in slotUopBuf and stripping out the full debug copy. The payoff here is not in cycle counts but in area, timing, and the cost of future extension: a redundant wide vector had sat in the ROB for generations, and every added field had to be widened along with it.
The second wires the floating-point issue queue into the ROB’s topdown counters: FP issue activity and queue state are reported to the topdown collector, including the source of issue cancellations; integer and FP issue queue states are both fed into ROB head classification, and issued FP uops are counted toward the issue count; cancellation-source history is sized by all encoded sources and cancellation time is updated only when the ROB head is classified as an issue cancellation. The third is a fix: when JumpUnit creates a backend redirect on misprediction or backend fault, it zeroes the redirect bundle without assigning io.in.bits.ctrl.isRVC to redirect.isRVC, so isRVC is cleared to false in backend redirects for compressed jump instructions (such as c.jr), affecting subsequent instruction-length determination.
Taken together, the three form a typical “structural governance + observability” combination. The point of topdown counters is to turn “why is the pipeline idling” from a judgment call into attributable data: frontend fetch starvation, backend issue stalls, and ROB-full stalls each get their own counter. For an open-source core like XiangShan that must run SPEC and real workloads, how fast bottlenecks can be localized determines iteration speed; meanwhile work like the ROB split and the isRVC fix produces no performance numbers yet directly determines the blast radius of adding new features.
Date: 2026-11-12 Source: OpenXiangShan/XiangShan
openRuyi: Five Commits — LLVM 23 Bumped to 23.1.5, Dual-Kernel Patch and Config Refresh, gcab and OVMF fwupd Support Land (11-12)
The openRuyi main repository landed five commits within the window, all confined to the SPECS directory, spanning 9:13 to 11:27 AM. The first bumps llvm23 to 23.1.5, another patch-level follow-up on the LLVM 23 branch this month. The next two, by the same committer, refresh the patchsets and configs for linux and linux-lts respectively — note that “refreshing patchset” and “refresh config” are done together, meaning the batch of downstream-maintained RISC-V patches must be rebased onto the new baseline while kernel config options are realigned; such paired refreshes typically follow a stable kernel point release and are necessary for the distribution to stay buildable.
The fourth adds the gcab package. gcab is the GNOME camp’s Cabinet file (.cab) read/write tool — small in itself, but its appearance in a dependency chain usually means some upper-layer component needs to handle firmware or driver packages. The fifth confirms exactly that: ovmf gains fwupd support. fwupd is the firmware update daemon on Linux, and reaching the OVMF (UEFI firmware) layer shows openRuyi is making “updatable firmware” a distribution capability rather than leaving users to flash manually. For RISC-V platforms this is an easily overlooked but very practical requirement: board vendors’ firmware release cadence is out of sync with the distribution, and only an fwupd channel allows BIOS and firmware fixes to be pushed without reinstalling the system.
Taken together, openRuyi shipped no new release and no major features in this window; what it did was “keep pace with the upstream baseline and complete the firmware update chain.” This continues the line of its two October announcements (Ruyi Firmware offering an open UEFI firmware solution, and its positioning as ecosystem infrastructure): the firmware solution is about “how to boot openly,” while fwupd support is about “how to keep fixing things after boot.”
Date: 2026-11-12 Source: openRuyi-Project/openRuyi
difftest: Two Commits — Image Decoding Separated from RAM Loading, tee Pipe Removed from gsim Build (11-12)
Both difftest changes in the window are engineering-level. The first separates image decoding from RAM loading: previously, parsing the format and writing to memory were coupled in the same logic when loading an image; the change introduces an independent decoding stage so that “understanding the image” and “loading it into memory” are each testable and replaceable. For a differential testing framework, the value of this decoupling is that adding an image format no longer requires touching memory-layout code, and conversely adjusting the memory model does not affect format parsing.
The second fixes an easily overlooked build trap: the gsim generation rule contained gsim-gen-cpp ... | tee gsim-gen-cpp.log, and since Make does not set -o pipefail by default, the pipeline’s exit code comes from tee’s status and is always 0. As a result, when gsim-gen-cpp is killed by a bug or OOM, make still proceeds to invoke gsim-gen-emu and ends with a puzzling SimTop.h not found error. The change simply removes the internal tee, with a clear rationale: verilator.mk does not do this, and users who want logs can add a pipe on the outer command line themselves. This is the kind of commit that “saves half an hour for everyone who comes later, fixed once.”
Date: 2026-11-12 Source: OpenXiangShan/difftest
Espressif: 22 Repositories Pushed in Batch — Module Libraries, Video and Voice Components, Chip Errata Updated Together (11-12)
Espressif had 22 repositories pushed within the window, the densest organization-level push in the entire window. The list falls into several categories: host-side and cloud-side include esp-hosted-linux and esp-cloud-common; chip support includes esp-chip-errata, esp-phy-lib, esp-coex-lib, esp-hal-3rdparty, and usb-pids; wireless and Bluetooth module libraries include esp32s31-bt-lib and esp32h4-bt-lib; multimedia includes esp-video-components, esp-vision, and esp-gmf; frameworks and tools include esp-matter, esptool-js, idf-eclipse-plugin, esp-board-manager, esp-idf-security-dashboard, and homebrew-eim; plus esp-claw (2,379 stars), esp-ptp, developer-portal, and musl.
Same-day multi-repo pushes are routine for Espressif, reflecting release trains and upstream sync, and this newsletter does not break them down individually. Worth noting is that the media and Bluetooth lines are still expanding: esp-video-components, esp-vision, and esp-gmf moving at once indicates the on-device vision and audio chain is still filling in components; esp32s31-bt-lib and esp32h4-bt-lib appearing in the same batch shows new-chip support libraries are catching up. For teams building products on the ESP32-C series (C2/C3/C5/C6/C61, RISC-V cores), the update cadence of these libraries largely determines whether new modules can get usable firmware on day one.
Date: 2026-11-12 Source: espressif organization
II. Product and Ecosystem Updates
Start9 Launches RISC-V Router: SpacemiT K1 Octa-Core, StartWRT Distro, $300 (11-12)
Open-source hardware and software vendor Start9 officially launched its RISC-V router on November 12, built jointly with DeepComputing. The main controller is the SpacemiT K1 (octa-core 64-bit RISC-V), running the OpenWrt-based StartWRT distribution, priced at $300 (approximately RMB 2,006), bundled with a TP-Link four-port switch. Hardware configuration: 4GB LPDDR4 memory, 16GB eMMC storage, wireless via a replaceable miniPCIe module (AsiaRF AW7916-NPD, Wi-Fi 6E, concurrent 2.4GHz and 5GHz, external antennas), and external interfaces consisting of 1 gigabit WAN, 1 gigabit LAN, 2 USB 3.0 Type-A, 1 USB-C, and a microSD card reader.
Start9’s official page is very detailed about “how open it is,” which is more informative than the specs. The open parts include: the RISC-V instruction set itself, the K1 reference schematics derived from this board (already published with the documentation), the boot stack OpenSBI and U-Boot, the Linux kernel and mt76 wireless driver, and the MIT-licensed StartWRT; the closed parts are also named explicitly—the wireless RF firmware (the vendor does not provide open-source firmware; the module is removable and the device still runs without it), and two early-stage binaries executed once at power-on (DRAM initialization and first-stage boot), the latter of which is being replaced. This approach of “listing each part that cannot be open-sourced” is more credible in open-source hardware products than a blanket claim of “fully open source.”
The product positioning is also worth explaining: this router’s selling point is not throughput, but “you decide which tunnel each device takes”—per-profile routing, the ability to chain two VPN providers so that no single vendor controls both identity and destination, independent DoH resolution specified per configuration, and device isolation at the network layer. Such needs were previously handled by x86 mini-PCs running OpenWrt, with unsatisfactory power consumption and size; with an octa-core RISC-V, it becomes feasible to pack a home gateway, private cloud entry point, and network isolation into a single low-power box. For SpacemiT, this is a complete-product-level deployment case for the K1 in consumer network equipment, and it also shows that the K1’s software stack (OpenSBI/U-Boot/kernel/distribution) can already support third parties in building complete products, not just development boards.
Date: 2026-11-12 Source: Start9: RISC-V Router product page, IT之家: Start9 launches RISC-V router
SingNova Showcases Full-Stack RISC-V AI Product Line at Singapore Security Expo, Desktop Model Rated at 256 TOPS (11-12)
SingNova (南洋星科技), incubated at Nanyang Technological University, Singapore, exhibited a RISC-V-based product line at Safety & Security Asia 2026, held November 10–12 at Marina Bay Sands, comprising four products: an intelligent video storage server, the AI Power Studio desktop AI computing system, an AI digital employee, and a storage workstation. The official press release framed the theme as “Open Minds, Open Systems, Open Future,” anchored in the tension between video data growth and bandwidth, storage, and analytics costs in security and smart city scenarios.
The most specific specs are for the AI Power Studio: a 12-core RISC-V CPU plus a dataflow architecture, rated at up to 256 TOPS (INT8), up to 128GB LPDDR5 ECC unified memory, support for up to 128 channels of 1080p@30fps video decoding, PCIe 4.0 x16 expansion and high-speed network interfaces, targeting real-time video analytics, local LLM inference, generative video creation, robotics, and AI agents. The video storage server focuses on the encoding side: based on RISC-V architecture and chip-level intelligent video encoding, it cuts storage requirements to one-third (3:1 compression ratio) while preserving HD image quality and the visual metadata needed for downstream analytics, supports board-level redundancy, cluster failover, and automatic recovery, and is compatible with RTSP/ONVIF and H.265/H.264.
Restraint is warranted here: this is an expo press release, 256 TOPS is a vendor-rated figure with no test conditions given beyond the precision specification (batch size, model, whether sparsity is included), and the specific microarchitecture of the 12-core RISC-V is not disclosed. What can be confirmed is the form factor and scenario choice: making “multi-channel video decoding + local inference” into a desktop-class box, using RISC-V cores for control and pre/post-processing and a dataflow architecture for dense compute—this is the same division of labor that a group of edge AI vendors in China and abroad is converging on. For the RISC-V ecosystem, the value of such complete systems is that they turn the CPU from a “supporting role” into the system’s main controller, with the software stack maintained in-house rather than hanging off a PCIe card in an x86 host.
Date: 2026-11-12 Source: FinancialContent: SingNova Showcases RISC-V-Based AI Computing and Intelligent Video Storage Portfolio
Baochip-1x: Bunnie Huang’s 22nm “Mostly Open” SoC, RV32IMAC Single Core at 350MHz (11-12)
Hackaday reported on November 12 on Andrew “Bunnie” Huang’s Baochip-1x: a 22nm “mostly open” SoC targeting high-assurance applications. The design descends from the FPGA-based CPU in the Precursor device, whose selling point was a fully verifiable communicator design. The Baochip-1x includes an MMU, so any software supporting the RV32IMAC instruction set can run; a Coder’s Guide is provided, and the SystemVerilog and Verilog sources are on GitHub. Specs: a single-core 350MHz VexRISC-V, 4MB resistive RAM (RRAM) plus 2MB SRAM, along with a set of security-related features; the companion “Dabao” evaluation board is shaped roughly like a Raspberry Pi Pico.
The value of this news lies in the alternative product logic it presents. Most SoCs compete on performance and cost; the Baochip-1x swaps in “every transistor can be accounted for”: a verifiable design, public RTL, non-volatile RRAM replacing external flash (reducing physical attack surface and bill of materials), plus a mature open-source RISC-V core. Hackaday’s comment section brought up DEF CON badges and the old Chumby story, in a less-than-friendly tone, but for teams building high-assurance devices, chips with auditable designs are genuinely scarce—especially now, as supply chain scrutiny tightens, “you can read the RTL” matters more than “high benchmark scores.”
Date: 2026-11-12 Source: Hackaday: Baochip-1x: A Mostly-Open, 22nm SoC For High Assurance Applications
MIPS’ Three Physical AI Reference Designs Detailed: Real-Time Control, 50 TOPS Inference, and a Security Domain, One Each (11-12)
Electronic Design published a detailed breakdown of MIPS’ three reference designs on November 12, filling in specs for products previously disclosed only by platform name. Actus-1 is a module targeting real-time motor control and communications, built around two MIPS Atlas M8500 real-time cores (RMV23 profile) plus an ARC VPX 512-bit vector processing unit, using GlobalFoundries’ FDX FD-SOI process, with MRAM for on-chip non-volatile storage, targeting ISO 26262 and ISO 21434 certification; for field-oriented control it pairs 100ps-class high-resolution PWM with sigma-delta ADCs, gigabit Ethernet via a three-port switch with Time-Sensitive Networking (TSN) support, plus USB, CAN-FD, SPI, and I2C.
Acies-1 is an M.2 module focused on inference: MIPS Atlas S8200, RVA23 platform, with RVV vector extension and VME vector-matrix extension, capable of running models such as llama.cpp and Qwen as well as vision-language-action (VLA) models; 8 cores, 50 TOPS, over 50GB/s bandwidth, just 10W power, supporting integer and floating-point weights including FP4. The software side is supported by the Atlas SDK and a graph performance simulator, including an IREE- and MLIR-based compiler that can interface with TensorFlow, PyTorch, and ONNX. Aegis-1 is the central controller: a multi-core ARC-V RPX-500 running Linux and agent-type workloads, plus an Atlas M8500 security processor, an ARC RHX-100 security processor, and an Atlas S8200 NPU, supporting ISO 26262, ISO 21434, and the EU Cyber Resilience Act (CRA).
The division of labor across the three designs deserves separate attention: one handles determinism (motor control, TSN, safety certification), one handles throughput (50 TOPS, 10W, FP4), and one handles the system (Linux, security domain, whole-device orchestration). This is precisely the “compute map” that embodied AI system builders most lack—previously everyone used one big SoC to shoulder all workloads, and ended up struggling on all three fronts: real-time performance, power, and certification. By making the three tiers into composable modules and M.2 cards, MIPS essentially turns “how to allocate compute” into a selection problem. Note that this platform was already covered as an entry on November 9; what’s new this issue is the detailed reference design specs; MIPS itself, after being acquired by GlobalFoundries, has repositioned as a RISC-V IP supplier, and its products are outside the domestic supply chain, so its reference value exceeds its procurement value.
Date: 2026-11-12 Source: Electronic Design: RISC-V SoCs Zero in on Physical AI Applications
III. Deep Dive into Domestic Teams
XuanTie XT-Link Three-Generation Line: The Evolving Specs and a Spec Check from XL100 to XL200 to XL300 (11-12)
Lining up public information from past years reveals the iteration cadence of XuanTie’s interconnect IP. In July 2024, XL100 was released as “the first multi-Cluster interconnect IP supporting cache coherency,” positioned to give customers an 8-to-32-core processor scaling solution with support for big.LITTLE configurations; its selling points were reducing the bandwidth requirements, software programming difficulty, and extra cache accesses of multi-Cluster system interconnect, while optimizing system power. At the XuanTie Ecosystem Conference in February 2025, XL200 appeared in the R&D roadmap, described as “providing larger-scale, higher-performance multi-cluster coherent interconnect,” with no specific parameters at the time. In the March 2026 C930 deep-dive, XL-300 was already the strongest interconnect IP paired with C930, with specs of up to 8 cores per Cluster, up to 23MB L3, a 20% frequency increase and doubled bandwidth versus XL-200 with only a 5% area increase, and mentions of an elastically configurable architecture, capacity and bandwidth allocation, and an additional acceleration design for discrete GPUs on the same ID. The new specs given in this issue’s report are 8-core asynchronous connection, 256-bit width, 32MB L3, over 90% linearity at 8 cores, roughly doubled bandwidth versus XL-200, and a new Direct Memory Transfer.
Three details worth checking. First, L3 capacity going from 23MB to 32MB is a span of nearly 40%, which doesn’t look like a typo; it could be the difference between single-Cluster and multi-Cluster configurations, or it could be that the same product has different ceilings under different SRAM libraries or process nodes. Second, “20% frequency increase” appears in the March specs but not in this issue, while this issue adds “over 90% linearity at 8-core concurrency”—the former is a process and physical-implementation metric, the latter an architecture and protocol-efficiency metric; the two outlets each picked one side, suggesting the official materials themselves emphasize different dimensions on different occasions. Third, the phrase “8-core asynchronous connection” needs some explanation: asynchronous means each core can run in a different clock domain, which is necessary when big and little cores are mixed, or when cores and IO belong to different power domains, but it also raises the implementation complexity of the coherency protocol.
From a product-logic perspective, the competitor to the XT-Link line isn’t other vendors’ RISC-V interconnect IP, but the temptation for “customers to just build one themselves.” A team building a server SoC could, in theory, take four C930s plus a self-developed NoC and save on licensing fees, but they’d have to grapple with the coherency protocol, deadlock avoidance, and post-silicon debug themselves—extremely high risk. By making XL100, XL200, and XL300 into a gradient from 8 to 32 cores, XuanTie essentially turns the question of “how to stitch multiple cores together” into a multiple-choice question of buy or not buy. What will really determine the success of this line going forward isn’t posting another bandwidth number, but whether any third-party chip company publicly announces a mass-produced multi-core SoC based on XL-300—the value of IP is ultimately cashed out on someone else’s silicon.
Date: 2026-11-12 Source: Tencent Cloud Developer Community: Alibaba XuanTie C930 Deep Dive, Xinhua Net: XuanTie’s First Server-Grade CPU to Ship Next Month, IT之家: XL-300 Released
SpacemiT llama.cpp 0.2.4: lead-2 Triple-Buffered F16 GEMM Scheduling and a Manual Release Channel (11-12)
Three commits landed in the mtmd-backend branch window of spacemit-com/llama.cpp, in a very tidy rhythm. At 10:36, the version number was bumped from 0.2.3 to 0.2.4, with a restrained commit message: this commit only changes VERSION_NUMBER, doesn’t touch inference code, and prepares for the next release, accompanied by LLM validation results on K3 and separately attached multi-stream ASR reference measurements; the source commit corresponding to the LLM test is 13b6136. At 11:47, a manual release channel was added: after GitHub Actions recovered, it allows manually publishing a missed automatic release from Run workflow, adding a publish_release boolean input defaulting to false, retaining the manual run of “build only, don’t publish”; manual publishing requires the branch to be mtmd-backend, reads VERSION_NUMBER without depending on the SHA before the push, and backfills a missing version tag or skips if it already exists. At 16:55, the actual optimization for this version landed: ggml-spacemit adds lead-2 triple-buffered F16 GEMM scheduling, with co-authors from SpacemiT.
Together, the three commits form a complete release-engineering sequence: first set the version and validation baseline, then fix the reliability of the release channel, and only then the performance optimization. The lead-2 triple-buffered scheduling means adding another stage of pipeline buffering to the GEMM main loop so that data movement and computation overlap more fully—such changes usually don’t bring order-of-magnitude gains, but on a memory-bandwidth-constrained platform like K3, the degree of GEMM overlap directly determines the token rate of large-model inference. The “manual release” commit is equally worth noting: it shows the team encountered a situation where Actions being down caused a missed release, and rather than manually tagging after the fact, they made backfilling a first-class citizen. For downstream users, this means greater trust in version numbers and tags—no need to worry that some version “only exists in the build logs.”
Date: 2026-11-12 Source: spacemit-com/llama.cpp
Nuclei Tool Guide Adds a Directory Ignore Rule; Sipeed SLogicWeb Closes Out Frontend Polish with 20 Commits in One Window (11-12)
Nuclei (Nuclei-Software) had only one commit in the window: the .gitignore of the nuclei-tool-guide repo adds a SCRATCH directory, with the commit message “ignore SCRATCH directory for local miscellaneous files.” On its own it’s a trivial bit of engineering housekeeping, but combined with the repo’s continuous updates since October to chapters such as GDB live variable watching, it’s clear the tool guide line is still being maintained, and the maintainer is in the habit of separating local drafts from formal documentation to avoid accidental commits. The activity level of such repos reflects an IP company’s investment in customer support better than the SDK itself does.
Sipeed’s SLogicWeb landed 20 commits in the same window, almost all focused on closing out frontend interaction details: toolbar tooltips re-reconciled after the toolbar re-renders, tooltips suppressed in detailed tab mode, tooltips aligned with tab mode, and tests pinned for the geometry layout of the channel operation column; the times run from 4:26 PM all the way to 5:44 PM, constituting a concentrated round of UI polish. SLogicWeb is the browser-based host software for Sipeed’s logic analyzer; the UX bottleneck for such tools is often not sampling performance but “how to keep the toolbar and channel columns from getting messy when dozens of channels are open at once”—pinning the geometry layout with tests shows the team already treats the UI as an engineering artifact that needs regression protection, rather than something to tweak with CSS on the fly.
Date: 2026-11-12 Source: Nuclei-Software/nuclei-tool-guide, sipeed/SLogicWeb
IV. Embodied Intelligence and Robotics Special Section
SpacemiT spacemit-robotics two repos: vslam initializes stereo localization and mapping application, humanoid aligns G1 navigation limits (11-12)
spacemit-robotics had one commit each in two repositories within the window. ros2_spacemit_vslam was committed at 18:21, titled “spacemit-vslam: initialize stereo localization and mapping application” — from the commit semantics, this packages stereo-vision localization and mapping (SLAM) into a launchable ROS 2 application package, rather than just a set of algorithm libraries. For teams building robot bodies, this distinction is practical: an algorithm library requires writing your own launch files and configuring camera calibration and the TF tree, whereas an application package means switching to another machine only requires changing parameters. The stereo approach is especially meaningful on RISC-V, because it is cheaper in compute than LiDAR and more reliable than monocular, at the cost of needing sufficient vector compute for feature extraction and matching.
ros2_humanoid was committed at 16:47, fixing “aligning G1’s Nav2 limits with native gait.” Nav2 is the ROS 2 navigation stack, which has its own velocity and acceleration limit parameters; a humanoid robot’s native locomotion controller also has its own set of limits. When the two sets of limits are inconsistent, the typical symptom is that navigation plans a path that the low level cannot execute — either slipping in place or emergency-stopping. Fixes like this never appear in any demo video, but they determine whether a humanoid robot can continuously complete navigation tasks in a real venue.
Together, the two commits show that SpacemiT’s robotics line is moving toward “deliverable”: on one side, packaging the perception side into applications; on the other, fixing parameter consistency on the control side. The three-year embodied intelligence action plan issued in Beijing on November 11 wrote “mean time between failures exceeding 2000 hours” into its targets, and a considerable share of such failures come from repeated emergency stops and overheating caused by parameter mismatches, not hardware breakage.
Date: 2026-11-12 Source: vslam repo, humanoid repo
Compute tiering for Physical AI seen through three reference designs: real-time cores, vector cores, and NPUs each in their place (11-12)
Looking at today’s two product news items together, they happen to form a comparison table. MIPS’s three reference designs split embodied intelligence compute into three layers: Actus-1 uses two real-time RISC-V cores plus a 512-bit vector unit to handle motor control and TSN communication, emphasizing determinism; Acies-1 uses 8-core RVA23 plus vector matrix extensions to reach 50 TOPS at 10W, emphasizing throughput per watt; Aegis-1 uses multi-core ARC-V to run Linux and agent workloads, plus a safety processor and NPU, emphasizing system orchestration and certification. SingNova takes another path: a desktop machine packing 12-core RISC-V plus a dataflow architecture, rated at 256 TOPS, 128GB unified memory, and 128-channel video decoding, covering real-time analytics and local inference in a “one big box does it all” manner.
Neither path is right or wrong; it depends on the delivery form. For automotive and industrial control, certification and determinism come first — better to split work across three chips than let a single scheduling jitter on the Linux side affect the motor loop; for security and edge video analytics, camera density and total cost of ownership come first — one box handling 128 cameras is more convincing than three chips each being excellent. The real watershed is software: the three-tier approach requires vendors to maintain three toolchains and three sets of debugging methods, while the single-box approach requires heterogeneous scheduling on one chip to be mature enough — both paths pay a price in software, just in different places.
For domestic RISC-V supply, the lesson from this comparison table is “don’t use one chip to answer every question.” SpacemiT K1 takes the low-power general-purpose route (the Start9 router is its full-device implementation), Sophgo SG and BM series take server and edge inference, Nuclei N1000/NI1000 cover real-time cores and vector matrix cores, and XuanTie C930 plus XL-300 take the server-class multi-core route — every tier has a corresponding product; what’s missing are integrators who assemble these tiers according to full-device requirements and complete certification. The Beijing plan’s target of “key component matching rate exceeding 60%” tests exactly this kind of integration capability.
Date: 2026-11-12 Source: Electronic Design: RISC-V SoCs Zero in on Physical AI Applications, FinancialContent: SingNova product line
V. Summary and Trend Observations
1. Interconnect IP has moved from supporting role to the main line of RISC-V serverization. Among the specs disclosed for XuanTie XL-300 this issue, the most telling is not the 32MB L3, but “8-core concurrency linearity above 90%” — a metric only given by someone who has actually run multi-core workloads. Single-core performance relies on microarchitecture; multi-core performance relies on coherence and the bus; domestic RISC-V can already hold a conversation at the SPECint2006 15/GHz level on single-core, and the deciding factor over the next two to three years will fall on multi-core scaling efficiency and post-silicon stability. The one-generation-per-year cadence from XL100 to XL300 shows this line is being run as a product, not a one-off delivery.
2. Full-device deployment is shifting from development boards to consumer products. Start9’s RISC-V router at $300, with eight-core K1, spelling out item by item “which parts are open source and which are not,” is how a mature hardware product is written; SingNova turning 12-core RISC-V plus dataflow into a desktop box that can handle 128 cameras is likewise a full-device mindset. What both share is maintaining their own software stacks (StartWRT, a self-developed dataflow runtime), no longer depending on some x86 host. What RISC-V needs to prove is not “can run Linux,” but “can support a device that will be sold to ordinary users or integrators.”
3. The open hardware side branch is moving toward “auditability.” Baochip-1x uses 22nm, RV32IMAC, public RTL, and RRAM in exchange for “every transistor can be accounted for” — a small market but a real need. It does not conflict with the high-performance server route; rather, it fills in the piece most easily missing from the RISC-V narrative: in scenarios with increasingly strict security review and supply-chain traceability, being able to read the RTL is worth more than a high benchmark score.
4. Toolchains and engineering hygiene became the hidden thread of the day. difftest removed a | tee so that build failures no longer masquerade as SimTop.h not found; XiangShan stripped debug-only uop copies out of the normal RTL; SpacemiT added a manual channel for a version that was missed in release; openRuyi wired firmware updates into fwupd — none of these four produce performance numbers, but all reduce “the cost to be paid on the next change.” Whether an ecosystem is mature often depends on how dense these commits that never make it into release notes are.
Appendix
Source Verification Table
| Source | Verification Result |
|---|---|
| Google News aggregation (24 Chinese and English queries, 24-hour window) | 4 substantive new items within the window: XuanTie XL-300 launch, Start9 RISC-V router, SingNova security expo product line, MIPS reference design deep dive; plus one Baochip-1x item from a direct Hackaday RSS pull. The four items — ESP32-S31 running Linux, Qwenputer, Infineon AURIX RC1, and CH32’s RISC-V-only clang — are all second-hand recirculation of old news, already covered in previous issues, and are not counted again; the Chinese query “香山” returned 44 hits, all unrelated content such as Beijing Fragrant Hills autumn foliage, Xiangshan Bookstore, and Xiangshan Holdings, and the entire batch was discarded |
| Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International | HN yielded one hit within the window on AMD Zen 6’s RISC-V TEE discussion; the original was published 11-06 and already covered in this newsletter on 11-09, so it is not repeated; the current Phoronix RISC-V channel headlines are Linux 7.4 RISC-V optimizations (10-29) and SpacemiT V100 (10-19), both already covered, with no new items in the window; CNX and LinuxGizmos current content is unrelated to RISC-V; the latest riscv.org blog post remains the 10-20 SoCMake project snapshot |
| GitHub organization and key repository checks (30+ organizations and repositories) | Active within the window: espressif 22 repos, spacemit-com 6 repos, spacemit-robotics 3 repos, sipeed 3 repos; on the XiangShan side, XiangShan 3 commits, difftest 2 commits, NEMU 0 commits, XSCache and GEM5 not retrieved; openRuyi 5 commits, CORE-V CVW 1 commit, Nuclei-Software 1 commit. The DAMO XuanTie (XUANTIE-RV) organization page shows one push record at 11-12 20:20 within the window, but after checking the default branch of each repository individually, no commits were found within the window; this is judged to be a branch- or tag-level operation and is not included in the main text; ESWIN (eswincomputing)’s most recent push was 10-19, with zero pushes in the window; radxa-linkr had zero pushes in the window |
| riscv100.top inspection | No changes: the homepage What’s New, Upcoming, committee members, and the six SPEC CPU2006/2017/2026 leaderboards under the cross-architecture and RISC-V channels all show no new records or score changes; no space is allocated in the main text |
| RISC-V member website inspection (391 sites) | 212 unchanged, 110 reporting content changes, 69 failed to fetch (including ISCAS, Beijing Institute of Open Source Chip, Tencent, Synopsys, ADI, DENSO). All change items were checked one by one and found to be homepage marketing rotations or announcement updates (Huawei, Phytium, Stream Computing, NVIDIA GTC keynote, etc.), with no new products, funding, or architecture launches, so none are included |
| openRuyi official news page and main repository | The news page’s latest items remain the two from 2026-10 (ecosystem infrastructure, Ruyi Firmware) and the 2026.09 release, with no new version in the window; the 5 commits in the main repository have been included per the log |
Complete Source List
- [1] IT之家: XuanTie releases high-throughput interconnect IP XL-300 — https://www.ithome.com/1/025/902.htm
- [2] Phoenix Tech reprint: XuanTie releases XL-300 — https://tech.ifeng.com/c/8xzKvFpZpjv
- [3] Sina Tech reprint: XuanTie releases XL-300 — https://finance.sina.com.cn/tech/digi/2026-11-12/doc-iniymwrc1204054.shtml
- [4] Tencent Cloud Developer Community: In-depth analysis of Alibaba XuanTie C930 — https://developer.cloud.tencent.com/article/2642562
- [5] Xinhua Net: XuanTie’s first server-class CPU to ship next month — http://www.news.cn/tech/20250228/f4fd7f4a23744e4f9763ed156429cd6c/c.html
- [6] OpenXiangShan/XiangShan — https://github.com/OpenXiangShan/XiangShan
- [7] OpenXiangShan/difftest — https://github.com/OpenXiangShan/difftest
- [8] openRuyi-Project/openRuyi — https://github.com/openRuyi-Project/openRuyi
- [9] espressif organization — https://github.com/espressif
- [10] Start9: RISC-V Router product page — https://start9.com/riscv-router
- [11] IT之家: Start9 launches RISC-V router — https://www.ithome.com/1/026/074.htm
- [12] FinancialContent: SingNova showcases RISC-V AI product line — https://www.financialcontent.com/article/marketersmedia-2026-11-12-singnova-showcases-risc-v-based-ai-computing-and-intelligent-video-storage-portfolio-at-safety-and-security-asia-2026
- [13] Hackaday: Baochip-1x 22nm open-source SoC — https://hackaday.com/2026/11/12/baochip-1x-a-mostly-open-22nm-soc-for-high-assurance-applications/
- [14] Electronic Design: RISC-V SoCs Zero in on Physical AI Applications — https://www.electronicdesign.com/technologies/embedded/digital-ics/processors/article/55418566/electronic-design-risc-v-socs-zero-in-on-physical-ai-applications
- [15] spacemit-com/llama.cpp — https://github.com/spacemit-com/llama.cpp
- [16] spacemit-com/mpp — https://github.com/spacemit-com/mpp
- [17] spacemit-com/ai-sdk — https://github.com/spacemit-com/ai-sdk
- [18] Nuclei-Software/nuclei-tool-guide — https://github.com/Nuclei-Software/nuclei-tool-guide
- [19] sipeed/SLogicWeb — https://github.com/sipeed/SLogicWeb
- [20] spacemit-robotics/ros2_spacemit_vslam — https://github.com/spacemit-robotics/ros2_spacemit_vslam
- [21] spacemit-robotics/ros2_humanoid — https://github.com/spacemit-robotics/ros2_humanoid
Limitations: This issue covers a 24-hour window from 05:00 on November 12 to 05:00 on November 13. The effective news-side increments are concentrated in four items — XuanTie XL-300, the Start9 router, the SingNova product line, and the MIPS reference design — while the rest are second-hand recirculation of old news, removed per the deduplication rules. The XL-300’s L3 capacity shows a discrepancy between 23MB (per the 2026-03 C930 analysis) and 32MB (per this issue’s reporting); no official spec sheet is available for verification, so the main text records both figures side by side rather than adopting one over the other. On the GitHub side, the anonymous API hit the 60-requests-per-hour rate limit this issue, so organization and repository activity was verified instead via organization pages and each repository’s Atom commit feed, consuming no API quota; therefore, the determination of “pushes within the window” is based on default-branch commit times, and non-default-branch pushes (such as branch development at T-Head) may be underestimated. The XuanTie organization page shows one push at 11-12 20:20 but it could not be traced to a specific repository and branch, so this newsletter did not include it in the main text. SpacemiT’s weekly rls 2026.11.12 commits across three repos occurred before 11-13 01:14, i.e., before the window start, and were already covered in the 11-12 report, so they are not repeated this issue. The member website inspection (391 sites) took 232 seconds: 212 unchanged, 110 reporting changes, 69 failed; after checking each change item one by one, there were no substantive product or architecture launches, so no entries were included. Chinese-language queries such as “香山”, “玄铁”, and “芯来” remain dominated by tourism, stock market, and unrelated content, and were removed based on title semantics.