RISC-V Weekly Report (2026-11-09 ~ 2026-11-13)
Reporting period: 2026-11-09 to 2026-11-13 (five working days this week)
Part 1 · Brief
This section is the weekly highlight selection, by default within 800 words (787 words this issue); full entries, itemized lists, and sources are archived in Part 2.
-
DAMO XuanTie releases XL-300 interconnect IP: Disclosed on 11-12: 8-core asynchronous, 256-bit, L3 up to 32MB, 8-core concurrency linearity above 90%, bandwidth roughly doubled versus XL-200, and Direct Memory Transfer added; the March specification paired with C930 listed L3 up to 23MB, and both records stand side by side.
-
Nuclei NI1000 brings matrix extensions into RVV: Released on 11-09 as the first general-purpose AI compute core supporting VME, with VLEN configurable from 512 to 4096 bits and native FP8 support; the company claims that a 32×32 FP32 matrix multiply at VLEN 1024 increases area by only 15% and delivers 2.5x performance, positioned to divide work with NPUs.
-
Automotive-grade RISC-V completes IP, silicon, and toolchain in three days: AURIX RC1 was released on 11-04; on 11-09 it was reported that the core comes from MIPS IP and is fabricated by GlobalFoundries; on 11-10 HighTec and TASKING announced support on the same day, with compilers available in December and certification in 2027, aligning with Infineon’s sampling timeline.
-
Beijing writes embodied intelligence requirements into assessment metrics: On 11-11 it issued a three-year action plan: by 2028, complete-machine capacity of 200,000 units per year, key component localization rate above 60%, a 50,000 PFlops dedicated cluster, dexterous hands under RMB 10,000 each, and mean time between failures above 2,000 hours.
-
Distributions converge their platform matrices in the same week: RevyOS covered six platforms with a single debian_defconfig on 11-09 and landed 101 commits in a single window on 11-11; openRuyi upgraded its kernel to 7.3.4 and 6.18.63, glibc to 2.44, and moved into security backporting.
-
T-Head’s spin-off for a Hong Kong listing moves from rumor to announcement: On 11-09 it was reported that a 2027 listing would raise USD 2 to 3 billion; on 11-10 it was confidentially filed; on 11-11 Alibaba announced it and received HKEX confirmation that the split is permitted; the listing entity is PPU and Yitian, while XuanTie IP was transferred to DAMO in 2024.
This week’s judgment: Progress centered on turning existing capabilities into deliverables—interconnect IP answers multi-core efficiency, toolchains answer certification timing, distributions answer whether things compile and reach production, and policy answers what parameters procurement will use. What remains scarce is post-silicon validation and system integration, not another core with a higher peak.
Part 2 · Archive
The following are the complete entries and verification records, for reference and traceability.
Source material: Five RISC-V daily reports this week (2026-11-09 to 2026-11-13), of which 11-09 is a Monday 72-hour window (covering from 11-06 05:00), so some entries in the archive are dated earlier than Monday; 11-10 through 11-13 are each 24-hour windows. Additional cross-checks included a 168-hour reinforcement check of GitHub organizations and key repositories, a dual-source check of openRuyi, a site inspection of member company websites, and a riscv100.top inspection.
Editorial note: Entry facts, version numbers, and dates are taken from the original text of the above daily reports or from this round of hands-on verification; developments from the evening of 11-13 and the early hours of 11-14 (after the end of the previous daily report’s window) are marked separately, with sources being repository commit records. Entries with divergent specifications or vendor-claimed values are noted in the text, and neither has been used to override the other.
I. Weekly Highlights
DAMO XuanTie Releases XL-300 High-Throughput Interconnect IP: 8-Core Asynchronous, 32MB L3, 90% Linearity (11-12)
IT Home reported on November 12 that XuanTie, Alibaba DAMO Academy’s RISC-V brand, has released the XL-300 high-throughput cache-coherent interconnect IP design: it supports asynchronous connection of up to 8 cores, a 256-bit data width, an L3 Cache configurable up to 32MB, and 8-core concurrent linearity of over 90%, with a bandwidth ceiling roughly double that of the previous-generation XL-200, plus a new Direct Memory Transfer feature designed to bypass traditional paths and break through the throughput bottlenecks caused by large-capacity L3 and high bus latency. Phoenix Tech, Sina Tech, and MyDrivers republished the story the same day with identical content.
The weight of this news lies in the “interconnect IP” position. RISC-V’s progress in single-core performance and vector and matrix extensions is verifiable, but stitching multiple cores into a usable SoC depends on the cache coherence protocol, cross-core bandwidth and latency, and L3 capacity and partitioning strategy—precisely where domestic RISC-V server chips are most prone to post-silicon problems. 8-core concurrent linearity of 90% means the marginal gain of the eighth core is still nine-tenths that of the first, a figure that says more about the actual level 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. The XL-300 is not making its first appearance: in a March 2026 in-depth analysis of the C930, XuanTie’s technical lead already listed it as the strongest interconnect IP paired with the flagship server core C930, with figures at the time of up to 8 cores per cluster, L3 up to 23MB, a 20% frequency increase and doubled bandwidth versus the XL-200, and only a 5% area increase. This week’s report brings a more complete public specification disclosure, but the difference in L3 capacity from 23MB to 32MB has not been explained officially; it may be different configuration tiers, or specifications raised over iterations. This publication records the two figures side by side as independent accounts.
Date: 2026-11-12 Source: IT Home report, Phoenix Tech reprint, Sina Tech reprint
Nuclei Releases NI1000: First RISC-V General-Purpose AI Compute Core Supporting the VME Matrix Extension (11-09)
Nuclei Technology released the NI1000 RISC-V CPU IP on November 9, which the company says is its first high-performance general-purpose AI processor IP supporting the Vector Matrix Extension (VME). The NI series 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 while maintaining cache coherence, and the NICE and VNICE interfaces allow users to define custom scalar and vector instructions.
The technical point lies in how VME is implemented: VME is tightly coupled with the CPU and RVV within the same core, reusing the RVV vector register file for source operands and the RVV execution units for computation, 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 parallelism, and native FP8 support is added. Nuclei’s figures are: running a 32×32 FP32 matrix multiplication on the NI1020 (VLEN 1024), overall area increases by only 15% while matrix multiplication performance improves 2.5×.
The positioning is stated with restraint: the company makes clear that the NI series is not intended to replace accelerator IP such as NPUs and GPUs, but to combine with them. Compared with IME, which has independent matrix accumulation hardware and higher energy efficiency, and with AME, which eliminates a full independent matrix register file, VME sits in the middle, trading controllable hardware overhead for matrix compute, at the cost of lower peak energy efficiency than dedicated matrix units. The real barrier is not the IP but whether operator libraries and compilers can actually put VME to use.
Date: 2026-11-09 Source: Nuclei Technology press release, Design And Reuse report
Automotive-Grade RISC-V Completes Three Steps in Three Days: Product Definition, IP Source, Toolchain (11-09, 11-10)
After Infineon’s AURIX RC1 launched on November 4 (up to 12 cores, 40MB storage, ASIL D, sampling in the first half of 2027), two more links in the chain were filled this week. On November 9, a Design And Reuse reprint analysis pointed out that the RC1 uses MIPS processor IP and is fabricated by GlobalFoundries, while also raising the question of “whether RISC-V will replace Infineon’s own TriCore architecture”; Infineon’s official line has consistently been “the first RISC-V-based automotive MCU platform, with the TriCore family continuing to ship,” without elaborating on the IP source.
On November 10, two European automotive software tool vendors issued support statements on the same day, landing precisely on the opening day of electronica 2026 in Munich. HighTec’s approach is to fold the RC1’s device-specific compiler options into a single multi-architecture compiler, while providing safety- and security-certified C/C++ and Rust toolchains, with support extending from AURIX to TRAVEO and PSOC; TASKING, meanwhile, folds the entire lifecycle into one tool suite—VX-toolset for compilation, winIDEA and BlueBox for debugging, LDRA for verification—with built-in Agentic AI, on a timeline of compiler availability in December 2026 and toolchain certification completed in 2027.
Reading the three days of news together, the sequence for automotive-grade RISC-V is IP, silicon, toolchain, certification, with clear external signals at each of the four steps, and the two tool vendors’ certification timing aligned with Infineon’s sampling timing. For teams working on domestic automotive-grade RISC-V, what deserves attention in this round is not that there are two more compiler vendors, but that “certification timelines” are becoming the standard format of customer questions—who provides the functional safety evidence chain directly determines whether a project can get off the ground.
Date: 2026-11-09, 2026-11-10 Source: Design And Reuse analysis, HighTec press release, TASKING press release
Beijing Issues Three-Year Embodied Intelligence Action Plan: 200,000 Units Annual Capacity and a 50,000 PFlops Cluster (11-11)
The Beijing Municipal Bureau of Economy and Information Technology issued the “Beijing Embodied Intelligence Robot Industry High-Quality Development Action Plan (2026–2028)” on November 11, replacing directional language with assessable numbers: by 2028, cultivate no fewer than 5 world-class and 10 domestically leading humanoid robot OEMs; the Beijing-Tianjin-Hebei region’s OEM capacity to account for 25% of the national total, with key component localization exceeding 60%; form an annual OEM production capacity of 200,000 units in Beijing; and newly build embodied intelligence dedicated compute clusters with a total scale exceeding 50,000 PFlops.
The manufacturing-side targets are written hard: rated payload of 40% or more of the robot’s own weight, mean time between failures exceeding 2,000 hours, annual failure rate of core motion components and end effectors below 10%, design service life exceeding 5 years, continuous-operation battery life of more than 4 hours, and the cost of a single high-DOF dexterous hand reduced to under RMB 10,000. On the governance side, the plan proposes “one code per machine” full-process management, a “one-on-one” service mechanism for OEM listings, and explicitly opposes involution-style competition.
For the RISC-V line, the plan’s value lies in turning the demand side from a vision into procurement parameters: the joint side needs hard real-time, low power, and deterministic networking, while the body side needs compute and bandwidth capable of running VLA inference, and both must be mass-produced under a 2,000-hour failure-free constraint. Domestic supply that currently matches these two tiers includes the SpacemiT K3 series, the Sophgo SG and BM series, Nuclei’s N1000 and NI1000, and joint-level real-time cores such as the MIPS Atlas M8500. The plan itself names no instruction set or supplier; the above correspondences are a comparison against the current state of the supply side.
Date: 2026-11-11 Source: Wallstreetcn report, Beijing Business Today report (Sohu reprint)
Distributions Converge Their Platform Matrices in the Same Week: RevyOS Covers Six Platforms with One defconfig, openRuyi Kernel Bumped Twice (11-09 to 11-13)
The RevyOS linux repository did two things this week that differ in weight but point in the same direction. In a single push of 20 commits in the early hours of November 9, it added debian_defconfig, folding the configuration items of six platforms—ESWIN EIC7700, UltraRISC DP1000, SpacemiT K3 and K1, T-Head TH1520, and Sophgo SG2042 and SG2044—into a single defconfig, and added CI for both cross-architecture and native builds; on November 11, another 101 commits landed in a single window, filling in item by item the K3’s GPU node and display power domain, USB3 PHY, and UFS WriteBooster, the EIC7700’s fan controller and thermal zones, and the device trees for the SG2044’s SRD3-10-LB and SRD3-12 boards.
The openRuyi main repository saw four consecutive days of high-density rolling: on 11-08 the kernel was bumped to 7.3.3 and LTS 6.18.62, glibc to 2.44 with the full test suite run on x86_64; 11-09 to 11-10 brought twenty routine upgrades (wlroots 0.20.2, wget2 2.3.0, Python bindings for llvm22 through llvm24 switched to explicit linking against libclang); 11-10 moved into a security backport batch (policycoreutils 3.11, cups, lxc, and libmtp backporting upstream patches, libssh 0.12.2); and 11-11’s thirty-six commits closed out OpenTelemetry and four Python authentication packages together, bumped the kernel to 7.3.4 and 6.18.63, and restricted the XTHEADVECTOR and XLINUXMEMALIAS options to the rva20 platform.
Neither is doing new features; both are about “whether others can compile it and dare put it into production.” The long-standing problem for RISC-V distributions is one config and one patch set per board, so upgrading the kernel means redoing the adaptation; folding six platforms into one defconfig and bringing security fixes into the mainline lowers downstream costs more than shipping another release. The official website still shows no new version this week—the openRuyi news page’s latest entries remain the two from October 2026 (Ruyi Firmware, ecosystem infrastructure), and the November release has not yet shipped.
Date: 2026-11-09 to 2026-11-13 Source: revyos/linux repository, openRuyi-Project/openRuyi main repository, openRuyi official news page
T-Head Spinoff Heads to Hong Kong: From IPO Rumors to Alibaba Announcement and HKEX Confirmation (11-09 to 11-11)
On November 9, multiple media outlets citing Reuters reported that T-Head Semiconductor plans to list on the Hong Kong Stock Exchange in 2027, targeting a fundraising of US$2 billion to US$3 billion. On November 10, Hong Kong’s Hong Kong Economic Times further reported that T-Head had confidentially submitted a listing application to the HKEX; on the evening of November 10, Alibaba issued an announcement of its plan to spin off wholly owned subsidiary T-Head for a separate listing on the HKEX Main Board, with the HKEX having confirmed that Alibaba may proceed with the spin-off, and T-Head having submitted its A1 application confidentially as early as September 11. The announcement also left blanks: neither the offering size nor the reduction in Alibaba’s shareholding has been determined, and because the highest applicable percentage ratio is below 5%, the spin-off does not even constitute a disclosable transaction for Alibaba.
Valuation and business figures come in multiple versions: fundraising of US$2 billion to US$3 billion, with valuation ranges from US$20 billion to US$30 billion given by one institutional source, to JPMorgan’s rough January benchmarking estimate of US$25 billion to US$62 billion; on the business side, as of April 2026 the Zhenwu series AI chips had cumulative shipments of 560,000 units, serving more than 650 enterprise customers, with external commercial customers accounting for over 60%; the Zhenwu V900 has 216GB of memory, 1,200GB/s inter-chip bandwidth, can connect up to 500,000 chips into a cluster, and is scheduled for mass production in Q1 2027.
The relationship with RISC-V needs restating, because this framing has been repeated by the media for several rounds: the listing entity is T-Head, whose product focus is PPU (AI accelerator cards) and Yitian (Arm server CPUs); the XuanTie RISC-V processor IP was spun off from T-Head in 2024 and transferred to DAMO Academy, and does only RISC-V processors. Media roundups often lump in “XuanTie’s cumulative shipments of 6 billion units,” but that is a group-level historical figure and does not mean the listed assets include the RISC-V IP business. What really determines the valuation are two splits in the prospectus: how much of the tens-of-billions-scale annualized revenue comes from supplying Alibaba Cloud versus procurement outside the group, and whether internal settlement is at fair value.
Date: 2026-11-09 to 2026-11-11 Source: EqualOcean news flash, 10jqka reprint of Caiwen report, 36Kr report, Eastmoney reprint of Time Weekly
Three Purchasable Form Factors Emerge on the System Side: Router, Desktop AI System, and Open-Source SoC (11-12)
Start9 released a RISC-V router on November 12, built jointly with Deepin, powered by the SpacemiT K1 (eight-core 64-bit RISC-V), running StartWRT based on OpenWrt, priced at US$300, configured with 4GB LPDDR4, 16GB eMMC, a replaceable Wi-Fi 6E miniPCIe module, and one gigabit WAN and one gigabit LAN port. The official page spells out item by item “how open it is”: the K1 reference schematic, OpenSBI and U-Boot, the Linux kernel and mt76 driver, and the MIT-licensed StartWRT are all open; the wireless RF firmware and the two early-stage binaries executed at power-on are not open, with the latter being replaced. This way of “listing each part that cannot be open-sourced” is more credible than a blanket claim of full open source.
SingNova (Nanyang Star Technology) exhibited four RISC-V-based products at Safety & Security Asia 2026 in Singapore, among them AI Power Studio, a 12-core RISC-V plus dataflow architecture with a claimed maximum of 256 TOPS (INT8), up to 128GB of LPDDR5 ECC unified memory, and up to 128 channels of 1080p decoding; its video storage server emphasizes chip-level intelligent encoding, claiming to cut storage requirements to one-third of the original. What needs restraint is that this is a trade-show press release, 256 TOPS is a vendor-claimed figure without test conditions such as batch size, model, or whether sparsity is included, and the specific microarchitecture of the 12 cores is not disclosed.
On the same day, Hackaday reported on Bunnie Huang’s Baochip-1x: a 22nm-process, “mostly open” SoC for high-assurance applications, with an MMU and capable of running RV32IMAC software, specified as a single-core 350MHz VexRISC-V, 4MB resistive memory plus 2MB SRAM, with SystemVerilog and Verilog source open, and a companion “Dabao” evaluation board shaped much like a Raspberry Pi Pico. Its competitive dimension is not performance and cost, but “every transistor can be accounted for”; in scenarios with tightening supply-chain scrutiny, “being able to read the RTL” matters more than a high benchmark score.
Date: 2026-11-12 Source: Start9 product page, IT Home report, SingNova report, Hackaday report
II. Instruction Set and Architecture Ecosystem
XiangShan Kunminghu V3: 18 TLB Refactoring Commits in One Pass, Debug Fields Stripped from ROB (11-10 to 11-12)
The kunminghu-v3 branch of the main XiangShan repo saw activity on each of three days this week. On 11-10, five commits: changed the VBTB prediction selection logic in MBTB to taken-first, fixed NewCSR not re-fetching when a CSR write changes breakpoint gating, preserved ROB pending-slot reset during redirect cleanup, aligned CBO invalidate events to cache line granularity, plus one codeowners maintenance commit. On 11-11, twenty-one commits, of which 18 were a refactor(TLB) series with the same day and same timestamp, none changing behavior — all structural governance: naming fields for TLB read results and PTW bypass results, sharing dynamic VPN index selection, extracting permission and PBMT checks into the TLB source, aggregating access-check information by port, and removing unused scaffolding such as blocking TLB support and the saveLevel configuration. On 11-12, three commits turned to observability: introduced RobSlotInfo, keeping only the three fields commitType, fuType, and fuOpType in normal RTL while stripping the entire debug uop copy; wired up the topdown counters for the floating-point issue queue and ROB; and fixed JumpUnit not assigning isRVC on backend redirect, which caused incorrect length determination for compressed jump instructions.
All three threads point to the same judgment: being able to do a “structure-only, behavior-unchanged” batch refactor of a module presupposes a sufficiently dense regression surface and stable behavior. TLB and PTW are the hardest-to-debug stretch of the memory access path; converging permission checks and bypass results to a single exit will shrink the change surface considerably when adding extensions like Svnapot and Svpbmt later. The topdown counters, meanwhile, turn “why is the pipeline idling” from a matter of experience into attributable data.
Date: 2026-11-10 to 2026-11-12 Source: OpenXiangShan/XiangShan, OpenXiangShan/difftest
NEMU: Matrix Extension Moves from “Semantic Completion” to “Memory Access Fast Path” (11-10, 11-11)
NEMU had six commits across two windows this week, splitting neatly into correctness and performance groups. On 11-10, fixed riscv64-side writes to SIP latching timer interrupts (#1242), a detail error on the interrupt virtualization path; the same day AME switched to CRC32C for computing the matrix completion hash (#1280). On 11-11, four commits: added a PF refresh enable bit to SMBLOCKCTL (#1285), turning this control bit from reserved into configurable behavior; fixed on the RVV side zero-length gather, slide, and mask-class operations not setting the dirty VS bit (#1286) — a missed vector state dirty bit would cause the host to misjudge whether context needs saving in virtualization scenarios; fixed on the MMU side that a non-canonical address on the second page should report the next page base; and on the performance side enabled fast paths for matrix load and store in the shared ref (#1284).
AME is the XiangShan-family matrix extension, and over the past year NEMU’s support for it has gone from completing instruction semantics to optimizing the memory access path — indicating the matrix unit has run enough in the simulator that performance has become the bottleneck. The same day, two supporting commits on the difftest side: removed the latest tag from the Firtool release in the nightly workflow (#1040), and vectorized the MMA CPU check vectors while reusing already-decoded operands (#1038). The focus of verification infrastructure is shifting from “is it functionally correct” to “does it run fast enough.”
Date: 2026-11-10, 2026-11-11 Source: OpenXiangShan/NEMU, OpenXiangShan/difftest
CORE-V CVW: After Removing Two Local Branch Predictors, Turning to Verifiability Closure (11-07, 11-09)
CORE-V Wally (cvw) merged 20 commits in a batch on 11-07, in three groups: removed the BP_LOCAL_AHEAD and BP_LOCAL_REPAIR local branch predictors and the BTB_ADDR_BITS parameter; fixed centered PWM sticky and deglitch behavior, APB read strobe, and AHB/APB bridge tests; documented RV32GC CoreMark/MHz of 4.33 in the README and extracted the common CoreMark compile options for reuse. On 11-09 to 11-10, 11 more commits merged, with the main thread becoming verifiability: added LSR error-bit behavior to the 16550 UART, made SPI read back zero when the receive FIFO is empty, added exclusions to branch-prediction redirect coverage statistics, gave lockstep mode an exemption rule for unaligned accesses, and removed the SHA Questa script and Docker regression script.
Removing predictors is unusual on a teaching/research-oriented core; it usually means the team is moving toward “simple structure, easy timing closure, explainable results” rather than chasing IPC. Combined with the CoreMark/MHz 4.33 published the same week and this round of protocol detail closure, CVW is putting both “reproducible scores” and “synthesizable RTL” on the table at once — more practical for teams using it for teaching, pre-tapeout verification, and CORE-V ecosystem chip prototypes.
Date: 2026-11-07, 2026-11-09 Source: openhwfoundation/cvw
Toolchain and Debug Visibility: Instruction-Level Trace, RISC-V-only clang, and LLVM Python Bindings (11-09 to 11-12)
Espressif pushed esp-riscv-trace to 0.2.0. This tool decodes hardware instruction traces captured on Espressif RISC-V chips, reading captures produced by OpenOCD or a core ELF with trace annotations, reconstructing program execution flow, resolving each PC to a function and source line, and outputting hotspot tables, pre-crash call history, or call stacks directly feedable to flame graphs, marking continuity losses as gaps rather than guessing to fill them. 0.2.0 adds four commands — decode, profile, stats, history — pushing this capability from “can capture” to “can analyze,” supporting ESP32-C5, C6, C61, H2, H21, H4, P4, and S31.
The community-organized CH32 RISC-V User Group released ch32rv-clang: a clang and lld built from upstream LLVM with only the RISCV backend retained, packaged as one archive per host system for ArduinoCore-CH32RV to install via the board manager, supporting WCH Qingke’s XW compressed instructions across five host platforms. It deliberately subtracts — not bundling the C library or compiler runtime — in exchange for an unzip-and-use install path, eliminating half of the most common deterrent: “can’t get the toolchain installed.”
On the openRuyi side, a class of “installs but doesn’t work” problems was fixed: the llvm22, llvm23, and llvm24 packages use sed to change clang’s Python bindings to link against libclang, preventing users on RISC-V from failing at the import stage when installing llvmlite or triton; on 11-12 llvm23 was bumped to 23.1.5. Nuclei’s nuclei-tool-guide had 12 commits in a single day breaking down known issues in the 2026.10 release by component, and on 11-11 added documentation for GDB live variable watch — debug experience has long been a weak spot in the RISC-V MCU ecosystem, and updates like these retain customers better than shipping new cores.
Date: 2026-11-09 to 2026-11-12 Source: espressif/esp-riscv-trace, Adafruit coverage, openRuyi-Project/openRuyi, Nuclei-Software/nuclei-tool-guide
Teaching and Open-Source Cores: Specula’s Out-of-Order Core, a Computer on Tang Nano 9K, Baochip-1x (11-06 to 11-12)
Developer Pranav published a long post on 11-06 about Specula, a dual-issue out-of-order RV32I processor he implemented from scratch in Bluespec SystemVerilog, breaking down register renaming and retirement, reservation stations and issue, store forwarding and speculation, and how comparison trees and register ports squeeze timing, item by item. The value of such content isn’t in code volume but in pulling the out-of-order core’s barrier out of the black box of papers and commercial IP — teams in China doing RISC-V teaching and pre-tapeout verification have long lacked an intermediate textbook “from in-order to out-of-order.”
The m9b project that appeared on Hacker News on 11-09 is another kind of layered sample: a complete small computer built on a roughly twenty-dollar Sipeed Tang Nano 9K using the VexiiRiscv soft core — 36MHz, 8MB PSRAM, 720×480 HDMI, 48kHz audio, two USB ports for keyboard and mouse, SD card for program storage, capable of running Doom (timedemo at about 22 fps), Descent (about 10 fps), Tyrian 2.1 (full speed at 35 fps), and five other DOS-era games. The project splits the SoC, runtime, five game ports, and SD card image into separate repositories, all open source, each documenting where it modified upstream.
Together with Baochip-1x on 11-12, the three threads show that RISC-V’s “non-high-performance” branch still has real demand: reproducible teaching cores, FPGA projects that make clear “what a computer minimally needs,” and high-assurance SoCs with auditable designs. None of these three produce benchmark scores, but they determine how newcomers and new scenarios enter this ecosystem.
Date: 2026-11-06 to 2026-11-12 Source: Specula long post, m9b-system repo, Hacker News discussion, Hackaday coverage
III. Operating Systems and Distributions
openRuyi: Two Kernel Bumps, Security Backports and Firmware Update Chain Filled In (11-08 to 11-13)
The openRuyi main repo followed a typical rolling-release cadence this week. On 11-08, ten substantive commits: the mainline kernel package moved to 7.3.3 and the LTS package to 6.18.62, both with refreshed patch sets and configs; glibc was bumped to 2.44 and ran the full test suite on x86_64, plus one commit adding the x86_64 architecture macro to rpm-config-openruyi. From 11-09 to 11-10, twenty commits returned to routine upgrades: desktop stack wlroots 0.20.2, networking tool wget2 2.3.0, Python-side updates to httpx2, httpcore2, idna, and langgraph-sdk, and python-pymongo now bundles BSON into the main package.
On 11-10 the focus shifted to security: policycoreutils was bumped to 3.11 with a backported fixfiles fix, lxc, libmtp, and cups each received upstream security patches (all commits tagged security-fixes), and libssh moved to 0.12.2. On 11-11, thirty-six commits closed out two supply chains at once: the entire python-opentelemetry family moved to 1.45.2 with two new OTLP exporter subpackages, and four auth packages — python-pyjwt, python-oauthlib, python-joserfc, and python-authlib — were upgraded the same day under security updates; the kernel also moved to 7.3.4 and 6.18.63 that day, the three OpenCL packages (headers, icd-loader, clinfo) entered the tree, and qemu-rustsbi was brought into the repo.
On 11-12, five commits wrapped up: llvm23 moved to 23.1.5, the linux and linux-lts patch sets and configs were refreshed in pairs, a new gcab package was added, and OVMF gained fwupd support — a step that makes “firmware updatable” a distribution capability, so board vendors’ firmware fixes no longer require reinstalling the OS. The official news page remains stuck on two posts from October 2026; the November release has not shipped, so the above counts only as repo-side progress.
Date: 2026-11-08 to 2026-11-12 Sources: openRuyi-Project/openRuyi, kernel-team-tools, openRuyi official news page
openRuyi Friday Evening Addendum: Kernel Tag 7.3.4-3 and the Ruyi Firmware Wave (11-13 evening to 11-14 early morning)
The previous briefing window closed at 11-13 05:00; activity after that up to press time is covered by this 168-hour repo check, recorded in Beijing time. Between 11-13 22:01 and 23:04, the openRuyi main repo landed the new python-torch-ruyiai package, cmake 4.4.5, and clang-wrap 20261113, and enabled PIE and generated clang resource directories for llvm22 through llvm24; from 22:55 the same day, the linux branch received SpacemiT K3 errata fixes (replacing WRS.NTO with WRS.STO), EIC7700 SD card clock and detection fixes, and P550 fan control and thermal zone nodes.
Tags and firmware closed out after midnight: from 11-14 00:00 to 00:16, kernel-team-tools pushed linux to v7.3.4-3 and linux-lts to v6.18.63-4, registering 7.2.15-5 and 7.0.14-11 in the manifest alongside, and on the config side switched r8169 and the Realtek PHY to modules and enabled LED support; ruyifirmware pushed continuously from 00:47 to 01:25, covering unified use of Linux ruyi/v7.3.4-3 across all boards, updated PLL1 definitions for K3 OpenSBI, integrated emulator fixes and a readable text console, a RAM recovery flow based on OpenOCD 0.12.0 for the HiFive Premier P550, and streamlined UART and JTAG recovery instructions — with pushes appearing simultaneously across six repos: edk2, opensbi, coreboot, edk2-ruyi, MultiArchUefiPkg, and unicorn-for-efi. This is a coordinated integration push following October’s Ruyi Firmware announcement, aimed at making both “firmware boot” and “firmware recovery” reproducible processes.
Note that the above tag and firmware actions landed in the early hours of 11-14, strictly outside the Monday–Friday cycle, and are a pre-press addendum; the main repo and kernel repo commits remain on the evening of 11-13 and count toward this week. Taken together, Friday night’s work centered on tagging the kernel and adding recovery paths to firmware, not on new features.
Date: 2026-11-13 evening to 2026-11-14 early morning Sources: kernel-team-tools, ruyifirmware repo, openRuyi-Project/openRuyi
RevyOS: From Six-Platform defconfig to 101 Board-Support Commits in a Single Window (11-09 to 11-11)
RevyOS climbed three steps in three days this week. On 11-09, twenty commits: a new debian_defconfig unified six platform configs, cross-arch and native build CI was added, and driver fixes landed for SpacemiT UFS entering FASTAUTO before switching to FAST mode, DRM shmem GEM mapping switched to writecombine, and USB power retained during handoff on T-Head Lichee Pi 4A. On 11-10, six commits: configs for the K3, K1, DP1000, and EIC7700 platforms entered the tree separately, plus a linuxboot config and a dual-build CI job — a batch closeout of the platform matrix. On 11-11, 101 commits in three chronological batches: platform features and FROMLIST backports (K3 GPU nodes and display power domains, Type-C Port A, USB3 PHY and the FUSB301 autonomous controller, PCIe asserting PERST# only after PHY init, UFS WriteBooster; EIC7700 hwmon fan controller, PWM, thermal zones and DVFS with max frequency set at 1.4GHz; P550 USB, SATA, PCIe, eMMC, and SD completed), general kernel changes (the full set of page-table accessor FROMLIST patches backported, KVM exposing all counters to VS-mode), and defconfig closeout plus packaging-side HACKs.
The SG2044 board line deserves separate mention: device trees for both the SRD3-10-LB and SRD3-12 boards entered the tree, with SRD3-12 also carrying a large-BAR variant, a new second C2C1 x4 PCIe port and eFUSE controller, and PCIe windows aligned with the EDK2 implementation. For teams building RISC-V servers, consistent board descriptions and firmware conventions matter more than an extra driver — inconsistent window partitioning shows up directly as incomplete device enumeration in multi-card scenarios.
Date: 2026-11-09 to 2026-11-11 Source: revyos/linux
Other Distributions and Firmware: No Substantive openEuler Updates This Week, qemu-rustsbi Enters the Tree (11-09 to 11-13)
Within the coverage of this week’s five daily briefings, openEuler had no verifiable version releases, kernel, or toolchain changes in the window, so no entry is expanded this week; this observation only means no substantive updates were seen on the briefing’s sources and does not imply the project is dormant. Two other distribution-side items are worth noting: openRuyi brought the Rust-written SBI reference implementation qemu-rustsbi into its repo, meaning multilingual firmware implementations are now being taken seriously at the packaging level; and RevyOS and openRuyi are both adding dual-build CI, with both build baselines moving toward “reproducible.”
Date: 2026-11-09 to 2026-11-13 Sources: openRuyi-Project/openRuyi, revyos/linux
IV. Chips and Hardware
XuanTie XL-300 and the XT-Link Three-Generation Line: From “Can It Connect” to “Throughput Under a Large L3” (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 as a processor scaling solution for 8 to 32 cores with support for big.LITTLE configurations; 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 in-depth analysis of the C930, XL-300 was already the strongest interconnect IP paired with the C930, stated as up to 8 cores per Cluster, L3 up to 23MB, a 20% frequency increase and doubled bandwidth versus XL-200, with only a 5% area increase; this issue’s report gives 8-core asynchronous connection, 256bit, 32MB L3, linearity above 90%, doubled bandwidth, and Direct Memory Transfer.
Three points warrant verification: the L3 capacity spans from 23MB to 32MB—nearly 40%, unlikely a typo, possibly a difference between single-Cluster and multi-Cluster configurations or an upper limit under different SRAM libraries; “20% frequency increase” appears in the March statement but not in this issue, while this issue adds “8-core concurrent linearity above 90%”—the former is a process and physical implementation metric, the latter an architecture and protocol efficiency metric; “asynchronous connection” means cores can run in different clock domains, which is necessary when big and little cores are mixed and cores and IO belong to different power domains, but it raises coherency protocol complexity.
Organizational corroboration: a 168-hour check shows the XuanTie organization (XUANTIE-RV) had only two pushes within the window, qemu (11-10) and the newly appeared zvfx-proposal repo (11-12 20:20). The latter is dated the same day as the XL-300 report, but its default branch has no readable commits or description, so no technical interpretation is offered; product-level new releases this week were zero, and XL-300 is an external specification disclosure rather than new silicon.
Date: 2026-11-12 Sources: IT之家 report, Tencent Cloud Developer Community: C930 in-depth analysis, Xinhua Net: XuanTie’s first server-class CPU to ship next month
Nuclei NI1000’s Implementation Trade-offs: RVV as the Base with Matrix Added on Demand (11-09)
NI1000’s technical route is worth viewing side by side with the other two matrix extension paths: IME has dedicated matrix accumulation hardware and higher matrix compute energy efficiency; AME omits a full dedicated matrix register file and hardware matrix accelerator; VME sits in the middle, reusing RVV for both source operands and compute units, adding only Matrix Tiles, trading controllable hardware overhead for matrix compute. Nuclei also judges that VME will become the earliest-standardized RISC-V matrix extension—whether this holds depends on how fast toolchains and operator libraries follow, not on the IP itself.
On product cadence, IP release is only the beginning. Nuclei’s nuclei-tool-guide had 12 commits in a single day on 11-09, all for the 2026.10 version: upgrade notes, QEMU and OpenOCD release notes, libncrt fix entries split by version, and IDE known issues, plus pushing the VS Code extension to 0.3.0; on 11-11 it added GDB real-time variable watching. Such a detailed known-issues list indicates the version has entered customer delivery—for an IP vendor, this kind of engineering investment retains customers better than shipping another core.
Date: 2026-11-09 Sources: Nuclei press release, Design And Reuse report, nuclei-tool-guide
Automotive-Grade and Security Cores: AURIX RC1’s Supply Chain, AMD Zen 6’s RISC-V TEE (11-06 to 11-10)
This week’s information on AURIX RC1 fills in the supply chain layer: the core comes from MIPS processor IP, fabricated by GlobalFoundries, and MIPS, after being acquired by GlobalFoundries, shifted its entire roadmap to RISC-V. For teams building domestic automotive-grade RISC-V, the core’s origin 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 itself), and which foundry handles capacity and process node. On the toolchain side, HighTec and TASKING take over on the same day, 11-10, with certification timing scheduled for 2027.
Another inconspicuous but clearly directional clue comes from the x86 camp: a November 6 linux-firmware commit adds AMD PMF trusted application firmware v4 for the 1AH_M80H platform, with the description stating that this TA now runs on a RISC-V-based trusted execution environment, uses a larger shared memory layout, and supports extended policy conditions, while 1AH_M80H corresponds to the upcoming Zen 6 client processors. The confirmable fact is that this one PMF trusted application runs on a RISC-V TEE; one cannot infer that the entire security processor has switched cores; but for the RISC-V industry, this is an x86 major vendor putting RISC-V on the safety-critical path in a production platform.
Date: 2026-11-06 to 2026-11-10 Sources: Design And Reuse analysis, Phoronix report, linux-firmware commit
Espressif: ESP32-S31 Moves from Specs to Platform, 22 Repos Pushed in One Batch (11-06 to 11-12)
ESP32-S31 is the dual-core RISC-V chip released in April this year: 320MHz, Wi-Fi 6, Bluetooth 5.3, 802.15.4, on-chip PSRAM up to 32MB, gigabit Ethernet MAC. Two things landed simultaneously this week: Waveshare launched the ESP32-S31-Zero board, continuing the Zero series’ tiny size, with 18 through-holes and half-holes breaking out GPIO plus a 40-pin high-density connector; on November 7, Hackaday reported under “Proper Linux, On An ESP32” that the chip has an MMU and can run full-featured Linux rather than a nommu version, and Espressif has opened the espressif/linux repo on GitHub. Together, these two developments push the chip’s positioning past “MCU plus wireless” to the boundary between MCU and SoC.
Peripheral form factors are also filling in: M5Stack Core Matrix uses ESP32-C61 plus a 16×16 LED matrix at $29.99; two SparkFun ESP32-P4 IoT RedBoards respectively use the ESP32-C5 as a Wi-Fi 6 companion chip and PoE Ethernet; LocalDot uses ESP32-S3 plus ESP32-P4 to turn a third-generation Echo Dot into a Home Assistant voice terminal. On 11-12 Espressif pushed 22 repos in one batch, with the media-side esp-video-components, esp-vision, and esp-gmf repos all moving simultaneously, and the Bluetooth-side esp32s31-bt-lib and esp32h4-bt-lib appearing in the same batch—companion libraries for the new chip are following up, and whether modules can get usable firmware on day one depends on the cadence of these libraries.
Date: 2026-11-06 to 2026-11-12 Sources: CNX Software board report, Hackaday report, M5Stack Core Matrix report, espressif organization
Sophgo: se-rag-core Ships Four Versions in a Row, Qwen3.5 MTP, and a 19-Repo Weekly Release (11-07 to 11-11)
Sophgo made progress on three fronts this week. sophon-tools merged four PRs on 11-07, pushing se-rag-core from 1.6.0 to 1.8.1: 1.6.0 added offline pure BM25 indexing and multi-index merged retrieval, 1.7.0 added exact-match boosting for identifiers and titles, fixed top-n truncation, and added doc-type filtering, 1.8.0 added registry-mode entity-based routing and index building, and 1.8.1 fixed type annotations and parameter rejection logic. RAG is sinking from “vector database plus embedding model” to “offline index plus rule-based boosting”—the latter is the form edge devices can afford.
On 11-09, LLM-TPU merged MTP (multi-token prediction) support for Qwen3.5 dense models, and three bootloader-riscv commits switched SG2042 to build with the SG2044 toolchain, changed to building from the devel-sophgo branch, and unified the BMC firmware packaging script. In the final seven minutes of that window on 11-11 (around 03:00 Beijing time on 11-12), Sophgo pushed to 19 repos simultaneously, with the cviruntime commit message reading weekly rls 2026.11.12—a weekly-aligned release train: bootloader, firmware, kernel, runtime, operator libraries, multimedia, and build system all align version numbers on the same day, so the SDK customers receive won’t have runtime and operator library version mismatches, at the cost of individual repo commit histories being flattened by the release cadence.
Date: 2026-11-07 to 2026-11-11 Sources: sophgo/sophon-tools, sophgo/LLM-TPU, sophgo/cviruntime
Edge Vision and Physical AI Reference Designs: VeriSilicon VS815V, MIPS Three Tiers (11-06 to 11-12)
VeriSilicon debuted its edge AI vision chip VS815V at the Security Expo from November 3 to 6—its first RISC-V chip: a fully domestic dual-core RISC-V core plus a self-developed dedicated small core, an NPU providing 1.6T@INT8 (3.2T@A8W4) of compute, with emphasis on supporting on-device natural language interaction for small models on the order of 0.3B, up to 4 MIPI cameras connected in parallel without an external switch, hardware PDAF achieving 4K@30 without frame drops, and, per the vendor, full domesticization of process, foundry, packaging and testing, and key IP in manufacturing, using a fully domestic 22nm process. The company’s stated volume production timeline is 2027.
On 11-12, Electronic Design filled in the specs of MIPS’s three Physical AI reference designs: Actus-1 targets real-time motor control and communications, with two Atlas M8500 real-time cores plus an ARC VPX 512-bit vector unit, FDX FD-SOI process, on-chip MRAM, 100ps-class PWM, and gigabit Ethernet attached to a three-port switch with TSN support; Acies-1 is an M.2 module with Atlas S8200, an RVA23 platform, RVV and VME, 8 cores at 50 TOPS, bandwidth over 50GB/s, 10W power, including FP4, with software side connecting TensorFlow, PyTorch, and ONNX via a compiler based on IREE and MLIR; Aegis-1 uses multi-core ARC-V to run Linux and agent workloads, paired with a security processor and NPU. The essence of the three tiers is turning “how to allocate compute” into a selection problem rather than using one big SoC to brute-force all workloads.
Date: 2026-11-06 to 2026-11-12 Sources: RFID World report, Electronic Design report
V. Open Source Community and Industry Events
RISC-V Day Tokyo 2026 Autumn: Design Tutorial, Chip Museum, and Chiplet Statement (11-11)
RISC-V Association (Japan) will hold RISC-V Day Tokyo 2026 Autumn on November 11 at the University of Tokyo’s Hongo Campus, with an expected attendance of about 400. The morning agenda features a chip design tutorial, while the afternoon includes technical talks and booths: Tenstorrent will speak on physical AI and the Open Chiplet Atlas, AIST will speak on 2nm edge AI, Andes Technology will host a live demo, and the venue will also feature a special Chip Museum exhibit showcasing the hardware specifications and software overview of Google TPU v3.
Viewed alongside October’s RISC-V China Summit, regional events are developing a division of labor: the China event leans toward industry deployment and domestic substitution, while the Japan event leans toward design methodology and education. Worth noting are Tenstorrent’s chiplet statement and AIST’s advanced-process edge AI direction—both will shape how the “RISC-V servers and edge compute” narrative unfolds next year.
Date: 2026-11-11 Source: RISC-V Day Tokyo 2026 Autumn
Espressif’s 5th Global Developer Conference Held in Milan, MicroPixel Presents AI-Native Application System (11-06)
The 5th Espressif Global Developer Conference was recently held in Milan, Italy, under the theme “Building the Next Generation of Innovative Connectivity,” covering AI and edge intelligence, security, developer tools, and ecosystem collaboration, with a technical ecosystem day held concurrently in Shanghai. At the event, Shifang Ronghai Chairman Huang Guan presented “MicroPixel: An Embedded Development Harness That Brings AI into the Physical World,” discussing his second open-source project after XiaoZhi AI: the system layer encapsulates display, touch, audio, storage, and application management; the creation layer connects to coding agents via natural language; the experience layer provides installation, switching, and application-level sandbox isolation; and the closed-loop layer feeds real-device logs, screen feedback, and debugging channels back into the AI development workflow. According to public data, XiaoZhi AI has accumulated over 30,000 GitHub stars, connected to over 2 million hardware units, attracted 150,000 developers and makers, and handles 9 million daily conversations.
The connection to Espressif here is not about “what Espressif released” but about ecosystem spillover: projects like MicroPixel push the ESP32 from “writing firmware” toward “installing applications,” and the chips they rely on are precisely the RISC-V variants of the ESP32 series. For domestic teams building low-cost smart hardware, this is a reusable path—rather than waiting for chips to become more powerful, first make application installation, isolation, and updates into system-level capabilities.
Date: 2026-11-06 Source: ZAKER report, Jiemian News report
Jamestown Foundation Assesses China’s RISC-V: Strategy Ahead of Capability (11-10)
On November 10, U.S. think tank the Jamestown Foundation published an article assessing China’s RISC-V progress, authored by Sunny Cheung. The article presents several verifiable factual threads: SpacemiT and the XiangShan team unveiled the 46-core server chip V100 at the 6th RISC-V China Summit on October 19, claiming it to be the first general-purpose chip fully compliant with the RISC-V server standard; the “15th Five-Year Plan for the Electronic Information Manufacturing Industry” released in September proposed advancing RISC-V R&D and industrialization and developing RISC-V chips for artificial intelligence—previously, national-level policy documents rarely named the instruction set directly.
The article’s core judgment is “strategy over capability”: expanded use and influence over international organizations are clear, but performance still lags proprietary chips; the article cites concerns from Chinese analysts about a “glass ceiling” and states that RISC-V chips’ share of substitution in the national IT substitution plan is “negligible.” Such external assessments are worth reading as a cross-reference—they test the domestically common claim that “the ecosystem is already established” against a different set of metrics, and the gap between the two sets of metrics is precisely the question domestic RISC-V must answer in the next phase.
Date: 2026-11-10 Source: Jamestown Foundation
Robotics Software Stack: SpacemiT Fills In Perception, SLAM, Control, and Robot Model Adaptation in One Week (11-06 to 11-13)
The commit density in spacemit-robotics this week illustrates the domestic RISC-V robotics stack’s approach. On 11-06, the ros2_spacemit_vslam repo was created with its first commit, and on 11-07 it was registered into the full-machine manifest via the manifest repo; on 11-09, unitree-go2 gained smallcore motor communication support, and perception first separated model files from the source repo and then packaged the inference runtime into the release artifact; on 11-10, the humanoid side split G1’s SONIC tracking, motion presets, and startup-readiness determination, humanoid_common aligned scheduling accordingly, and control_whole_body handled feedback position clamping and command boundaries separately; on 11-11, SuperMimic expanded from LingLong to G1, adding four optional actions to standby and pushing them down to the common-layer action directory; on 11-12, vslam committed a stereo localization and mapping application, and ros2_humanoid fixed G1’s Nav2 constraints and native gait alignment.
The value of this line is not in any single algorithm but in “completing a reproducible full-machine reference implementation”: with all four layers—perception, localization and mapping, whole-body control, and robot model adaptation—in place, and the manifest in CI, upgrades won’t break others’ dependencies. As of the evening of 11-13, perception and vslam still had commits, part of the same round of continuation. The Beijing plan writes 2,000 hours of fault-free operation into its targets, and a considerable share of such faults come from repeated emergency stops and overheating caused by parameter mismatches, not hardware failure.
Date: 2026-11-06 to 2026-11-13 Source: vslam, perception, humanoid_unitree_g1, ros2_humanoid
Member Company Website Inspection and Source-Side Observations (11-09 to 11-13)
The five daily briefings conducted daily inspections of riscv.org member company websites, covering 391 companies: 66 to 69 failed to be fetched each day (including ISCAS, Tencent, Ingenic, MediaTek, DENSO, etc.), and 8 to 110 per day were judged to have content changes; upon item-by-item verification, all were homepage marketing rotations, announcement updates, university homepage rotations, or security verification pages, with no new products, funding, or architecture announcements, so no entries were included from the inspections this week.
The structure on the news side is also worth noting: Google News yielded about 20 valid new items this week, about one-third of which were re-circulations of old news (VeriSilicon VS815V, AMD RISC-V TEE, ESP32-S31 running Linux, Infineon AURIX, MIPS Actus, CH32’s clang, etc.), which were not counted again after being included on their respective dates; on the Chinese side, queries for “XiangShan,” “XuanTie,” “Nuclei,” “Yiwei,” and others had high noise—the term XiangShan alone had 44 hits on 11-13, all about red leaves, studies, and listed companies. What truly produced stable output was repository-side commits and vendor toolchain documentation.
Date: 2026-11-09 to 2026-11-13 Source: RISC-V International, openRuyi official news page
Leaderboard Watch: riscv100.top This Week
This week riscv100.top had no new score records and is judged to have no significant changes; however, the site completed a structural redesign on 11-10 that is worth recording separately.
After the redesign, the leaderboard page no longer outputs a full score table, but instead “one top card per architecture + an Overall Ranking overview list”: the cards give platform, compiler, Ratio/GHz, Run rank, Tests, and Run date, while the overview list retains only the best row for each platform. Taking cross-architecture SPEC CPU2006 as an example, the cards correspond to AWS EC2 Graviton5 (GCC 15.1.0, 24.91), XiangShan KMHv3 RTL (XSCC 2.0, 22.80, marked NEW, record date 2026-10-17), and AMD Ryzen 9 9950X (GCC 15.1.0, 17.72); the overview separately lists Yitian 710 (XSCC 1.1, 17.69) and SpacemiT K3 X100 (XSCC 1.1, 8.54). The homepage’s What’s New now has six items, including SpaceMiT K3 X100 XSCC 1.1 SPEC CPU2006, Version Progress Now Available, and Run Comparison Now Available.
Two caveats must be stated: first, the XiangShan KMHv3 record of 22.80 is dated 2026-10-17, earlier than this week—it is a site backfill onto the leaderboard, not a new score this week, and this publication recorded it on 11-11 under that interpretation; second, the site refreshed both the page structure and the snapshot baseline in the same redesign, making it impossible to distinguish “score changes” from “display contraction,” so no additions or removals are judged this week. Practical advice for readers: when specific scores are needed, read the cards and the overview; when historical trends are needed, use the site’s Version Progress and Run Comparison sections instead.
Date: 2026-11-10 Source: riscv100.top cross-architecture SPEC CPU2006, riscv100.top homepage
VI. Trend Observations
-
Interconnect IP has moved from supporting role to the main thread of RISC-V serverization: Among the specs disclosed this cycle for XL-300, the most telling item is not the 32MB L3 but the 8-core concurrency linearity of over 90%—a metric only someone who has actually run multi-core workloads would dare to give. Single-core performance depends on microarchitecture, multi-core performance on coherence and 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.
-
The competitive point for automotive-grade has shifted to toolchains and certification timing: The three steps of AURIX RC1 over three days—11-04 product definition, 11-09 IP source, 11-10 toolchain support—lay out clearly the sequence for advancing automotive-grade RISC-V: chip vendors define the platform, IP vendors supply the core, tool vendors provide compilation and debugging, and certification comes last. The “tools available in December, certification in 2027” given by the two tool vendors dovetails exactly with Infineon’s “sampling in the first half of 2027,” indicating the timeline was coordinated. The question domestic teams will most likely be asked next is not “do you have a RISC-V core” but “when will your toolchain pass certification.”
-
Distributions are shifting from “one board, one config” to “platform matrix plus reproducible CI”: A single debian_defconfig in RevyOS covers six platforms, and 101 commits in a single day consolidate board support into combinations; openRuyi over four consecutive days did version upgrades, security backports, kernel tagging, and firmware integration, and on Friday pushed kernel-team-tools tags to 7.3.4-3 and 6.18.63-4. Neither is shipping new features; both are about “whether others can build it and dare put it into production.” This lowers downstream costs more than shipping one more release.
-
Verification and observability have become the new front line of toolchain competition: NEMU adds a fast path for AME matrix memory access, difftest vectorizes MMA verification vectors and removes the tee pipeline that produced false successes, XiangShan wires topdown counters into the ROB, and Espressif turns instruction-level trace into four pip-installable commands—four actions pointing to the same judgment: once cores and extensions reach a certain complexity, the bottleneck shifts from “is the design correct” to “can we quickly see what it is doing.” Teams underinvesting in this line will pay multiplied time in the integration phase.
-
Demand-side policy is quantified, writing compute tiers into procurement parameters: The Beijing plan gives 200,000 units of annual complete-machine capacity, a 60% localization rate for key components, dexterous hands under RMB 10,000, and 2,000 hours mean time between failures—essentially writing “which chip, what cost, what reliability” into acceptance conditions. The joint side needs hard real-time and deterministic networking, the body side needs compute and bandwidth to run VLA, and the training side needs 50,000 PFlops clusters—none of these three tiers can be solved by a single big chip. MIPS’s three-tier reference designs and SingNova’s single-box solution represent exactly two answers, and the watershed is where software maintenance costs are paid.
-
Capital and external assessments are simultaneously anchoring domestic RISC-V: T-Head’s spin-off and Hong Kong listing puts a comparable target on the table—”already commercialized, with external customers, with revenue”—while the Jamestown Foundation uses another set of metrics—single-core performance, software maturity, and actual substitution rate—to conclude that “strategy precedes capability.” Both point to the same homework: when the narrative is dominated by shipment volumes, summit scale, and standards compliance, the external revenue breakdown and measured performance data in the prospectus are what will truly be tested in the next stage.
Appendix: Sources and Verification Notes
Sources: 5 RISC-V daily briefs this week — 2026-11-09 (Monday, 72-hour window, covering 11-06 05:00 to 11-09 05:00), 2026-11-10, 2026-11-11, 2026-11-12, 2026-11-13 (24-hour windows each). No missing dates within the period.
Supplementary verification:
| Check item | Result |
|---|---|
| GitHub organizations and key repositories (168-hour window) | Pushes within window: espressif 96 repos, sophgo 28 repos, spacemit-com 21 repos, openRuyi-Project 17 repos, sipeed 11 repos, openhwfoundation 8 repos, spacemit-robotics 7 repos, XUANTIE-RV 2 repos, revyos 2 repos, Nuclei-Software 1 repo; eswincomputing and radxa-linkr had zero pushes. Window extends to the morning of 11-14; 11-14 commits are excluded from the main text except for the noted early-morning addenda |
| openRuyi dual-source | No new releases on the official news page this week; the latest remain the two 2026-10 posts on Ruyi Firmware and ecosystem infrastructure. Commits to the main repo and firmware repo are included per the logs; activity on the evening of 11-13 and early morning of 11-14 is listed separately |
| Member company website sweep | 391 sites, 66 to 69 failures per day; all changes were marketing rotations, announcement updates, or security verification pages — no substantive product or architecture releases |
| riscv100.top | No new results this week; the site was redesigned on 11-10 into cards plus an overview, breaking the original table parser, so no new entries or removals were determined |
| Default branches of key repositories | XiangShan kunminghu-v3, revyos/linux’s revyos/7.3.y, spacemit-com/llama.cpp’s mtmd-backend, and openRuyi’s main all had commits within the window; the XuanTie organization’s 11-12 push was located to the zvfx-proposal repo, whose default branch has no readable commits for now |
Limitations: GitHub activity determinations in several of this week’s briefs rely on organization pages and repository Atom feeds (the anonymous API is rate-limited to 60 requests per hour), so pushes to non-default branches may be underestimated; the “zero pushes” conclusions for XuanTie and ESWIN are likewise subject to this method. XL-300’s L3 capacity appears as two differing figures — 23MB (2026-03 basis) and 32MB (the basis reported in this issue); no verifiable official spec sheet has been published, so both are recorded side by side in the main text. SingNova’s 256 TOPS is a vendor-claimed figure with no test conditions given. Score changes during the riscv100.top redesign cannot be observed continuously. The supplementary items from the evening of 11-13 and early morning of 11-14 are sourced from repository commit records; the 11-14 early-morning portion is noted as outside the period and has not been confirmed by vendor announcements.
Suggested next period: 2026-11-16 to 2026-11-20.