Research window: 2026-10-14 04:30 ~ 2026-10-15 04:30 (Beijing time, past 24 hours).

Sources: Full-repository audit of three GitHub organizations (RuyiAI-Stack, ruyisdk, openRuyi-Project commits, pull requests, and releases via both Atom and API channels, with every item in the window individually verified; upstream fork commits filtered by RISC-V keywords, irrelevant commits excluded; openRuyi-Project firmware repository matrix opensbi / coreboot / edk2-ruyi / ruyifirmware and others checked repo by repo), ruyi package manager and PyPI version verification, openRuyi distribution news page, RuyiSDK developer community forum, RuyiAI official website, ruyiai-daily intelligence repository, arXiv search, Google News RSS multi-query in Chinese and English (via proxy; external hits in 4.2).

Index

  • Today’s Highlights: K3 firmware second release — UEFI setup introduces CPU mode selection (X100/A100, persistable), OpenSBI K3 v4 series and coreboot / EDK2 land in coordination on the same day
  • I. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V adaptation)
    • 1.1 pytorch: riscv branch completes another full-stack rebuild (23 commits, head b870a71a); 8 open PRs batch-refreshed (10-14)
    • 1.2 llvm-project: riscv branch overnight replay still not resumed; two FPGA PRs remain open (10-14 status)
    • 1.3 triton-riscv: nightly channel stays at issue 10-11; community optimization PR awaiting review (10-14 status)
    • 1.4 Intelligence and site infrastructure: ruyiai-daily pushes issue 10-14; site repository routine update (10-14)
  • II. RuyiSDK Toolchain and Distribution (build, toolchain, IDE, openRuyi)
    • 2.1 K3 firmware line: Ruyi Firmware 20261014 released — in-UEFI CPU mode selection and persistence land (10-14)
    • 2.2 OpenSBI and coreboot K3 branches: v4 series 13 commits land as a batch, CPU tier selection enters SPI configuration (10-14)
    • 2.3 openRuyi main repository: mesa bumped to 26.2.4 and PowerVR optimization backport initiated (10-14)
    • 2.4 ruyi-tuner: v1.x merged into main, compile comparison result persistence enhanced (10-14)
    • 2.5 Others: ruyi and PyPI remain at 0.53.0, Imager nightly refresh, website auto-update PR and ceph-ci maintenance (10-14)
  • III. Vector Extensions and Operator Optimization (VecIntrinBench / IntrinTrans / FBGEMM)
    • 3.1 All three repos silent; matrix extension documentation PR and support matrix line see no new pushes (10-14 status)
    • 3.2 External developments: RISE releases SpeexDSP RVV optimization, measured speedups on K1/K230 (10-14)
  • IV. Community and Events (forum, external news)
    • 4.1 No new topics in forum window; biweekly report 078 launch post remains latest (10-14 status)
    • 4.2 External news: direct queries return zero hits; openRuyi news page and arXiv see no new entries (10-14 status)
  • V. Trend Observations
    • 5.1 K3 firmware enters “configurable” phase: trade-offs of CPU mode selection and coordination watchpoints
    • 5.2 Dual-track routine maintenance: PyTorch morning rebuild cadence continues, LLVM riscv replay still absent
    • 5.3 openRuyi deep maintenance surface expands: graphics stack, storage CI, and cross-architecture tuning advance in parallel

Today’s Highlight: Second K3 Firmware Release — UEFI Setup Introduces CPU Mode Selection (X100/A100) with Persistence; OpenSBI K3 v4 Series and coreboot / EDK2 Land in Coordination on the Same Day

Date: 2026-10-14　Source: Ruyi Firmware 20261014 Release, OpenSBI K3 v4 series integration commit, edk2-ruyi commit ba4ff0ad, coreboot k3/port commit list, opensbi k3/port commit list

The openRuyi firmware packaging repository ruyifirmware released its second firmware version, “Ruyi Firmware 20261014,” on the afternoon of 10-14 (14:42). Its core feature is bringing choice to K3’s CPU configuration: the UEFI setup interface now offers CPU mode selection, with the setting persisted and taking effect on the next boot. Tools such as fastfetch can identify the mode as k3 / k3-x100-a100 / k3-a100. The three modes are: X100-only (default, 8 X100 cores, RVV VLEN 256, H extension retained); X100 + A100 mixed (16 cores, both RVV and H extensions disabled); A100-only (8 A100 cores, VLEN 1024, H extension disabled).

This release was completed through coordinated work across four K3 firmware stack repositories on the same afternoon: OpenSBI’s k3/port branch first landed the K3 v4 series (13 commits; 6 with the FROMLIST prefix at 09:50, 7 with the RUYI prefix at 13:05); coreboot’s k3-pico-itx selects the CPU tier via SPI configuration (14:11) and moves device tree fixup earlier to execute before payload handoff; on the EDK2 side, edk2-ruyi adds “persistent CPU selection mode” (14:11); ruyifirmware then integrates the OpenSBI K3 v4 series (13:05) and the K3 CPU selection mode (14:24) in sequence, completing packaging and release.

Only one day separates the first firmware release (20261013) from this one. K3’s delivery focus has shifted from link bring-up (the previous day’s batch debut of 35 EDK2 platform commits) to the hardware configuration surface — CPU cluster combinations and extension toggles are now left to firmware-layer selection. Points to watch: the performance and compatibility trade-offs of each mode (especially the mixed mode with RVV / H disabled and the big-core mode) under real workloads, how setting persistence behaves across firmware upgrade paths, and how it coordinates with the openRuyi distribution image (kiwi-image) and the toolchain under development (K1/K3 adaptation).

1. RuyiAI Stack Progress (PyTorch / Triton / TileLang RISC-V Adaptation)

Window Overview: The RuyiAI-Stack organization has 19 repos. Three actions within the window: pytorch (riscv branch underwent another full-stack rebuild with 23 commits; 8 open PRs batch-refreshed on the morning of 10-14; main synced 20 upstream commits), ruyiai-daily (10-14 issue), site repo (routine contributors data update, 1 commit); llvm-project (no pushes within the window, see 1.2 for riscv status), triton-riscv, harness, tilelang-riscv, mxnet, caffe, FBGEMM, IntrinTrans, VecIntrinBench, cpuinfo, torq-tile, dsh-osc, skills, pytorch-ci, SyncBots-DeepAgents, etc. had no pushes within the window.

1.1 pytorch: riscv branch completes another full-stack rebuild (23 commits, head b870a71a); 8 open PRs batch-refreshed (10-14)

Date: 2026-10-14　Source: riscv branch commit list, head commit b870a71a, previous head a17584a3, PR list

The riscv branch completed another full-stack rebuild: 23 local commits were re-landed wholesale with new committer timestamps (concentrated around 03:00 on 10-14), with the head updated to b870a71a “[SyncBots] Integrate PyTorch at e188cec” — an update over the previous head a17584a3 (same-titled rebuild, SHA refreshed per round); the stack composition maintains the existing sequence (#44 bytes_to_scalar fix, #50 blocklist expansion, #42 ULP tolerance, three SyncBots integration commits, #40 / #41 CI parsing, #20 / #21 CI split, #18 cpuinfo replacement, etc.), continuing the morning rebuild cadence (continuously visible since 10-08); no later push batch on this branch within the window. On the open PR side, 8 PRs (#23 Enable XNNPACK, #35 test PR, #36 timeout relaxation, #38 DataLoader temporary disable, #43 patch streamlining, #46 ATen RVV support, #48 RVV CPU capability, #51 riscv64 local image) had their baselines batch-refreshed on the morning of 10-14 (09:10), all remaining open with no merges or closures. The main branch synced 20 upstream commits within the window (including general changes to inductor, CUDA, CI, MPS, etc., not individually listed per filtering rules).

1.2 llvm-project: riscv branch overnight replay still not resumed; two FPGA PRs remain open (10-14 status)

Date: 2026-10-14 (status)　Source: riscv commit list, head commit 0fb2da86, PR #25, PR #28

The riscv branch remains without a new replay round: head 0fb2da86 “[RISCV] Add A100 HP/SP and pack/unpack lowering with calibrated scheduling (#31)” (committer timestamp around 10-12 23:00), with no new round seen over the two nights of 10-13 and 10-14 — the overnight replay cadence has now gone uncontinued for a second consecutive window (previously it ran three consecutive nights on 10-09, 10-11, and 10-12). The two FPGA backend PRs (#25 “BOSCAME FPGA backend”, #28 “xboscame FPGA”) remain open, with last updates stalled at 10-12. The main branch continues upstream sync as usual (not individually listed per filtering rules).

1.3 triton-riscv: nightly channel remains at the 10-11 build; community optimization PR awaiting review (10-14 status)

Date: 2026-10-14 (status)　Source: Releases page, PR #85

No new commits on the main branch within the window (head remains fe554ff, 10-10); the nightly channel produced no new build — the latest is still “Nightly 2026-10-11-fe554ff” (published 10-12). Community optimization PR #85 “softmax and matmul optimization” remains open (last updated 10-08, no activity within the window). Follow-up watch points: the next round of nightly artifacts and review progress on #85.

1.4 Intelligence and site infrastructure: ruyiai-daily publishes the 10-14 issue; site repo routine update (10-14)

Date: 2026-10-14　Source: 10-14 issue file, site repo commit list

The tech intelligence repo’s 10-14 issue went live on the morning cadence (pushed 10-14 06:20), covering a self-reported window of 10-13 05:00 to 10-14 06:00: top stories include Google releasing Gemini 4 Argon, NVIDIA open-sourcing the Run:ai platform, IBM’s Spyre compiler stack migrating to Inductor (code size reduced to roughly one-seventh), RISE SpeexDSP RVV optimization, the LLVM fpconvert proposal, and Liquid AI releasing LFM2.5-24B-A2B. The RISC-V section covers: RISE SpeexDSP RVV (validated on K1/K230, see also 3.2), gpSP’s Zephyr dynamic recompilation cache optimization (Kirby slow scenarios reduced from 30-100 ms per frame to 11-13 ms), a retrospective on the riscv.org mentorship program and ISA Explorer tool, RuyiSDK biweekly report 078 progress (0.54.0 expected end of month), new Sscofpmf tests in riscv-arch-test, sail-riscv virtualization GVA/D semantics fixes, and a XiangShan StoreQueue forwarding fix. The site repo performed a routine contributors data update (10-15 03:58, no code changes).

II. RuyiSDK Toolchain and Distributions (Build, Toolchain, IDE, openRuyi)

Window Overview: Within the ruyisdk organization window, three sets of activity: ruyi-tuner (v1.x merged into main, 2 commits persisting per-project compile comparison results), ruyisdk-website (auto data update PR #595 opened), ruyi-imager (nightly channel refresh); no new commits in the window for ruyi, packages-index, wechat-articles, support-matrix, board-docs, ruyici, ruyisdk-vscode-extension, riscv-gcc, riscv-binutils, riscv-gnu-toolchain, qemu, sail-riscv, riscv-isa-sim, ruyi-backend, AME-operator, attached-matrix-extension, linux-xuantie-kernel, and other repositories. The openRuyi-Project organization saw 23 commits across six repositories: opensbi (k3/port 13 commits), coreboot (k3/port 2 commits), edk2-ruyi (1 commit), ruyifirmware (2 commits + 20261014 release), openRuyi main repo (mesa 3 commits), ceph-ci (2 commits); kiwi-image, TakoPack, lookout, abaci-bot, homepage, k3firmware, k1firmware, etc. had no commits in the window.

2.1 K3 Firmware Line: Ruyi Firmware 20261014 Released — CPU Mode Selection and Persistence Land in UEFI (10-14)

Date: 2026-10-14　Source: Ruyi Firmware 20261014, ruyifirmware commit list, edk2-ruyi commit ba4ff0ad

ruyifirmware released its second firmware version (feature and mode details in today’s highlights): UEFI settings add CPU mode selection, settings persist and take effect on next boot, and the mode can be identified by tools such as fastfetch as k3 / k3-x100-a100 / k3-a100. Three modes: X100-only (default, 8×X100, RVV VLEN 256, H extension retained); X100+A100 mixed (16 cores, RVV and H disabled); A100-only (8×A100, VLEN 1024, H disabled). Two integration commits were completed before release: OpenSBI K3 v4 series integration (13:05) and K3 CPU selection mode integration (14:24); on the EDK2 side, edk2-ruyi added “K3: Add persistent CPU selection modes” the same day (14:11). The previous day (10-13) the repository had just established the packaging chain with 5 commits and released its first version — two versions in two days, closed out same-day, with K3 firmware delivery cadence steady at “same-day integration, same-day release.”

2.2 OpenSBI and coreboot K3 Branches: v4 Series Lands 13 Commits in a Batch, CPU Tier Selection Enters SPI Configuration (10-14)

Date: 2026-10-14　Source: opensbi k3/port commit list, coreboot k3/port commit list, OpenSBI K3 v4 integration commit 14687d0a

The OpenSBI k3/port branch added 13 commits in the window: 6 with the FROMLIST prefix (a batch at 10-14 09:50) — K1 platform support refactoring (hart init moved down, cache flush operations renamed), K3 platform support added with cold_boot_allowed override, SBI selecting the expected trap handler per hart; 7 with the RUYI prefix (a batch at 10-14 13:05) — K3 auxiliary cluster warm-up retention, native warmboot handoff, CCI switch wait upper bound, restoring the P1 bus before I2C reset, K3 ACLINT binding support, Kconfig output and driver array refresh, ACLINT timer quirk naming. The coreboot k3/port branch had 2 commits the same day: k3-pico-itx selects CPU tier via SPI configuration (14:11, echoing the mode selection on the UEFI / EDK2 side), device tree fixup moved earlier to execute before payload handoff (10:54). The four repositories (opensbi / coreboot / edk2-ruyi / ruyifirmware) formed a coherent “base → platform → packaging → release” delivery on the same afternoon.

2.3 openRuyi Main Repo: mesa Bumped to 26.2.4 and PowerVR Optimization Backport Started (10-14)

Date: 2026-10-14　Source: main repo commit list, mesa 26.2.4 commit 5bc42d25, PVR backport commit c1b1bb15

The main repo had 3 commits in the window (author Icenowy Zheng), all around the mesa graphics stack: mesa upgraded to 26.2.4; mesa-26.2.4-pvr-backport.patch added to start the first round of PowerVR (PVR) optimization backports; mesa-gallium’s dependency recommendation on mesa-vulkan-drivers adjusted. After the previous day’s batch updates to Go modules and llvm-snapshot / GnuPG, distribution maintenance attention shifted to the graphics stack. The main repo has 71 open PRs, among which #1421 (enabling qemu graphics test functionality), #1431 (rdma-core dependency adjustment), #1425 (CODEOWNERS), etc. had updates in the window.

2.4 ruyi-tuner: v1.x Merged into main, Compile Comparison Result Persistence Enhanced (10-14)

Date: 2026-10-14　Source: ruyi-tuner commit list, commit 14706ae1, merge commit ba0e26f5

Compiler auto-tuning tool ruyi-tuner had 2 commits in the window: writing “per-project actual compile comparison results” to txt and updating step 3’s output message (author shining1984, 16:21; touching CHANGELOG, Readme, and real_compile.py / run_csibe.py); then the v1.x branch was merged into main (16:28). The tool has recently been continuously expanding along the CSiBE code-size benchmark chain (riscv / aarch64 / x86 three-architecture comparison); this commit completes persistent output of actual compile results; after the v1.x merge, the main branch carries the tool’s latest form.

2.5 Others: ruyi and PyPI Hold at 0.53.0, Imager Nightly Refresh, Website Auto-Update PR and ceph-ci Maintenance (10-14)

Date: 2026-10-14　Source: ruyi Releases page, Ruyi Imager nightly build, ruyisdk-website PR #595, ceph-ci commit list

The ruyi package manager main repo and PyPI both hold at 0.53.0 (uploaded 09-24), with no new version in the window (0.54.0 expected by month-end per biweekly report 078). ruyi-imager nightly channel refreshed routinely (10-14 06:46, no code changes). ruyisdk-website opened auto data update PR #595 “Auto-update: API data to latest” (created 10-14 07:30, routine). ceph-ci had 2 commits: fork-patches adding intarith.h header for 1039 test_DiffIterate.cc (10:27), and a cleanup commit (18:23). kiwi-image, packages-index (open PR #228 with no in-window updates), support-matrix (all 4 open reports stalled at 10-12), board-docs, ruyici, VS Code extension (0.1.7 remains latest), etc. had no new pushes in the window.

III. Vector Extensions and Operator Optimization (VecIntrinBench / IntrinTrans / FBGEMM)

3.1 All three repos silent; no new pushes on matrix extension doc PR or support matrix thread (10-14 status)

Date: 2026-10-14 (status)　Source: VecIntrinBench, IntrinTrans, FBGEMM fork, attached-matrix-extension PR #1

The three target repos — covering vector intrinsic benchmarks, intrinsic translation, and low-precision operator libraries — remain silent (last commits: VecIntrinBench 02-13, IntrinTrans 06-25, FBGEMM fork 07-13); the companion cpuinfo and torq-tile repos also saw no activity. On the matrix extension documentation side: the “ztt intrinsic doc V0.6” PR #1 in attached-matrix-extension remains open, with its last update stalled at 10-12. The four pending support-matrix reports (#409 ESP32-C6 RT-Thread, #405 Zephyr Milk-V Duo S, etc.) saw no updates in this window.

3.2 External developments: RISE releases SpeexDSP RVV optimizations, measured speedups on K1/K230 (10-14)

Date: 2026-10-14　Source: RISE blog post, Hacker News discussion

The RISE project published progress on RVV optimizations for SpeexDSP (author: Tristan Matthews): introducing Meson builds, Checkasm tests, and hand-written RVV kernels for SpeexDSP, validated on K1 (VLEN 256) and K230 (VLEN 128) — K1 floating-point resampling achieved roughly 3.71x / 3.59x speedups over Clang 18 / GCC 13 baselines, respectively. This work extends RVV optimization into foundational multimedia libraries such as VoIP and audio resampling, directly relevant to the SpacemiT platform tracked by RuyiSDK / openRuyi (entry captured via the intelligence repo’s 10-14 issue; classified as external ecosystem development).

IV. Community and Events (Forum, External News)

4.1 No new forum topics in window; biweekly report 078 launch post remains latest (10-14 status)

Date: 2026-10-14 (status)　Source: RuyiSDK Developer Community, Topic #2891

No new topics or replies in the community forum window: latest activity remains the 10-13 “RuyiSDK Biweekly Progress Report Issue 078” launch post (#2891, 17:51) and the summit participation announcement (#2889); the daily news post series shows no new issue after the 10-13 edition (#2890). With several days remaining before the summit opens (10-18), the community side is in a quiet preparation phase; watch for: pre-summit release cadence and booth demo teasers.

4.2 External news: direct queries return zero hits; no new entries on openRuyi news page or arXiv (10-14 status)

Date: 2026-10-14 (status)　Source: openRuyi distro news page, arXiv

On the external news side: Google News queries (when:7d) across multiple Chinese and English combinations for RuyiSDK / RuyiAI / openRuyi / “如意社区” / “Ruyi Software” continue to return zero relevant hits (channel functioning normally); the openRuyi distro news page has no new articles (the 2026.09 release announcement remains latest; the 2026.10 release page is not yet live); arXiv (RISC-V compilation, vector, and edge AI directions) shows no new entries in the window (most recent relevant items remain the 10-10 on-demand data layout transformations, etc.). Summit-related preview coverage continues to be tracked by the RISC-V Daily; this newsletter does not duplicate it (RuyiSDK participation details in 4.1 and Biweekly Report 078).

V. Trend Observations

5.1 K3 Firmware Enters a “Configurable” Phase: Trade-offs in CPU Mode Selection and Points for Coordinated Observation

With only one day between the first firmware release (10-13) and the second (10-14), the focus of the K3 firmware line has shifted from link bring-up to the hardware configuration surface: CPU cluster combinations (X100 / A100) and extension switches (RVV, H) are now selected and persisted by the firmware. Points to watch: first, the performance and compatibility trade-offs of the three modes under real workloads — the applicable scenarios for mixed mode and for disabling RVV / H in big-core mode need to be defined through testing; second, how settings persistence behaves along the firmware upgrade path and in recovery scenarios; third, the coordination cadence with the openRuyi distribution image (kiwi-image) and the toolchain under development (K1/K3 adaptation) — with the summit week (starting 10.18) approaching, whether K3-related results will be unveiled in a concentrated fashion.

5.2 Routine Maintenance on Two Lines: PyTorch Morning Rebuild Cadence Continues, LLVM riscv Replay Still Has Not Returned

The two fork maintenance lines continued to diverge in this window: the PyTorch riscv branch completed its morning rebuild as usual (23 commits, head b870a71a), with 8 open PRs batch-refreshing their baselines on the morning of 10-14; the LLVM riscv branch saw no new round of nightly replay for the second consecutive window (head remains at the 10-12 version) — whether replay returns, and the relationship between the ruyisdk/llvm-project rebase line (XTHeadVector fix already merged) and the fork replay line, remain under observation.

5.3 openRuyi’s Deep Maintenance Surface Expands: Graphics Stack, Storage CI, and Cross-Architecture Tuning Advance Together

Distribution maintenance goes beyond bulk dependency updates: in this window, mesa (26.2.4 upgrade + PVR optimization backports), ceph-ci (continuous fork-patches), and ruyi-tuner (compilation comparison results persisted to disk, v1.x merged into mainline) each made incremental progress across three directions — graphics stack, storage stack, and toolchain — and together with the daily release cadence of the firmware line, openRuyi’s “deep maintenance surface” continues to expand ahead of the summit. Points to watch: how these dedicated workstreams couple with the 0.54.0 release and the 2026.10 image release cadence.

Source Verification Table

Source Verification Result
RuyiAI-Stack organization (all repos) pytorch: a new full-stack rebuild (23 commits, head b870a71a) + 8 PRs batch-refreshed + 20 main upstream sync commits; llvm-project: no new riscv replay (head 0fb2da86) + two FPGA PRs open; triton-riscv: no new commits (nightly channel issue 10-11); ruyiai-daily issue 10-14; site repo 1 routine commit; no pushes to harness, tilelang-riscv, mxnet, caffe, FBGEMM, etc.
ruyisdk organization (all repos) ruyi-tuner: v1.x merged into main + 2 commits persisting comparison results; ruyisdk-website: automated update PR #595 opened; ruyi-imager: nightly channel refreshed; no in-window commits to ruyi, packages-index, wechat-articles, support-matrix, board-docs, ruyici, VS Code extension, etc.
openRuyi-Project organization (all repos) 23 commits across six repos: opensbi (k3/port 13) + coreboot (k3/port 2) + edk2-ruyi (1) + ruyifirmware (2 + 20261014 release) + openRuyi main repo (mesa 3) + ceph-ci (2); no in-window commits to kiwi-image, TakoPack, lookout, abaci-bot, homepage, k3firmware, etc.
RuyiSDK community forum No new topics or replies (latest is the 10-13 biweekly report 078 post; no 10-14 issue of the daily news)
openRuyi news page No new articles (the 2026.09 announcement remains the latest; the 2026.10 release page is not yet online)
ruyi package manager / PyPI 0.53.0 remains the latest (09-24); 0.54.0 expected at month-end (per biweekly report 078)
Tech intelligence repo (ruyiai-daily) Issue 10-14 pushed on the morning cadence (06:20); the RISC-V section includes SpeexDSP RVV, gpSP, Sscofpmf, etc.
arXiv (RISC-V compilation and vector) No new entries in the window (most recent 10-10)
Google News RSS (multiple Chinese and English queries) Zero hits for RuyiSDK / RuyiAI / openRuyi / 如意社区 / Ruyi Software; channel functioning normally
RuyiAI official site and blog No new blog posts (1 routine data update in the site repo)

Complete Source List