Skip to main content

Changelog

2026-06-10​

📰 Added 3 domestic chip feature blog posts​

  • Ascend 910C detailed specs and deployment status:
    • Created blog/2026-06-10-ascend-910c-specs-and-deployment.mdx
    • Includes detailed spec parameters, deployment cases, performance comparisons, and deep technical architecture analysis
    • Covers inference/training performance, energy efficiency, software ecosystem, and use cases
  • Kunlun P800 performance data:
    • Created blog/2026-06-10-kunlun-p800-performance.mdx
    • Includes architectural innovations (Tianchi architecture), performance data, benchmark tests, and use cases
    • Covers inference/training performance, energy efficiency, and the Tianchi supernode solution
  • New domestic chip progress in 2026:
    • Created blog/2026-06-10-2026-domestic-chip-progress.mdx
    • Summarizes progress on new products such as Huawei Ascend 950/960/970 and Kunlun M100/M300
    • Analyzes technical breakthroughs, market dynamics, ecosystem building, and future outlook

📊 Comprehensive chip data proofreading​

  • Scope: all 121 chips in static/chips.json
  • Method: initial proofreading based on the AI knowledge base (as of February 2026)
  • Results:
    • ✅ Chips with correct data: 108 (89.3%)
    • ⚠️ Chips requiring verification: 13 (10.7%)
    • 🔧 Errors found and corrected: 3 (already corrected previously)
  • Previously corrected errors (on record):
    1. AMD MI355X: TDP from 750W → 1400W ✅
    2. Huawei Ascend 910C: BF16 TFLOPS from 780 → 800 ✅
    3. NVIDIA B300 Ultra:
      • TDP from 1400W → 1200W ✅
      • Memory bandwidth from 8 TB/s → 12 TB/s ✅
      • FP16 from 3.5 PFLOPS → 4 PFLOPS ✅
  • Created proofreading report: CHIP_DATA_PROOFREADING_REPORT.md, with detailed results and recommendations

🔗 Fixed Broken Links​

  • Fixed all broken links (583 → 0)
  • Fixed the tags.yml permalink format issue (Docusaurus auto-adds /blog/tags/ prefix)
  • Removed the broken nvlink-6 link from the Vera CPU doc

2026-06-07​

🏠 Homepage ranking algorithm optimization (combined options A+B+C)​

  • Option A: weight allocation 60%/40% → 50%/50% (compute weight raised, ranking focuses more on actual performance)
  • Option B: Huawei vendor weight 0.85 → 0.9 (on par with AMD, reflecting the rising status of domestic chips)
  • Option C: inference category weight 0.8 → 0.9 (reflects the 2026 market trend of inference demand surpassing training)
  • Files modified: src/pages/index.tsx, src/data/home-content.zh.ts, src/data/home-content.en.ts

🔧 Fixed Cloudflare Pages build failure​

  • Problem: src/pages/home-content.zh.ts and home-content.en.ts were treated by Docusaurus as page components, causing the build to fail (missing default export)
  • Fix: moved the data files from src/pages/ to src/data/ and updated the import paths in index.tsx
  • Commit: 6632b88

📝 Fixed blog post build errors​

  • Issue 1: the authors: [mirrorfrog] field on all 4 new blog posts was invalid (the valid author is aicomputecards)
  • Issue 2: the title of blog/2026-06-07-ai-hardware-inflection-point-2026.mdx contained unescaped double quotes, breaking YAML parsing
  • Fix: changed all authors: [mirrorfrog] to authors: [aicomputecards]; wrapped titles in single quotes
  • Commit: 7bf64fa

📰 Added industry news blog posts​

  • Added 2 Chinese blog posts:
    • 2026-06-01-global-ai-computing-report-2026.mdx — Major release: 2026 Global AI Computing Report and the top 10 trends in the computing industry
    • 2026-06-07-ai-hardware-inflection-point-2026.mdx — AI hardware enters the "deployment era": 5 major shifts and survival rules for 2026
  • Added 2 English blog posts (mirroring the Chinese versions)
  • Updated static/news.json so the homepage "Latest News" shows these 2 new articles
  • Commit: c93aefd

💻 Added NVIDIA Rubin chip docs​

  • Created the Chinese doc: docs/cards/nvidia/rubin.mdx
  • Created the English doc: i18n/en/docusaurus-plugin-content-docs/cards/nvidia/rubin.mdx
  • Specs: 336B transistors, 288GB HBM4, 50 PFLOPS FP16, mass production planned for Q3/Q4 2026
  • Commit: aacc6c7

🔄 Created the interactive chip comparison tool​

  • Created src/pages/compare.tsx — interactive chip comparison page
  • Created src/pages/compare.module.css — comparison tool styles
  • Features:
    • Supports selecting 2-4 AI chips for side-by-side comparison
    • Fetches chip data from /chips.json
    • Shows specs: vendor, architecture, process, release date, compute (FP8/FP16/FP32/INT8), memory (type/capacity/bandwidth), power (TDP)
    • Bilingual support (Chinese/English) via useDocusaurusContext()
  • Access at: /compare (Chinese), /en/compare (English)
  • Commit: 6c012d6

🔗 Fixed Broken Links (H20/H800 docs)​

  • Created the NVIDIA H20 chip doc:
    • Chinese: docs/cards/nvidia/h20.mdx
    • English: i18n/en/docusaurus-plugin-content-docs/current/cards/nvidia/h20.md
    • Specs: Hopper-architecture export-compliant version, FP16 148 TFLOPS, 96GB HBM3, 4.0 TB/s, released 2024
  • Created the NVIDIA H800 chip doc:
    • Chinese: docs/cards/nvidia/h800.mdx
    • English: i18n/en/docusaurus-plugin-content-docs/current/cards/nvidia/h800.md
    • Specs: China-specific H100 variant (2022), FP8 1,979 TFLOPS, 80GB HBM3, 3.35 TB/s, released 2022
  • Corrected links (H200 → H20, the right comparison):
    • docs/cards/others/hygon-dcu-k100.md — related card
    • docs/cards/others/kunlun-p800.md — related card
    • Same corrections applied to the English versions
  • Updated docs/gpu/nvidia-cuda.mdx (Chinese and English) to add the H800 page link
  • Commits: c13db84 (H20), 8167da0 (H800)

🏷️ Defined blog tags (tags.yml)​

  • Problem: several tags used by blogs were missing from blog/tags.yml, causing "not defined" build warnings
  • Fix:
    • Added Chinese-language tag keys (新闻, AI芯片, 国产芯片, 昇腾, 推理算力, 产业报告, 算力趋势, 智算驱动, AI硬件, 边缘AI, 专用芯片, 能效散热)
    • Each key uses a unique permalink (with a -zh suffix to avoid conflicts)
    • Added English tag keys (selection-guide, deep-learning, announcement, site-update, new-feature, architecture-comparison, inference, domestic-chips, domestic-substitution, ai-chips, ai-chip, selection, etc.)
    • Fixed the casing of an English blog tag key (AI-chip → ai-chip; YAML keys are case-sensitive)
    • Removed the duplicate news tag from the English blog
  • Commit: 7a6b5b2

💻 Added more chip data (Vera CPU & SN50)​

  • Created the NVIDIA Vera CPU doc (announced at GTC 2026-03-18):
    • Chinese: docs/cards/nvidia/vera-cpu.md
    • English: i18n/en/docusaurus-plugin-content-docs/current/cards/nvidia/vera-cpu.md
    • Specs: 88-core Armv9.2 custom architecture, 176 threads, 1.5TB LPDDR5X, 1.2 TB/s, NVLink-C2C 1.8 TB/s, the world's first CPU with native FP8 support
  • Updated the SambaNova SN50 doc (replacing the old version):
    • Chinese: docs/cards/others/sn50.md
    • English: i18n/en/docusaurus-plugin-content-docs/current/cards/others/sn50.md
    • Specs: fifth-generation RDU, 10T+ parameter models, 10M+ context, 5× inference speed (Llama 3.3 70B: 895 vs B200 184 tokens/s), 3× TCO advantage, strategic partnership with Intel, released 2026-02-24
  • Updated static/chips.json:
    • Added the vera-cpu entry (121 chips)
    • Updated the sn50 entry with accurate data (from speculation to official figures)
  • Commit: 849df78

2026-06-06​

🎉 Full-site English translation completed​

  • All 220 pages (197 docs + 23 blogs) translated into English
  • Covers all chip cards, architecture deep-dives, category indexes, and type guides
  • UI localized with bilingual support

🔗 Fixed 636 broken links​

  • Added the /docs/ prefix to legacy routes such as /gpu/, /asic/, /tpu/
  • Fixed in both the Chinese and English versions

📝 README improvements​

  • Added badges, contribution guide, directory structure, and latest updates
  • Added translation quality and formatting notes

📋 Changelog page launched​

  • This very page

🛠️ Automation​

  • Added a blog translation automation script
  • Added a scheduled chip data validation job

2026-06-05​

🔍 On-site search launched​

  • Installed @easyops-cn/docusaurus-search-local
  • Supports Chinese tokenization and offline search
  • Invoked with the / key on the right side of the navbar

🆕 New chip cards​

  • Intel Crescent Island: Xe3P architecture, 480GB LPDDR5x, 350W inference GPU
  • Huawei Ascend 950 series: 950PR/950DT, 1 PFLOPS FP8, self-developed HBM

📊 Data updates​

  • Added Crescent Island and 950PR/950DT to the comparison table
  • Corrected Trainium 3 FP8 compute, GB200 FP4 sparse, and B300 Ultra comparison table units
  • Fixed GitHub push authentication issues

✍️ New blog​

  • 2026-06-05-computex-2026-roundup: Computex 2026 roundup
  • English version of this post translated as well

2026-06-04​

🔬 Online data verification​

  • Huawei Ascend 910B: HBM2e → HBM3 correction
  • Huawei Ascend 910C: 780 TFLOPS BF16 confirmed, TDP added
  • Moore Threads MTT S5000: compute corrected from 50 TFLOPS → 1000 TFLOPS
  • NVIDIA H100: FP32 corrected from 60 → 67 TFLOPS
  • AMD MI355X: TDP 750W → 1400W, FP8 4.6 → 5 PFLOPS
  • AWS Trainium 3: FP8 362 PFLOPS → 5.7 PFLOPS

📝 Roadmap update​

  • Reorganized the timeline (2025 Q2 → 2026 H2 → 2027)
  • Added Vera Rubin mass production, RTX Spark launch, and Ascend 950 delivery

✍️ New blogs​

  • 2026-06-04-huawei-ascend-950-mass-production: Ascend 950 and the Chinese chip ecosystem
  • 2026-06-01-nvidia-rtx-spark: NVIDIA RTX Spark launch

2026-06-03 and earlier​

🏗️ Site foundation​

  • Built on Docusaurus 3.10
  • Initial data for 111 AI compute cards
  • Blog system launched (24 initial posts)
  • Cloudflare Pages deployment
  • Custom domain mirrorfrog.com
  • Dark mode support
  • Chinese/English i18n architecture
  • JSON-LD structured data / Sitemap SEO optimization
  • Complete .gitignore configuration