Evidence Library
Every key performance figure on this site comes from the signed / official test reports below (R1–R9). Each report’s scope, test platform and headline measurements are published free — click a report title to read them. The report PDF itself (full method, per-step raw data, charts, screenshots and signatures) is sold per report. The test code, load clients, orchestration scripts and raw data (R8) allow independent third-party reproduction of all conclusions. We label every number by provenance: measured / vendor / public / estimated.
Multi-model token-supply capacity and acceptance report (external official V2.3)
Full text is free. It is not one of the paid R1–R9 PDFs in the evidence library.
Specification charts are on the architecture page. Vendor-published specifications. Not our measured throughput, and not a commercial multiple versus NVIDIA.
FX100 Comprehensive LLM Inference & Training Benchmark (8× AMD MI308X)
LMCache parallel-read patch, multi-GPU KV tiering scale-out, concurrent model loading, model-switch effective TPS, training checkpoint concurrent writes (Qwen2.5-32B/7B).
Original file: test/AISSD5000-综合性能测试报告-AMD-MI308X.pdf
FX100 KV-Cache Benchmark (480B, TP8 long-context, signed)
Qwen3-Coder-480B-FP8, single 8-GPU instance (TP=8 standard production topology), long-context cold recovery, three-way comparison across concurrency levels.
Original file: test/AISSD5000-KVCache性能测试报告（480B·TP8长上下文·正式版）.pdf
FX100 KV-Cache Benchmark Summary (480B, TP4×2, all metrics, brand-unified edition)
Dual 4-GPU instances, nine load levels, full metrics (TTFT p50/p90/p99, TPOT, throughput, disk bandwidth); fs:// shared-pool cross-instance hot sharing verified; two independent runs within 5% deviation.
The hosted PDF is the brand-unified edition: the product is named FX100 throughout, the layout was rebuilt, and charts plus a numerical-verification appendix were added. All nine load levels match the signed original cell for cell — no measured value was added, changed, or recomputed. The signed original uses the legacy product name and is not hosted publicly to avoid confusion with the FX naming; it is available on request. Its MD5 (A66F5491F075F502BB0B2E2EFB0C101C) is printed inside this edition, so you can confirm you received the same file and then re-verify every number with fx_report/verify_r3.py.
Original file: test/AISSD5000-KVCache性能测试报告（480B·TP4x2·全指标汇总·正式版）.pdf
FX100 KV-Cache Benchmark (480B, multi-instance, official, No.-006)
Official compilation of R2 + R3 issued by a third-party testing organization.
Original file: test/AISSD5000-KVCache性能测试报告（480B·多实例形态·正式版）.pdf
FX100 KV-Cache Benchmark (14B, HBM efficiency, official, No.-004)
Seven-way comparison (GDS direct read / two re-compute HBM tiers / low-HBM direct read); three-step HBM-equivalence argument. Platform: MetaX N260 single GPU — methodology portable across platforms, numbers not mixed with MI308X.
Original file: test/AISSD5000-KVCache性能测试报告（14B·显存效益·正式版）.pdf
ComfyUI + LTX-Video 2.3 Full Deployment & Adaptation Report (V2)
All 7 delivered models (3 VAE / fp8 text encoder / text projection / 2 LoRA) running end-to-end on MI308X / ROCm 7.2 with no dtype, operator, or kernel errors.
Original file: test/ComfyUI-LTX2.3-完整报告（全模型·部署适配使用·V2）.pdf
ComfyUI + LTX-Video 2.3 Model Adaptation Report (AMD MI308X)
Per-model runs and adaptation notes for 3 VAE + Singularity LoRA.
Original file: test/ComfyUI-LTX2.3-模型适配报告（AMD-MI308X）(1).pdf
FX100 KV-Cache AMD Code Export + Raw Benchmark Data
LMCache parallel-read patch (git patch + full before/after), load clients, orchestration and forensic scripts, raw data — enables full third-party reproduction (contact us for access).
Original file: test/AISSD5000-KVCache-AMD代码导出包.zip; test/KVCache测试结果数据.rar
Xinyuanyi FX100-HBMM vs NFS Baseline on Huawei Ascend 910B
Huawei Atlas 910B ×8 (Kunpeng-920): model-serving load (DeepSeek-32B/70B), training weight/checkpoint I/O (Qwen-7B), training-data acceleration (YOLOv8/COCO) — three groups against an NFS baseline (Ascend platform, labeled as such; contact us for access).
Original file: test/微算WS-HBMM5000与华为910B模型推理与训练性能测试.docx
Each paid report includes two files: the signed Chinese original and an English reference translation (AI-produced, with machine-verified number fidelity — every numeric value in the source is asserted to appear in the translation; QC logs are published in the open repository under dmkt/report_i18n/qc/). The signed Chinese originals remain the authoritative versions — figures, screenshots, signatures and stamps are only in the originals. One piece of research is never charged twice by language. R8 and R9 have no hosted file; email us for access.
Naming note: FX100 appears in historical report filenames as AISSD5000 (also WS5000 / GP5000) — all refer to the same product; this site uses the unified FX naming while report entries keep original filenames for verification. 中文版：/evidence