铭信

Evidence Library

Every key performance figure on this site comes from the signed/official test reports below (R1–R9). Reports are directly downloadable; test code, load clients, orchestration scripts and raw data (R8) allow third parties to independently reproduce all conclusions. Figures are labeled by type: measured / vendor / public / estimated.

Primary test platform (R1–R4): 8 × AMD Instinct MI308X(每卡 192 GB HBM,gfx942); 2 × AMD EPYC 9654(384 线程),≈1.5 TB 内存; ROCm 7.2; vLLM 0.20.1+rocm721; LMCache(上游主线 2026-06-29 源码编译). Device under test: 铭信 FX100 全闪 NVMe-oF 阵列:4 盘 RAID0(14 TB, XFS),RoCEv2,单口 100 GbE. Model: Qwen3-Coder-480B-FP8(MoE,权重 ≈450 GB).
R12026-07-03

FX100 大模型推理与训练 综合性能测试报告(AMD MI308X ×8)

Report text is in Chinese; tables and metrics are readable without translation.

R22026-07-05

FX100 KV Cache 性能测试报告(480B·TP8 长上下文·正式版)

Report text is in Chinese; tables and metrics are readable without translation.

R32026-07-06

FX100 KV Cache 性能测试汇总报告(480B·TP4×2·全指标·正式版)

Report text is in Chinese; tables and metrics are readable without translation.

R42026-07-06

FX100 KV Cache 性能测试报告(480B·多实例形态·正式版,编号-006)

Report text is in Chinese; tables and metrics are readable without translation.

R52026-07-03

FX100 KV Cache 性能测试报告(14B·显存效益·正式版,编号-004)

Report text is in Chinese; tables and metrics are readable without translation.

R62026-07-07

ComfyUI + LTX-Video 2.3 全模型部署·适配·使用完整报告(V2)

Report text is in Chinese; tables and metrics are readable without translation.

R72026-07-07

ComfyUI + LTX-Video 2.3 模型适配报告(AMD MI308X)

Report text is in Chinese; tables and metrics are readable without translation.

R82026-07

FX100-KVCache AMD 代码导出包 + KVCache 测试结果数据

Report text is in Chinese; tables and metrics are readable without translation.

R92026-05-30

铭信 FX100-HBMM 与华为 910B 模型推理与训练性能测试(vs NFS 基线)

Report text is in Chinese; tables and metrics are readable without translation.