Mingxin FX100-HBMM vs NFS Baseline on Huawei Ascend 910B
Scope
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 filename: test/微算WS-HBMM5000与华为910B模型推理与训练性能测试.docx
Measured figures backed by this report (free)
These figures are not behind the paywall — they are the performance claims this site makes publicly, so anyone should be able to check them for free. What is paid for is the report itself: full method, per-level raw data, charts, screenshots and signatures.
Huawei Atlas 910B platform: DeepSeek-32B serving load 691 s → 112 s (6.2×), DeepSeek-70B 1399 s → 150 s (9.3×).
Provenance: Measured, R9 (Ascend platform)
Solutions backed by this report
Get the complete report
This entry is a code export package and raw data (or a report not hosted on this site); it is not listed as a standard product and is provided on request. The benchmark code and part of the raw data are already public and free onGitHubandHuggingFace.
Contact for accessEnglish PDFs are AI-produced reference translations with machine-verified number fidelity; the signed Chinese originals remain authoritative — figures, screenshots, signatures and stamps are only in the originals. Nothing on this page is investment advice.