# Huawei Ascend 950DT Test Document Claims 86% of Nvidia H200 Performance

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/60136
Published: 2026-11-18T05:00:46.000Z
Updated: 2026-11-18T05:00:46.000Z
Section: Technology

> The document lists 89% of estimated H200 throughput in one GLM 5.2 scenario and 82% in another.

A circulating October-dated test report gives Huawei’s Ascend 950DT 86% of Nvidia’s H200 inference performance overall. The document is dated October and describes an Oct. 10 test of GLM 5.2 and DeepSeek V4 Pro on a 64-card Ascend 950DT supernode.

For GLM 5.2 with 16,000 input tokens and 2,000 output tokens, the document lists single-card total throughput of 1,951 tokens per second for the 950DT, 483 for the Ascend 910C and 169 for the Ascend 910B4. It estimates H200 throughput at about 2,200 tokens per second in that scenario, putting the 950DT at about 89% of the estimate. A second comparison, using 7,000 input tokens and 2,000 output tokens, puts it at about 82%.

The document’s overall conclusion gives the 950DT an 86% comparison with the H200, separate from those two scenario figures. Its cost analysis puts the 950DT’s inference performance per unit of cost at 3.1 times the 910C’s and 2.9 times the 910B4’s, and estimates a roughly 30% cost-performance edge over the H200.


