PulseAugur
EN
LIVE 20:04:21

Qwen3-8B model shows 15% token agreement on Ascend vs CUDA hardware

By PulseAugur Editorial · [1 sources] ·

A technical comparison revealed significant differences in output between the Qwen3-8B model when run on Huawei Ascend hardware using DashScope versus NVIDIA GPUs with CUDA. Across 61 prompts, there was only a 15% agreement on the top predicted token, indicating substantial divergence in model behavior. This contrasts with a much smaller divergence rate observed between two NVIDIA GPUs using the same CUDA stack. AI

IMPACT Highlights potential inconsistencies in LLM outputs across different hardware and software configurations, impacting reproducibility and deployment.

RANK_REASON Technical comparison of model output across different hardware/software stacks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

Qwen3-8B model shows 15% token agreement on Ascend vs CUDA hardware

COVERAGE [1]

  1. dev.to — LLM tag TIER_1 Nederlands(NL) · Ruitong ·

    Ascend vs CUDA: Same Qwen3-8B Model, Same Seed, 15% Token Agreement

    <p>We ran Qwen3-8B on Ascend (DashScope) and NVIDIA A40 (CUDA) with identical parameters across 61 prompts. The result: only 15% top-1 token agreement. 98% of prompts diverged.</p> <p>Results:</p> <ul> <li>Top-1 token agreement: 15.02%</li> <li>Prompts diverging: 60/61 (98.4%)</l…

RELATED ENTITIES

RELATED TOPICS