Running Kimi K3 on MI355X at Better Performance per Dollar Than B300

MI355X 上运行 Kimi K3，性价比超越 B300

HN 131 分 · 41 条评论 · 作者 ilreb · 来源 www.wafer.ai · HN 讨论

【摘要】

Open source models have rapidly improved in recent months, with DeepSeek V4-Pro and GLM5.2 reaching near-Opus intelligence levels, establishing them as cost-efficient alternatives to closed models. However, Kimi K3 marks a new era, promising Fable/Sol-level intelligence but coming at a massive 2.8T parameters and requiring over 1.5TB of VRAM before KV caching, which exceeds even a B200 node's capacity.

⋯ 继续阅读请登录会员 ⋯