以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Ascend 910C

9 atoms · 跨 4 天 · 首见 2026-06-04 · 最近 2026-07-03

三色: 🟦 fact 6 · 🟥 take 3 · stance ▲2/▼1/◆1

来源: X 9

时态: fresh:9

标签: 好数字:6 · 好信源:4 · 好观点:1

2026-07-03 · 性能对比
910C is roughly 90% of p800
tags: 好数字 · → daily
value: qty=90% of p800 · direction=na
2026-07-03 · 训练稳定性
that's not really how you treat your flagship, especially if you've already developed confidence in training stability
tags: 好观点 · → daily
value: qty=na

2026-07-03 · 训练效率对比
910C is a cluster ≈10X larger than DeepSeek's at the time of training V3 (55 days), for maybe a 3x pretrain. But I can't see it taking >5-6 weeks.
tags: 好数字 · → daily
value: qty=10X larger cluster than DeepSeek V3, 3x pretrain, <5-6 weeks
2026-07-03 · 训练时间
even with FP16 it's < 2 months
tags: 好数字 · → daily
value: qty=< 2 months using FP16

2026-06-30 · 训练规模
Chyna can also train an 1.6T 48AB (35T tokens) MoE on 25K Ascend 910C. This was completed by late April.
tags: 好数字·好信源 · → daily
value: qty=25K chips · date=2026-04
2026-06-04 · 训练1.6万亿参数模型
Shenzhen team trains 1.6 trillion-parameter AI model on domestic Ascend 910C cluster
tags: 好数字·好信源 · → daily
value: qty=1.6万亿参数 · date=2026-06-05
2026-07-03 · NPU核心数据格式支持
Ascend 910C的NPU核心部分不支持FP8、BF16，这两个数据格式的浮点数是通过FP16模拟计算得到的结果。
tags: 好信源 · → daily
value: qty=na
2026-06-07 · 芯片封装配置
Ascend 910Cが 2 die + 8 HBM の構成ですね。
tags: 好信源 · → daily
2026-06-07 · HBM 容量
HBM2e / 64GB x 2 = 128GB ?
tags: 好数字 · → daily
value: qty=128GB