Chinese AI firm DeepSeek posted a cryptic message, hinting its V3.1 model’s UE8M0 FP8 scale is "for upcoming domestic chips"—fueling speculation about Huawei, Cambricon, Muxi, and Mofang Threads.
FP8 Breakthrough Highlights
- UE8M0 FP8 cuts memory use by up to 75% via dynamic mantissa allocation
- Block-level scaling boosts dynamic range (dozens of times) for large-model training
- Simplified critical path: only exponent shifting (no floating-point multiplication/normalization)
DeepSeek V3.1 becomes China’s first FP8 large-model training case, proving domestic hardware feasibility.
Domestic Chip Makers’ FP8 Moves
- Cambricon Siyuan 590: 40% higher computing density, FP8 support
- Mofang Threads: First domestic GPU maker with native FP8
- Hygon Deep Computing: 30% lower memory, 20% higher efficiency
- Huawei Ascend 910D: Integrates CloudMatrix 384 system, full-stack ecosystem
Capital Market Reaction
- Cambricon, Hygon shares hit +20% daily limit
- SMIC Hong Kong jumps 10.1% (Huawei Ascend/Kirin contract manufacturer)
This shows strong expectations for domestic AI computing power substitution.
ICgoodFind: FP8 advances speed domestic AI chip ecosystem closure; hardware-software synergy breaks computing power blockades.
