Huawei Ascend 970 (Roadmap)
Product Overview
Huawei Ascend 970 is the long-range roadmap flagship AI chip announced at HUAWEI CONNECT 2025, planned for Q4 2028, continuing the 950 → 960 → 970 one-chip-per-year, compute-doubling cadence.
The Ascend 970 aims squarely at trillion-parameter MoE architectures and the AGI era, with order-of-magnitude leaps in FP4/FP8 compute, interconnect bandwidth, and memory bandwidth — Huawei's planned "ultimate flagship" AI chip.
📌 This card is a roadmap preview: the single-chip specs below are official Huawei planning targets and may change after formal launch.
Core Specifications (roadmap)
| Parameter | Value (planned) |
|---|---|
| Architecture | Da Vinci v7 (presumed, SIMD + SIMT) |
| Process | N+3 (SMIC domestic, presumed) |
| FP8 Compute | 4 PFLOPS |
| FP4 Compute | 8 PFLOPS |
| BF16/FP16 Compute | 2 PFLOPS (planned) |
| INT8 Compute | 4 POPS (planned) |
| Memory | 288 GB HBM (in-house) |
| Memory Bandwidth | 14.4 TB/s (~18x the 910C's 784 GB/s) |
| Interconnect Bandwidth | 4 TB/s (~10x the 910C's 400 Gbps) |
| Package | Quad-die co-packaging (660mm², planned) |
| TDP | 800 W (presumed, roadmap) |
| Debut | 2028 Q4 (roadmap) |
| Systems | Atlas 960 SuperPoD (15488-card cluster) |
Generational Comparison with 950/960
| Model | FP8 | FP4 | Memory | Memory Bandwidth | Interconnect | Planned |
|---|---|---|---|---|---|---|
| Ascend 950PR/DT | 1 PFLOPS | 2 PFLOPS | 128-144 GB | 1.6-4 TB/s | 2 TB/s | 2026 |
| Ascend 960 | 2 PFLOPS | 4 PFLOPS | 288 GB | 9.6 TB/s | 2.2 TB/s | 2027 Q4 |
| Ascend 970 | 4 PFLOPS | 8 PFLOPS | 288 GB | 14.4 TB/s | 4 TB/s | 2028 Q4 |
Key Technology Directions
- HiF4 data format: Huawei's in-house 4-bit precision, with better inference accuracy than industry FP4 schemes
- Dynamic sparse computing + MoE adaptation: for trillion-parameter mixture-of-experts architectures
- Quad-die co-packaging: breaks single-die area limits, doubling compute
- UnifiedBus interconnect: interconnect bandwidth ~10x the 910C
Use Cases (planned)
- ✅ Trillion-parameter MoE LLM training / inference
- ✅ National intelligent computing centers (15488-card Atlas 960 SuperPoD)
- ✅ AGI-era hyperscale AI infrastructure
Vendor Information
| Parameter | Value |
|---|---|
| Manufacturer | Huawei Technologies Co., Ltd. (HiSilicon) |
| Website | https://www.hiascend.com |
| Announced | HUAWEI CONNECT 2025 (Xu Zhijun keynote) |
| Debut (planned) | 2028 Q4 |
Related Products
- Huawei Ascend 960 - Previous roadmap generation (15488-card SuperPod)
- Huawei Ascend 950DT - Current training/Decode mainstay plan
- Full comparison table