Huawei Ascend 950 (950PR / 950DT)
Product Overview
The Huawei Ascend 950 series (950PR / 950DT) is Ascend's fourth-generation AI chip, officially released and mass-produced in early 2026. It introduces FP8 / FP4 low-precision formats for the first time, with 1 PFLOPS of FP8 compute per card. It is also the first to adopt Huawei's in-house HBM (950PR uses HiBL 1.0, 950DT uses HiZQ 2.0), completely removing dependence on external HBM supply.
The 950PR (Prefill inference) and 950DT (Decode + training) are Huawei Ascend's first attempt at a "workload segmentation" strategy — same architecture, different memory configurations, precisely matched to the needs of different AI workloads.
Core Specifications
| Parameter | Ascend 950PR | Ascend 950DT |
|---|
| Architecture | Da Vinci v5 (4th-generation Ascend) | Da Vinci v5 (4th-generation Ascend) |
| Process | N+2 (SMIC domestic) | N+2 (SMIC domestic) |
| Programming Model | SIMD + SIMT dual model | SIMD + SIMT dual model |
| HBM Type | HiBL 1.0 (in-house, cost-optimized) | HiZQ 2.0 (in-house, bandwidth-optimized) |
| HBM Capacity | 128 GB | 144 GB |
| HBM Bandwidth | 1.6 TB/s | 4 TB/s |
| FP8 Compute | 1 PFLOPS (HiF8 format) | 1 PFLOPS (HiF8 format) |
| FP4 Compute | 2 PFLOPS | 2 PFLOPS |
| BF16/FP16 Compute | ~500 TFLOPS | ~500 TFLOPS |
| INT8 Compute | ~2,000 TOPS | ~2,000 TOPS |
| TDP | ~400 W | ~500 W |
| PCIe | Gen 5 ×16 | Gen 5 ×16 |
| Interconnect | LingQu 2 TB/s | LingQu 2 TB/s |
| Price | ¥70,000 per card (1/3 of the H200) | ~¥120,000-150,000 per card (estimated) |
| Launch | 2026 H1 | 2026 H1 |
Differentiated Positioning: 950PR vs 950DT
| Dimension | 950PR | 950DT |
|---|
| Target Scenario | Inference Prefill (first-token generation) | Inference Decode (token-by-token generation) + training |
| HBM Capacity | 128 GB (smaller; Prefill has low memory demand) | 144 GB (larger; training needs more memory) |
| HBM Bandwidth | 1.6 TB/s (moderate bandwidth demand for Prefill) | 4 TB/s (high bandwidth demand for Decode and training) |
| Typical Applications | Video recommendation, real-time interaction, search | Dialogue generation, text continuation, model training |
| Pricing | Lower (cost-optimized) | Higher (high performance) |
Key Technical Breakthroughs
1. In-House HBM (HiBL 1.0 / HiZQ 2.0)
- HiBL 1.0 (950PR): low-cost solution that resolves the HBM supply chokepoint problem
- HiZQ 2.0 (950DT): high-bandwidth solution, 4 TB/s rivaling NVIDIA HBM3e
- Fills a domestic gap, completely removing dependence on SK Hynix / Samsung
- Huawei's in-house FP8 variant with precision close to FP16 and 2x the compute of FP16
- Resolves the "low precision → accuracy loss" trade-off in large-model training
- Combined with the FP4 format, overall compute utilization improves by 30%+
3. SIMD + SIMT Dual Programming Model
- SIMD: efficient vector computation (continuing the Da Vinci core strength of the 910C)
- SIMT: newly added model supporting flexible scheduling
- Memory access granularity goes from 512 bytes → 128 bytes
- Discrete memory access efficiency improves 4x
- Offers twin ASIC (Ascend Core) and GPGPU variants
4. CloudMatrix 384 System
- 384 950 chips form a supernode
- Total compute: 384 × 1 PFLOPS FP8 ≈ 384 PFLOPS
- AI cluster performance surpasses the NVIDIA GB300 NVL72
Comparison with Previous-Generation 910C
| Metric | Ascend 910C | Ascend 950 | Improvement |
|---|
| Architecture | Da Vinci v4 (SIMD) | Da Vinci v5 (SIMD + SIMT) | Dual model |
| Process | 7nm-class | N+2 | More advanced |
| HBM | HBM2e (externally sourced) | In-house HiBL / HiZQ | Independent and controllable |
| HBM Capacity | 128 GB (dual chip) | 128-144 GB | Comparable |
| FP8/BF16 Compute | 780 TFLOPS BF16 | 1 PFLOPS FP8 | ~2.5x |
| FP4 Support | ❌ Not supported | ✅ Supported | New |
| TDP | ~310 W | ~400-500 W | +29-61% |
| Programming Model | Proprietary SIMD | SIMD + SIMT + ASIC/GPGPU twins | Open |
| In-House HBM | ❌ Externally sourced | ✅ In-house | Milestone |
Use Cases
- ✅ Large-model training (950DT, 1 PFLOPS FP8)
- ✅ Inference Prefill (950PR, video recommendation / real-time interaction)
- ✅ Inference Decode (950DT, dialogue generation / text continuation)
- ✅ Domestic AI clusters (CloudMatrix 384)
- ✅ Government and enterprise AI infrastructure (in-house HBM, supply chain security)
- ❌ CUDA ecosystem lock-in (migration to CANN / MindSpore required)
- ❌ FP4 training (the 950 supports FP4 for inference; training remains primarily FP8/BF16)