Huawei Ascend 9100 + Atlas 950 SuperNode Deep Dive: 300PFlops Training Chip, UnifiedBus 2.0 Optical Interconnect for 8192 NPUs

Abstract: In July 2026, Huawei officially unveiled the Ascend 9100 training chip (300 PFlops FP16 per card) alongside the Atlas 950 SuperNode—64 cards per cabinet, up to 8,192 NPU interconnects, and 16.3 PB/s interconnect bandwidth. This article dissects the chip architecture, supernode interconnect protocol, unified memory addressing, liquid cooling, and national substitution strategy, with complete Go/Python implementations.

1. Ascend 9100: The Performance Ceiling of Domestic Training Chips

1.1 Core Specifications

1.2 Mixed Precision Implementation

class MixedPrecisionEngine:
    """Ascend 9100 mixed precision engine supporting FP16/FP8/FP4"""
    
    def auto_select_precision(self, layer_type: str) -> str:
        if layer_type in ["attention_qkv", "attention_out"]:
            return "FP16"  # Precision-sensitive
        elif layer_type in ["ffn_gate", "ffn_up", "ffn_down"]:
            return "FP8"   # Compute-intensive
        elif layer_type == "moe_router":
            return "FP16"
        return "FP8"

2. Atlas 950 SuperNode: The World’s Most Powerful Computing System

2.1 Benchmark Comparison

2.2 UnifiedBus 2.0 Optical Interconnect

Atlas 950 SuperNode Topology:

  ┌──────────────┐     UB-Mesh     ┌──────────────┐
  │  Cabinet 0   │◄──────────────►│  Cabinet 1   │
  │  ┌─64 NPU─┐  │                 │  ┌─64 NPU─┐ │
  │  │ 950DT  │  │                 │  │ 950DT  │ │
  │  └────────┘  │                 │  └────────┘ │
  └──────┬───────┘                 └──────┬───────┘
         │                               │
         │    ┌─────────────────────┐    │
         └────│  Optical Mesh Switch │────┘
              │     16.3 PB/s       │
              └─────────────────────┘
  
  1 cabinet: 64x 950DT / 144GB each / 4TB/s BW
  Full cluster: 128 cabinets = 8,192 NPUs

2.3 Unified Memory Addressing

type UnifiedAddressSpace struct {
	nodeMemoryMap    map[uint64]*NodeMemory
	globalAddressMap map[uint64]uint64
	totalCapacity    uint64  // 1,152TB
}

func (uas *UnifiedAddressSpace) RemoteLoad(globalAddr uint64, size uint64) ([]byte, error) {
	// Direct load/store semantics for remote NPU memory
	// No serialization-network-deserialization pipeline
	// Single-hop latency: 200ns (vs 2μs for traditional networking)
	nodeID := uint32(globalAddr >> 48)
	localAddr := globalAddr & 0xFFFFFFFFFFFF
	return uas.nodeMemoryMap[uint64(nodeID)].RemoteRead(localAddr, size)
}

3. CANN Ops-Transformer: Hardware-Aware Operator Library

4. DeepSeek V4 Validation on Ascend

DeepSeek V4’s fine-grained expert parallelism verified on Ascend NPUs:

• General inference: 1.50-1.73x speedup vs non-fusion baseline
• RL rollout/Agent scenarios: up to 1.96x speedup

DeepSeek expects V4-Pro pricing to drop significantly once Ascend 950 SuperNodes enter mass production in H2 2026.

Sources:

• Ascend 9100 launch Weibo
• Atlas 950 SuperNode 证券时报
• Architecture deep dive CSDN

• ←Previous
• Next→