SetupAI
EN
來源較早收集於 2m

Huawei 晶片成功訓練 DeepSeek-V4-Pro 模型

Huawei 晶片成功訓練 DeepSeek-V4-Pro 模型
PostLinkedIn
🇭🇰閱讀原文： SCMP Technology
#semiconductors#china-techascend-910chuaweideepseekascend 910cnvidia

💡Huawei 晶片現已能訓練複雜模型，顯示全球 AI 硬體供應鏈出現重大轉變。

⚡ 30 秒速覽

有什麼變化

Huawei Ascend 910C 晶片被用於 DeepSeek-V4-Pro 模型的後訓練。

為什麼重要

此發展顯示中國國產硬體正成為訓練大型模型的可行替代方案，可能改變該地區 AI 基礎設施的競爭格局。

下一步行動

監控 Ascend 910C 與 Nvidia H100 的效能基準測試，以評估您的基礎設施策略是否需要考慮非美國硬體的替代方案。

誰應關注：Researchers & Academics

關鍵要點

  • •Huawei Ascend 910C 晶片被用於 DeepSeek-V4-Pro 模型的後訓練。
  • •此里程碑展示了中國在國內進行複雜 AI 訓練能力的進步。
  • •這降低了先進 AI 開發對美國製造的高階 GPU 的依賴。

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 24 個來源。

🔑 增強重點摘要

  • •The DeepSeek-V4-Pro model is a Mixture-of-Experts (MoE) architecture featuring 1.6 trillion total parameters and 49 billion activated parameters, designed to support a substantial 1 million token context window.
  • •Huawei's Ascend 910C chip utilizes a dual-die packaging design, integrating two Ascend 910B SoCs, and is manufactured using SMIC's 7nm (N+2) process technology.
  • •DeepSeek's AI models offer native support for Huawei's Ascend processors, facilitating a straightforward conversion from CUDA to CUNN, which is vital for seamless integration of Huawei's hardware into AI development workflows.
  • •The United States has recently intensified its export controls, issuing new guidance to prevent Chinese companies from circumventing restrictions by accessing advanced American AI chips, such as Nvidia's Blackwell and Rubin, and AMD's MI350X, through overseas subsidiaries.
  • •China has officially recognized AI training and inference chips, including Huawei's Ascend 910, within its 'secure and reliable' technology assessment framework, underscoring a state-backed initiative to promote domestic alternatives.
📊 競品分析▸ Show
Feature/MetricHuawei Ascend 910CNvidia H100 (Hopper)Nvidia A100 (Ampere)
FP16 Performance800 TFLOPS~1000 TFLOPS (implied from 80% comparison)320 TFLOPS
Inference Performance~60% of H100Higher than 910CMore suitable for LLMs beyond 10B parameters
Memory Bandwidth3.2 TB/sHigher than 910C (e.g., HBM3/HBM3e)Higher than 910C
Process TechnologySMIC 7nm (N+2)TSMC 4N (custom 5nm class)TSMC 7nm
Logic Die Area~60% larger than H100 for 80% performanceSmaller than 910C for higher performanceEfficient design
TDP550 WPrioritizes raw power, higher energy use
InterconnectHCCL, PCIe 4.0, 512 GB/s meshNVLink, high die-to-die bandwidthNVLink
Software EcosystemMindSpore (growing)CUDA, TensorRT (well-established)CUDA, TensorRT (well-established)
Training ReliabilityCritical weakness for long-term trainingUndisputed leadStrong
Estimated Price~$28,000 (similar to H100)~$28,000

🛠️ 技術深入

  • DeepSeek-V4-Pro Model:
    • Architecture: Mixture-of-Experts (MoE) with a novel Hybrid Attention mechanism, combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).
    • Parameters: 1.6 trillion total parameters with 49 billion activated parameters per token.
    • Context Window: Supports a maximum context length of 1 million tokens.
    • Training Data: Pre-trained on over 32 trillion diverse and high-quality tokens, with a focus on long documents and agentic execution traces.
    • Post-training: Employs a two-stage pipeline: initial independent cultivation of domain-specific experts (via SFT and RL with GRPO), followed by unified model consolidation through on-policy distillation.
    • Optimizer: Utilizes the Muon optimizer for enhanced convergence speed and training stability.
    • Efficiency: Achieves significant efficiency gains, requiring only 27% of single-token inference FLOPs and 10% of KV cache compared to DeepSeek-V3.2 at a 1M-token context.
    • Precision: MoE expert parameters are in FP4 precision, while most other parameters use FP8.
  • Huawei Ascend 910C Chips:
    • Architecture: Part of the DaVinci family, integrating multiple AI cores, each comprising high-throughput matrix-multiply cube units ('AIC'), wide vector SIMD engines ('AIV'), scalar units, and multi-level scratchpad memories.
    • Process Technology: Manufactured using SMIC's 7nm (N+2) process.
    • Transistor Count: Features approximately 53 billion transistors.
    • Packaging: Employs a dual-die packaging design, effectively co-packaging two Ascend 910B SoCs, interconnected via an organic substrate.
    • Peak Throughput: Delivers 800 TFLOPS (BF16/FP16), 100 TFLOPS (FP32), and 800 TOPS (INT8) for tensor operations.
    • Memory Bandwidth: Provides 3.2 TB/s memory bandwidth.
    • Interconnects: Features a 512 GB/s bidirectional mesh network across AI cores, PCIe 4.0 ×16 (64 GB/s) for host-NPU communication, and Distributed HCCL (Huawei Collective Communication Library) over PCIe or TCP for multi-node clusters; Huawei Collective Communication System (HCCS) is designed to rival NVIDIA's NVLink.
    • Thermal Design Power (TDP): Rated at 550 W.

🔮 前景展望基於引用來源的 AI 分析

China will significantly accelerate its domestic AI chip and model development, further reducing reliance on foreign technology.
The successful training of a large model like DeepSeek-V4-Pro on Huawei chips, coupled with tightening US export controls and China's 'secure and reliable' technology assessments, will intensify domestic efforts to achieve self-sufficiency.
Huawei's Ascend ecosystem will capture a substantial share of the AI training and inference market within China.
The demonstrated capability of Ascend 910C for complex model training, alongside native software support from major Chinese AI developers like DeepSeek, positions Huawei as a leading domestic alternative to Nvidia.
The global AI landscape will experience increased fragmentation in hardware and software ecosystems.
As China develops its own full-stack AI solutions, including chips, frameworks like MindSpore, and advanced models, it will foster a distinct ecosystem, potentially leading to reduced interoperability with Western-dominated platforms such as CUDA.

⏳ 時間線

2018
Huawei announces Ascend 910 and Ascend 310 AI chips.
2019-08
Huawei officially launches the Ascend 910 processor.
2022-10
US implements new export controls on semiconductor exports to China.
2023-07
DeepSeek, an AI company, is founded by Liang Wenfeng.
2025-03
Huawei officially launches the Atlas 900 A3 SuperPoD, incorporating Ascend 910C chips.
2026-04-24
DeepSeek-V4-Pro model is released under the MIT license.
🇭🇰閱讀原文： SCMP Technology
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #semiconductors

同產品

更多關於 ascend-910c

同來源

最新來自 SCMP Technology

Alibaba 將 AI「數位員工」部署至競爭對手應用程式

Alibaba 將 AI「數位員工」部署至競爭對手應用程式

SCMP Technology•Sep 9

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： SCMP Technology ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週電子報

每週一封，可隨時退訂。