AI Practices AI实践 5h ago Updated 1h ago 更新于 1小时前 53

Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72 在 NVIDIA GB300 NVL72 上设定 MoE 预训练世界纪录

NVIDIA GB300 NVL72 achieved a world record pre-training performance of 1,648 TFLOPs per GPU on the DeepSeek-V3 671B model, representing a ~3x improvement over previous generations. The system leverages fifth-generation NVLink to provide 1.8 TB/s per-GPU bandwidth and 130 TB/s non-blocking all-to-all bandwidth, critical for efficient Mixture of Experts (MoE) scaling. MoE architectures like DeepSeek-V3 offer massive computational efficiency by activating only a subset of parameters (~37B active ou NVIDIA GB300 NVL72在预训练DeepSeek-V3 671B模型时创下世界纪录,单GPU吞吐量达1,648 TFLOPs,较上一代GB200提升约3倍。 系统采用第五代NVLink技术,提供1.8 TB/s的单GPU带宽和130 TB/s的非阻塞全对全带宽,实现机架内极低延迟的内存语义通信。 MoE架构成为前沿模型主流,其通信密集型特性使得网络带宽和低延迟成为决定扩展效率的关键瓶颈,而非单纯算力。 通过硬件(Blackwell Ultra GPU)、互连(NVLink/ConnectX-8)与软件(Megatron Core/TorchTitan/JAX)的深度协同设计,实现了

75
Hot 热度
70
Quality 质量
85
Impact 影响力

Analysis 深度分析

TL;DR

  • NVIDIA GB300 NVL72 achieved a world record pre-training performance of 1,648 TFLOPs per GPU on the DeepSeek-V3 671B model, representing a ~3x improvement over previous generations.
  • The system leverages fifth-generation NVLink to provide 1.8 TB/s per-GPU bandwidth and 130 TB/s non-blocking all-to-all bandwidth, critical for efficient Mixture of Experts (MoE) scaling.
  • MoE architectures like DeepSeek-V3 offer massive computational efficiency by activating only a subset of parameters (~37B active out of 671B total), shifting the primary bottleneck from compute to communication.
  • Success relies on extreme hardware-software co-design, integrating NVIDIA Megatron Core, TorchTitan, and JAX optimizations with ConnectX-8 SuperNICs and Quantum-X800 InfiniBand for scale-out networking.
  • Per-GPU throughput remains above 97% when scaling from 256 to 1,024 GPUs, demonstrating the effectiveness of the tightly coupled scale-up and scale-out infrastructure.

Why It Matters

This achievement highlights that as AI models converge on Mixture of Experts (MoE) architectures, communication efficiency has become the primary determinant of training scalability, surpassing raw compute power. For AI practitioners and researchers, it underscores the necessity of investing in specialized, co-designed infrastructure that optimizes both intra-rack (scale-up) and inter-rack (scale-out) networking to handle the heavy all-to-all communication patterns inherent in MoE training.

Technical Details

  • Hardware Architecture: The NVIDIA GB300 NVL72 utilizes 72 Blackwell Ultra GPUs connected via fifth-generation NVLink, providing 1.8 TB/s per-GPU bandwidth and 130 TB/s non-blocking all-to-all bandwidth across the rack.
  • Networking Stack: Scale-out connectivity is handled by NVIDIA ConnectX-8 SuperNICs at 800 Gbps per GPU, integrated with Quantum-X800 InfiniBand or Spectrum-X Ethernet to ensure predictable performance across multiple racks.
  • Model Specifications: The benchmark used DeepSeek-V3, a 671B parameter MoE model that activates approximately 37B parameters per token, significantly reducing per-token compute costs compared to dense models.
  • Software Optimization: Performance was driven by NVIDIA Megatron Core, with additional support from TorchTitan and JAX, achieving 3x-10x performance improvements over previous generations through continuous software-hardware co-design.
  • Infrastructure Processing: NVIDIA BlueField DPUs offload infrastructure tasks such as virtual networking, storage access, and telemetry, reducing host CPU overhead and improving overall training job efficiency.

Industry Insight

  • Communication is the New Bottleneck: As MoE adoption accelerates, infrastructure planning must prioritize network bandwidth and latency, particularly for all-to-all collectives, over pure FLOP counts.
  • Co-Design is Essential: Isolated hardware upgrades yield diminishing returns; maximum efficiency requires tight integration between silicon, interconnects, and training frameworks (e.g., Megatron Core).
  • Scalability Validation: The ability to maintain >97% throughput scaling from 256 to 1,024 GPUs suggests that current architectural trends can support even larger frontier models without linear increases in training time.

TL;DR

  • NVIDIA GB300 NVL72在预训练DeepSeek-V3 671B模型时创下世界纪录,单GPU吞吐量达1,648 TFLOPs,较上一代GB200提升约3倍。
  • 系统采用第五代NVLink技术,提供1.8 TB/s的单GPU带宽和130 TB/s的非阻塞全对全带宽,实现机架内极低延迟的内存语义通信。
  • MoE架构成为前沿模型主流,其通信密集型特性使得网络带宽和低延迟成为决定扩展效率的关键瓶颈,而非单纯算力。
  • 通过硬件(Blackwell Ultra GPU)、互连(NVLink/ConnectX-8)与软件(Megatron Core/TorchTitan/JAX)的深度协同设计,实现了从256到1024 GPU扩展时超过97%的单GPU吞吐量保持率。

为什么值得看

本文揭示了AI基础设施竞争的核心已从单一芯片算力转向“计算+通信+软件”的全栈协同能力,特别是针对MoE架构特有的通信瓶颈提供了工程解决方案。对于AI从业者和基础设施提供商而言,理解如何通过非阻塞网络和内存语义互连来优化All-to-All通信模式,是构建下一代高效大模型训练平台的关键。

技术解析

  • 性能基准与突破:在DeepSeek-V3 671B(激活参数~37B,总参数671B)的预训练中,GB300 NVL72在256个GPU上达到1,648 TFLOPs/GPU,相比GB200 NVL72的606 TFLOPs/GPU实现了近3倍的性能飞跃,证明了新平台在处理大规模MoE模型时的卓越效率。
  • 第五代NVLink架构:核心创新在于第五代NVLink提供的1.8 TB/s单GPU带宽和130 TB/s机架级非阻塞全对全带宽。这种架构允许每个GPU在一个跳数内与其他所有GPU通信,且支持内存语义操作(直接读写对端HBM),消除了传统软件发送带来的延迟开销,使All-to-All通信能隐藏在计算之后。
  • MoE通信挑战与优化:MoE模型每层都需要进行复杂的Token分发和聚合(All-to-All),这构成了训练的关键路径。GB300 NVL72通过高带宽、低延迟的机架内Scale-up连接处理高频的MoE通信,同时利用ConnectX-8 SuperNICs(800 Gbps)和Quantum-X800 InfiniBand/Spectrum-X Ethernet处理跨机架的Scale-out梯度通信,确保整体训练步长不受限于网络延迟。
  • 软硬协同与软件栈:除了硬件,NVIDIA通过Megatron Core、TorchTitan和JAX等软件框架的深度优化,释放了硬件潜力。BlueField DPUs被用于卸载虚拟化、存储和安全等基础设施任务,减少主机CPU开销,进一步提升了大规模集群的训练稳定性和效率。

行业启示

  • 通信带宽即算力:随着MoE架构的普及,模型扩展的效率瓶颈已从计算转移至通信。未来的AI基础设施竞争将聚焦于如何提供更宽、更低延迟的非阻塞网络拓扑,以掩盖All-to-All等高通信开销操作的延迟。
  • 全栈协同设计的必要性:单纯堆砌GPU芯片无法实现线性扩展。必须像NVIDIA GB300 NVL72那样,将硅片、互连、网络交换机和训练软件作为统一平台进行联合设计与优化,才能在大规模集群中维持高吞吐量(如>97%的扩展效率)。
  • 基础设施专业化趋势:大型AI训练工厂需要专用的基础设施处理单元(如DPUs)来隔离和管理元数据、网络虚拟化和存储访问,从而让主计算节点专注于模型训练。这种分层解耦的设计将成为超大规模AI训练的标准配置。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPU GPU Training 训练 Research 科学研究