お知らせ • Jun 18
CoreWeave Sets New AI Training Records In MLPerf Training v6.0 And Trains DeepSeek-V3 In Approximately Two Minutes
CoreWeave, Inc. announced record-breaking results in the MLPerf Training v6.0 benchmark suite. Running on the same CoreWeave Cloud infrastructure available to customers, CoreWeave delivered the fastest DeepSeek-V3 671B training performance in the benchmark, training one of the most computationally demanding models ever benchmarked in 2.02 minutes on 8,192 NVIDIA GB300 NVL72 GPUs — the largest GB300 cluster submitted in this round. CoreWeave submitted three GB300 NVL72 configurations on DeepSeek-V3 671B, the benchmark's most demanding workload, achieving the fastest results across all Closed/Available-cloud submissions. On 8,192 GPUs across 2,048 nodes, CoreWeave hit target quality in approximately two minutes. Scaling down to 4,096 GPUs across 1,024 nodes, training was completed in 3.09 minutes. At 2,048 GPUs across 512 nodes, the result was 5.54 minutes. As the cluster size doubled at each step, training time improved predictably — a consistent, near-linear scaling efficiency that reflects full-stack optimization across every layer of the CoreWeave platform. CoreWeave was the only submitter in the v6.0 round to scale a GB300 platform beyond 2,048 GPUs on DeepSeek-V3. The scaling story is as significant as the result demonstrating that full-stack optimization delivers more usable performance per GPU than raw scale alone. For AI teams operating under compute budgets, that scaling curve translates directly into faster training runs, shorter development cycles, and quicker time to production. CoreWeave's MLPerf Training v6.0 results demonstrate that full-stack infrastructure advantages extend across deployment sizes, not just at frontier scale. On NVIDIA GB300 NVL72, CoreWeave's 4,096-GPU deployment reached the Llama-3.1-405B reference quality target in 9.77 minutes, achieving near-parity with larger GB200 deployments while using 20% fewer GPUs. The run was built on NVIDIA NeMo Framework Release 26.04, with full CUDA graphs, Tensor/pipeline/context-parallel sharding tailored to the GB300 NVL72 topology, and NVIDIA Spectrum-X Ethernet running RoCE for scale-out fabric. On a compact 8-node, 64-GPU NVIDIA HGX B200 cluster connected via InfiniBand, CoreWeave trained GPT-OSS-20B in 26.98 minutes and Llama-3.1-8B in 16.54 minutes. Through optimizations in orchestration, communication libraries, and distributed training configuration, CoreWeave delivered performance from the B200 platform that rivals larger and newer-generation deployments. This validated that CoreWeave's engineering advantages benefit customers at every scale, not just the largest clusters. CoreWeave's MLPerf Training v6.0 results reflect optimizations across every layer of the stack: Fleet-Wide Performance Consistency: CoreWeave Mission ControlTM continuously performs health checks across the latest rack scale systems like GB300, validating hardware, firmware, network, and thermal health before and during large-scale training jobs. This reduces stragglers and ensures workloads run on a consistent, performance-qualified infrastructure baseline. NVLink-Domain-Aware Scheduling: CoreWeave SUNK is topology-aware by design, intelligently placing workloads to maximize locality and co-locating expert-parallel groups within the same NVL72 domain to minimize inter-rack communication for MoE workloads. Optimized Network Performance: CoreWeave employs a rail-aware networking strategy that balances traffic, ensuring bandwidth is utilized efficiently and preventing hotspots from developing within the fabric at multi-thousand-GPU scale. CoreWeave's MLPerf Training v6.0 results were achieved on the same production infrastructure available to customers. The networking fabric, scheduler, storage architecture, and CoreWeave Mission Control orchestration platform used in MLPerf are the same systems customers use to run real-world workloads. This was not a benchmark-only environment, it was a validation of the platform customers can access.