Engineering
Cloud
November 16, 2026

Crusoe's MLPerf Training v6.1 results on NVIDIA GB300 NVL72 and B300

Crusoe's first MLPerf Training v6.1 submission benchmarked FLUX.1, Llama 2 70B LoRA, Llama 3.1 8B, and GPT-OSS 20B on NVIDIA GB300 NVL72 and HGX B300. Every result ran on the same virtualized Crusoe Cloud instances customers use today.

Young Jeong Photo
Young Jeong
Staff Solutions Engineer
Sanchit Pathak headshot
Sanchit Pathak
Senior Solutions Engineer
Alex Akesson photo
Alex Akesson
Director, Product Management
November 16, 2026
Lime illustration of converging lines fanning out, representing MLPerf Training scaling on NVIDIA GPUs on Crusoe Cloud

Today, MLCommons published Crusoe’s MLPerf® Training v6.1 results. On NVIDIA GB300 NVL72 and NVIDIA HGX B300, we submitted 10 results across FLUX.1, Llama 2 70B LoRA, Llama 3.1 8B, and GPT-OSS 20B. Every result ran on virtualized Crusoe Cloud instances, the same ones customers run on today.

Three results stand out:

  • FLUX.1 in 23.6 minutes on a full GB300 NVL72 rack. All 72 Blackwell Ultra GPUs, running virtualized, trained FLUX.1 in 23.63 minutes, on par with NVIDIA’s own 72-GPU submission at 23.68 minutes.
  • GPT-OSS 20B on 128 Blackwell Ultra GPUs in 15.4 minutes. This is the largest HGX B300 configuration submitted for GPT-OSS 20B in the round.
  • Bare-metal-class performance on virtualized nodes. GPT-OSS 20B on 64 Blackwell Ultra GPUs ran on par with bare-metal, on-premises HGX B300 systems at the same scale.

MLPerf Training v6.1 results on NVIDIA GB300 NVL72 and HGX B300

All results are MLPerf Training v6.1, closed division, on virtualized Crusoe Cloud instances. MLPerf Training reports time-to-train: wall-clock minutes to reach the benchmark’s target quality. The faster the time-to-train, the better the result is.

BenchmarkSystemNodes x GPUsTime-to-train (minutes)
FLUX.1GB300 NVL722 x 4 (8 GPUs)132.54
GB300 NVL724 x 4 (16 GPUs)96.91
GB300 NVL7218 x 4 (72 GPUs, full rack)23.63
Llama 2 70B LoRAGB300 NVL722 x 4 (8 GPUs)5.82
HGX B3001 x 8 (8 GPUs)6.82
HGX B3002 x 8 (16 GPUs)3.93
HGX B3004 x 8 (32 GPUs)2.55
Llama 3.1 8BHGX B3001 x 8 (8 GPUs)75.72
GPT-OSS 20BHGX B3008 x 8 (64 GPUs)22.94
HGX B30016 x 8 (128 GPUs)15.41

Published results are available in the MLCommons MLPerf Trining v6.1 results.

Software configuration

ComponentGB300 NVL72B300
Training stackPyTorch, NVIDIA NGC MLPerf Training v6.1 containerNVIDIA NeMo™ Framework (Megatron-Bridge) on NGC PyTorch 26.04
Node4x NVIDIA GB300, 2x NVIDIA Grace™ CPU (Arm Neoverse V2)b300-288gb-sxm-ib.8x: 8x NVIDIA B300 SXM 288GB, 2x Intel Xeon 6776P, 120 vCPUs
FabricNVIDIA NVLink™ across the 72-GPU rackNVIDIA Quantum-X800 InfiniBand switching between nodes
VirtualizationKVM, full 1:1 GPU passthroughcloud-hypervisor/KVM, driver 580.167.08 on every node

Full-rack GB300 NVL72: FLUX.1 at NVIDIA reference speed

On a full GB300 NVL72 rack, Crusoe trained FLUX.1 in 23.63 minutes, matching NVIDIA’s own 72-GPU submission (23.68 minutes). The rack is 18 nodes of 4 GPUs each, all 72 GPUs in a single NVIDIA NVLink domain, so gradients move over NVLink instead of crossing a network fabric.

Scaling from 8 GPUs to the full rack took FLUX.1 from 132.54 minutes to 23.63 minutes. Each GPU is passed through 1:1 to the VM so virtualization costs nothing measurable at rack scale. Customers training on Crusoe Cloud get the same performance NVIDIA measures on its own systems.

On 8 Blackwell Ultra GPUs on GB300 NVL72, Llama 2 70B LoRA finished in 5.82 minutes, in line with other 8-GPU GB300 submissions in the round (5.62 to 5.90 minutes). LoRA fine-tuning is how most teams adapt a foundation model to their own data, so this number maps directly to iteration speed.

Scaling on NVIDIA B300 to 128 GPUs

On HGX B300, Crusoe submitted GPT-OSS 20B at 64 and 128 GPUs, the largest HGX B300 configuration for that benchmark in the round. Time-to-train dropped from 22.94 minutes on 8 nodes to 15.41 minutes on 16 nodes. The 64-GPU result runs on par with bare-metal, on-premises B300 systems at the same scale, on virtualized cloud instances.

Llama 2 70B LoRA scaled from one node to four: 6.82 minutes on 8 GPUs, 3.93 on 16, and 2.55 on 32. That is 2.7x faster on four times the GPUs.

Scaling results matter more to most teams than a single headline number. Customers rarely start at full size: they prove out a model on one node, then grow. Consistent scaling across NVIDIA Quantum-X800 InfiniBand means the performance a team measures early holds as the cluster expands.

Benchmark results start with fleet health

Every result in this submission was produced on virtualized Crusoe Cloud instances, the same ones customers provision today. There was no special bare-metal benchmark configuration.

Results like these start long before benchmark day. Crusoe clusters pass comprehensive burn-in before delivery, and continuous fleet health monitoring keeps it performing throughout its production life. A training run that spans 72 or 128 GPUs is only as fast as its slowest component, so catching an underperforming GPU or link before it reaches a customer is what makes these numbers repeatable.

On top of healthy hardware, Crusoe Managed Kubernetes, Crusoe Managed Slurm, and fault-tolerant AutoClusters remove the operational overhead of running AI at scale.

What comes next

This submission shows the NVIDIA Accelerated Computing Platform on Crusoe Cloud at two scales: a full GB300 NVL72 rack matching NVIDIA’s own FLUX.1 result, and HGX B300 clusters scaling GPT-OSS 20B to 128 GPUs. We plan to continue working with NVIDIA on performance across GB300 NVL72, NVIDIA Vera Rubin NVL72, and beyond, and to publish results for both Training and Inference.

Crusoe is the energy-first AI factory company, designing, building, and operating the AI factories, energy, and cloud platform behind next-generation AI. The configurations submitted here are available today. Talk to our team about training on NVIDIA GB300 NVL72 or NVIDIA B300 on Crusoe Cloud.

Get Crusoe stories in your Google results

Frequently
asked questions

What is the MLPerf Training benchmark?

The MLPerf Training benchmark is an industry benchmark suite from MLCommons that measures how long a system takes to train a model to a fixed target quality. Submitters review each other’s results before publication, so results can be compared across vendors and systems.

Which NVIDIA systems did Crusoe submit in MLPerf Training v6.1?

Crusoe submitted NVIDIA GB300 NVL72 at 8, 16, and 72 GPUs, and NVIDIA HGX B300 at 8 to 128 GPUs. The benchmarks were FLUX.1, Llama 2 70B LoRA, Llama 3.1 8B, and GPT-OSS 20B.

Can I run on the same configurations?

Yes. NVIDIA GB300 NVL72 and B300 instances are available today on Crusoe Cloud, with Crusoe Managed Kubernetes, Managed Slurm, and AutoClusters. Talk to our team to get started.

Latest articles

Chase Lochmiller - Co-founder, CEO
November 16, 2026
Crusoe's MLPerf Training v6.1 results on AMD Instinct MI355X
Chase Lochmiller - Co-founder, CEO
November 16, 2026
Crusoe's MLPerf Training v6.1 results on NVIDIA GB300 NVL72 and B300
Chase Lochmiller - Co-founder, CEO
November 10, 2026
NVIDIA and Crusoe cut GLM-5.3 startup from 8 minutes to 31 seconds with NVIDIA Snapshot

Are you ready to build something amazing?