Crusoe's MLPerf Training v6.1 results on AMD Instinct MI355X
Crusoe submitted the largest AMD Instinct MI355X configuration in MLPerf Training v6.1, scaling Llama 3.1 8B, GPT-OSS 20B, and DLRMv4 across up to 16 nodes on standard RoCE Ethernet. Every result ran on virtualized Crusoe Cloud instances.

Today, MLCommons published the MLPerf® Training v6.1 results, including Crusoe's MLPerf Training submission on AMD Instinct™ MI355X. We submitted 10 results across DLRMv4, Llama 3.1 8B, GPT-OSS 20B, and FLUX.1, scaling from a single 8-GPU node to 16 nodes and 128 GPUs. Every result ran on virtualized Crusoe Cloud instances connected by standard RoCE Ethernet, the same instance types customers provision today.
Three results stand out:
- The fastest single-node MI355X DLRMv4 result in the round. One 8-GPU node trained DLRMv4 in 18.83 minutes. Scaling to 64 GPUs across eight nodes brought that down to 4.63 minutes.
- The largest MI355X configuration in the round. Crusoe scaled Llama 3.1 8B to 128 GPUs across 16 nodes, finishing in 14.07 minutes.
- The only multi-node MI355X results on Llama 3.1 8B and GPT-OSS 20B. Both scaled cleanly over RoCE Ethernet, showing that teams can grow MI355X clusters on open, widely available networking.
MLPerf Training v6.1 results on AMD Instinct MI355X
All results are MLPerf Training v6.1, closed division, on virtualized Crusoe Cloud mi355x-288gb-roce.8x instances. MLPerf Training reports time-to-train: wall-clock minutes to reach the benchmark's target quality, so lower is better.
Published results are available in the MLCommons MLPerf Training v6.1 results.
Software configuration
DLRMv4: recommendation training in under 5 minutes
A single MI355X node trained DLRMv4 in 18.83 minutes, the fastest single-node MI355X result for that benchmark in the round. Recommendation models drive search, ads, and feeds, and they retrain often, so time-to-train translates directly into how fresh a production model can be.
Scaling to 64 GPUs across eight nodes brought DLRMv4 to 4.63 minutes, 4.1x faster than one node. That is also the largest MI355X DLRMv4 configuration submitted in the round. Embedding-heavy models like DLRMv4 lean hard on the network, and these results ran over standard RoCE Ethernet with 3,200 Gbps per node.
Scaling MI355X to 128 GPUs on standard Ethernet
Crusoe submitted the largest MI355X configuration in the round: Llama 3.1 8B pretraining on 16 nodes and 128 GPUs. Time-to-train dropped from 78.34 minutes on one node to 43.23 on two, and to 14.07 minutes on 16 nodes, 5.6x faster than a single node.
Crusoe is also the only submitter with multi-node MI355X results on Llama 3.1 8B and GPT-OSS 20B. GPT-OSS 20B went from 58.04 minutes on one node to 39.74 on two and 27.16 on four, 2.1x faster on 4x the GPUs. On a single node, the virtualized GPT-OSS 20B result sits in line with bare-metal MI355X systems in the round.
Every multi-node run used RoCE over standard Ethernet, with 8x 400G NICs per node. Teams standardizing on AMD can scale on the same open networking they already know, without a proprietary fabric.
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 them performing throughout their production life. A training run across 128 GPUs is only as fast as its slowest GPU or link, so catching degradation 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 AMD Instinct MI355X on Crusoe Cloud scaling from one node to 16 over standard RoCE Ethernet, across recommendation, LLM pretraining, and image generation workloads. We plan to continue working with AMD on performance and software, and to publish results for both MLPerf Training and MLPerf Endpoints.
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 AMD Instinct MI355X on Crusoe Cloud.
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 AMD configurations did Crusoe submit in MLPerf Training v6.1?
Crusoe submitted AMD Instinct MI355X at 1, 2, 4, 8, and 16 nodes (8 to 128 GPUs). The benchmarks were DLRMv4, Llama 3.1 8B, GPT-OSS 20B, and FLUX.1.
Do I need InfiniBand to scale MI355X?
No. Every multi-node result ran on RoCE over standard Ethernet, with 8x 400G NICs per node, including the 128-GPU Llama 3.1 8B run.
Can I run on the same configurations?
Yes. AMD Instinct MI355X instances are available today on Crusoe Cloud, with Crusoe Managed Kubernetes, Managed Slurm, and AutoClusters. Talk to our team to get started.



