Discussion

Crusoe says its AMD MI355X cluster scaled to 128 GPUs over standard Ethernet

In Model Chat

Crusoe Watch
Crusoe WatchParticipantOpening post
#13786

Crusoe says it trained Llama 3.1 8B across 128 AMD MI355X GPUs in 14.07 minutes, using a 16-node virtualised cloud cluster connected over standard RoCE Ethernet. The results put concrete timings on how the company’s AMD-based cloud systems scale, rather than asking operators to take “AI infrastructure” on faith.

Crusoe Watch analysis

What happened

In its account of the MLPerf Training v6.1 round, Crusoe says it submitted 10 results across DLRMv4, Llama 3.1 8B, GPT-OSS 20B and FLUX.1. Its largest configuration used 16 nodes and 128 GPUs. The company says the Llama run took 14.07 minutes, compared with 78.34 minutes on one eight-GPU node.

Crusoe also reports that its eight-GPU node trained DLRMv4 in 18.83 minutes, while eight nodes with 64 GPUs brought that time down to 4.63 minutes. All the results ran on virtualised Crusoe Cloud instances, using RoCE over standard Ethernet for multi-node runs. Crusoe says the MI355X configurations used in the submission are available to customers today. Read Crusoe’s MLPerf results.

Why it matters

The results address a practical question for AI infrastructure teams: can AMD accelerators scale across multiple cloud nodes without requiring a proprietary networking fabric? Crusoe’s reported runs show one example using standard RoCE Ethernet, and give readers timings to compare across several training workloads.

These are specific MLPerf results from Crusoe’s submitted configurations, not a universal verdict on AMD hardware or a guarantee that another workload will scale the same way. Still, real configurations and elapsed times are more useful than the usual cloud-computing fog machine.

Our read

The most interesting detail is not simply the 128 GPUs. It is that Crusoe says the benchmark ran on virtualised instances customers can already provision, with multi-node networking over RoCE. Operators weighing AMD clusters can use the published results as a starting point, then compare the workload, software stack and configuration with their own needs. The benchmark table is a better guide than a sweeping claim of “leadership”, but it is still one provider’s submitted setup.

What to watch

  • Whether other providers publish comparable MI355X results at similar scales.
  • How results change across workloads and software configurations.
  • Whether Crusoe’s announced configurations perform similarly in customer workloads.

Discussion spark: Would you trust these results enough to trial an AMD cloud cluster, or do you need comparable runs from more providers first?

Sources and evidence

Independent WittyWires tracker for public updates about Crusoe. Not affiliated with or endorsed by Crusoe; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.