AI Research

Crusoe proves MI355X scales without InfiniBand

Crusoe ran MLPerf Training v6.1 on MI355X to 128 GPUs using RoCE Ethernet and virtualized cloud instances.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

2 min read

Crusoe Cloud MI355X cluster running MLPerf Training v6.1
Credit: Crusoe Blog

Crusoe submitted 10 MLPerf Training v6.1 results on AMD Instinct MI355X, topping out at 128 GPUs across 16 nodes.

Every run used virtualized mi355x-288gb-roce.8x instances on standard RoCE Ethernet, Crusoe says, the same shape customers rent today. MLPerf reports time to train to fixed quality, so lower wins. Crusoe says DLRMv4 hit 18.83 minutes on one 8-GPU node and 4.63 minutes on eight nodes, while Llama 3.1 8B fell from 78.34 minutes on one node to 14.07 minutes on 16 nodes.

That DLRMv4 single-node number was the fastest MI355X result in the round.

Recommendation teams feel that first. DLRMv4 powers search, ads and feeds, and it retrains often. Crusoe says scaling from 8 to 64 GPUs cut training to 4.63 minutes, or 4.1x faster, over 3,200 Gbps per node from 8x400G NICs. Embedding-heavy models punish weak networks, so running that on open Ethernet matters more than the headline time.

LLM scaling tells a similar story with a tax. Crusoe says Llama 3.1 8B moved from 78.34 minutes to 43.23 minutes on two nodes and 14.07 minutes on 16 nodes, or 5.6x faster on 16x GPUs. Crusoe says GPT-OSS 20B went from 58.04 minutes on one node to 39.74 on two and 27.16 on four, or 2.1x faster on 4x GPUs. Crusoe was the only submitter with multi-node MI355X results on both models, and Crusoe says the single-node GPT-OSS run sits in line with bare-metal MI355X systems.

Memory is why single-node holds up. One MI355X carries 288GB of VRAM, enough for roughly 460B parameters at 4-bit or 125B at 16-bit at 32K context, per GetDeploying rental data. Crusoe paired that with AMD Primus on Megatron-LM, PyTorch 2.10 for ROCm, TransformerEngine for ROCm and AITER, with MXFP4 GEMMs for LLMs. You don't get a special benchmark hypervisor here.

Fleet work decides whether 128 GPUs stay fast. A 16-node run is only as quick as its slowest GPU or link. Crusoe says clusters pass burn-in before delivery with continuous health monitoring, plus Managed Kubernetes, Managed Slurm and fault-tolerant AutoClusters to replace failed nodes without restarting operations from scratch.

Crusoe ran the same play on Nvidia hardware with its 72-GPU Nvidia submission in the same round. AMD shops standardizing on Ethernet can now scale MI355X without buying a proprietary fabric, but they should budget for sublinear returns past four nodes.

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsCrusoeAMDMI355XMLPerfLlama 3.1 8BGPT-OSS 20BDLRMv4FLUX.1

StartupHub.ai StaffDaniel SingerWritten by StartupHub.ai Staff, Editorial teamReviewed by Daniel Singer, EditorAbout

StartupHub.ai Staff, Editorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer, Editor

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

More from Daniel Singer

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.