Your file-sharing layer is the bottleneck.
Here's the data.
We benchmarked Fusion SMB against Samba, NFSD, NFS Ganesha, and Windows Server on AI training, checkpointing, and high-throughput workloads. The report shows exactly where each one breaks.
See how much throughput you're losing
FIO results at 100GbE, 200GbE, and 2×200GbE. At 2×200GbE: Fusion SMB hits 46.86 GB/s. Samba flatlines at 4.17.
FIO results at 100GbE, 200GbE, and 2×200GbE. At 2×200GbE: Fusion SMB hits 46.86 GB/s. Samba flatlines at 4.17.
Find out if your storage can feed your GPUs
MLPerf 3D U-Net passing accelerator units across Samba, Ganesha, NFSD, and Fusion SMB. And a LLaMA3-8B checkpoint that takes 6.7 seconds on Fusion vs 28.1 on Samba.
MLPerf 3D U-Net passing accelerator units across Samba, Ganesha, NFSD, and Fusion SMB. And a LLaMA3-8B checkpoint that takes 6.7 seconds on Fusion vs 28.1 on Samba.
A real IBM Storage Scale deployment: four Fusion SMB nodes replaced eleven Samba nodes. Same 150 GB/s, 30–35% lower 3-year TCO.