AI Infrastructure · Unified Storage
Your GPUs Shouldn’t Wait for Data
Most AI teams copy training data onto every GPU server’s local drive, because shared storage is assumed to be too slow. We tested that assumption in the TrndX Specialized Data Center (SDC), on NVIDIA H100s and on the industry-standard MLPerf benchmarks. The result changed how we size storage for training.
At a glance
Parity
A VM reading from shared storage kept pace with a VM reading from its own local NVMe drive, within a few percent.
On real H100 GPUs, a VM training from shared storage passed MLPerf Training at 95–96% of the per-GPU speed of bare metal with local NVMe.
Spreading 64 H100 GPUs across four VMs reading one shared dataset kept GPU utilization at 95%.
Preview
Read throughput, 32 simulated H100s (MLPerf Storage)
Note: Results are for the ResNet-50 model on the MLPerf Storage benchmark, in samples/sec.
Training speed per GPU, real H100s (MLPerf Training)
Full charts, test setup and scaling results are in the report.
In the full report
- Side-by-side results: shared storage vs local NVMe, VM vs bare metal
- How performance held up as we scaled out
- How TrndX Unified Storage combines parallel, block and S3 storage on one network
- How to size storage for your GPU count
Enter your details and we’ll send you the full report.
Early results from MLPerf benchmarks run by TrndX in its Specialized Data Center, not verified by MLCommons.