Quobyte has delivered the fastest unverified results in the 3D U-Net benchmark, significantly outperforming all competitors in one of the most demanding AI workloads in the industry.
By delivering the fastest results in this benchmark, Quobyte categorically shows that it’s built for real-world, production-scale AI, and does so with the best efficiency.
AI and some high-performance computing (HPC) depend on GPUs for critical processing tasks. GPUs are very expensive resources. Cost estimates for H100 GPUs are in excess of USD $90 per hour. Maximizing their utilization is not only critical but an absolute imperative to obtain timely results and the best ROI. If GPUs are not utilized at peak efficiency results suffer in the form of longer processing times, overprovisioning of GPUs in an attempt to compensate, higher capital costs, and increased power consumption, all of which result in a huge waste of resources. Storage cannot be overlooked, because slow performing storage systems waste expensive GPU cycles.
3D U-Net is the most IO intensive of the MLPerf® storage benchmarks. With high I/O requirements and sensitivity to latency, this benchmark separates true high-performance storage solutions from the rest. Quobyte outdistances all other storage solutions in the market, and does so with unparalleled efficiency, and total reliability.
A Quobyte mid-tier all-flash 4 node cluster was able to drive 6 H100 GPUs per client at 90.29 % utilization, over a 200 Gb/s RoCE network. This showcases the tremendous performance advantage of Quobyte over alternative solutions. This is illustrated in the graph above.
Why Quobyte for AI Workloads?
Quobyte is 33% faster than the closest competitive alternative and orders of magnitude better than the remaining ones. It does so almost at line speed, with efficient power consumption and using the least amount of hardware. Quobyte also has the best availability with no maintenance windows required for any reason including failed components, upgrades, or software updates. This total availability also increases GPU utilization, which in other solutions are kept idle for maintenance operations. Quobyte scales to 1000s of nodes, maximizing ROI, makes your users happy and keeps your operations productive and optimal.
Ready to Accelerate Your AI Workloads?
If you’re building or scaling deep learning infrastructure, now is the time to rethink your storage strategy. With Quobyte, you get:
- Performance that keeps up with your GPUs
- A true software-defined solution that runs anywhere – on x86, ARM, Nvidia Grace, on prem, GPU hosters and the public clouds
- Complete enterprise feature including snapshots, replication, erasure coding, and multi-tenancy
- Unparalleled availability with no maintenance windows for any reason
Want to learn more or see Quobyte in action? Book a demo or schedule an appointment to talk with our experts
