loading
loading
Off the Wire Press Releases Sept. 2, 2026 — Everpure has announced that FlashBlade//EXA ranks number one across 405B parameter and 1.25T parameter model checkpointing and KV cache performance categories in the newly released MLPerf Storage v3.0 benchmark. The benchmark is designed to measure storage performance against demanding AI workloads, rather than testing storage in isolation. When a training run needs to checkpoint, or an inference workload needs to retrieve KV cache data, the storage system moves beyond an afterthought and becomes part of the equation. If storage can’t keep pace, GPUs and accelerators sitting idle start becoming costly. “These results yet again demonstrate that FlashBlade//EXA delivers the highest performance and scale on the market, staying ahead of the increasing demands of AI workloads,” said Rob Lee, Chief Technology and Growth Officer, Everpure. “Organizations advancing the bleeding edge of AI development need infrastructure that can deliver the utmost performance and scalability, without making tradeoffs against reliability and operational simplicity. FlashBlade//EXA stands alone in being able to deliver that.” Behind the Numbers The checkpointing results are compelling when put at scale. With 1,024 simulated accelerators, FlashBlade//EXA delivered 877.52 GiB/s write B/W (17.74s duration) and 588.28 GiB/s of Read B/W (28.99s duration) using 30 data nodes for a 1.25T parameter model. More importantly, performance scaled predictably from 327.65 GiB/s of write throughput with 10 data nodes to 877.52 GiB/s with 30 data nodes. This predictable scaling matters because AI workloads don’t stay static. As organizations add compute power, increase model sizes, bring in more data and expand capacity, the infrastructure underneath i.e. storage needs to keep pace. The same issue is emerging on the inference side. As organizations move from training increasingly large models to serving them at scale, KV cache is becoming an increasingly important part of the data path. In MLPerf testing, FlashBlade//EXA delivered 85,736 tokens/second for the Llama 3.1 8B workload using storage only, 67,642 tokens/second for an 8B workload using storage plus memory, and 33,403 tokens/second for a 70B storage-only workload. The numbers point to a broader shift in how AI infrastructure should be thought about. Storage is no longer simply the place where data lives between compute operations. It increasingly has to participate in the data path at the speed the AI workload demands. That is also what makes the architecture behind these results important. FlashBlade//EXA uses a disaggregated architecture that separates metadata management from data transport. In the benchmark configuration, 30 FlashBlade//EXA blades with 120 DirectFlash Modules dedicated to metadata worked alongside 30 Linux/NVMe data nodes, providing approximately 866 TB of usable capacity through a unified file system. The architecture uses NFSv4.1/pNFS over TCP for file layout negotiation and NFSv3 over RDMA (RoCEv2) for high-speed data movement. In other words, data can move across the underlying network fabric rather than forcing every workload through a centralized storage path. Why This Matters That distinction becomes more important as AI clusters grow. This isn’t about building a storage system that is fast for a particular configuration and then redesigning it when the AI environment changes. It’s about having an architecture where capacity and performance can scale alongside the compute environment without adding another layer of complexity for engineering teams. This ties back to Everpure’s journey of broadening its vision beyond the traditional storage array toward the Enterprise Data Cloud—a software-driven approach designed to help enterprises manage and use data across increasingly complex environments. AI is one of the clearest tests of that vision because the infrastructure behind it can’t really be treated as a collection of disconnected components anymore. GPUs, networking, storage and data all have to work together, and the performance of one increasingly depends on how well the others can keep up. And the ability to scale that data movement predictably may be just as important as the peak performance number itself. More from HPCwire: MLCommons Releases New MLPerf Storage v3.0 Benchmark Results About Everpure Everpure (NYSE: P) allows organizations to take control of their data with an industry-leading, ever-evolving storage and data management platform. We help companies unleash the power of their data by ensuring it is accessible, intelligent, and ready to perform in the AI era. We make data management effortless while simultaneously scaling performance and significantly reducing energy consumption. With one of the highest Net Promoter Scores for over a decade, Everpure is the choice of the world’s most innovative organizations. For more information, visit www.everpuredata.com. Source: Mike Adams, Everpure
document fetched 2026-09-03 02:58 · canonical https://hpcwire.com/off-the-wire/everpure-leads-mlperf-storage-tests-for-large-model-checkpointing-and-kv-cache