Everpure는 FlashBlade//EXA 스토리지 아키텍처가 405B 및 1.25T 파라미터 모델 체크포인트 및 KV 캐시 성능에서 MLPerf 벤치마크 1위를 달성했다고 발표했습니다.
On September 2, 2026, Everpure announced that FlashBlade//EXA ranked first in both the 405B and 1.25T parameter model checkpointing and KV cache performance categories in the newly released MLPerf Storage v3.0 benchmark. Designed to evaluate storage performance under demanding AI workloads rather than in isolation, the benchmark underscores how infrastructure directly impacts compute efficiency.
When a training job requires a checkpoint or an inference workload retrieves KV cache data, storage shifts from an afterthought to a critical component. If storage cannot keep pace, idle GPUs and accelerators translate directly into wasted capital and lost throughput.
“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.”
The checkpointing results are particularly compelling at scale. Using 30 data nodes to handle a 1.25T parameter model across 1,024 simulated accelerators, FlashBlade//EXA achieved 877.52 GiB/s write bandwidth over a 17.74-second duration and 588.28 GiB/s read bandwidth over 28.99 seconds. Performance scaled predictably from 327.65 GiB/s of write throughput with 10 data nodes to 877.52 GiB/s with 30 nodes. This linear scaling is critical because AI workloads are inherently dynamic; as organizations expand compute power, increase model sizes, ingest more data, and grow capacity, their underlying storage must scale accordingly.
A similar shift is occurring on the inference side. As organizations transition from training increasingly large models to serving them at scale, KV cache management has become a vital part of the data path. In MLPerf testing, FlashBlade//EXA delivered 85,736 tokens per second for a Llama 3.1 8B workload using storage alone, 67,642 tokens per second for an 8B workload combining storage and memory, and 33,403 tokens per second for a 70B storage-only workload. These figures underscore a broader industry shift: storage is no longer merely a passive repository between compute operations. It must actively participate in the data path at speeds dictated by AI workloads.
This performance is enabled by FlashBlade//EXA’s disaggregated architecture, which separates metadata management from data transport. In the benchmark configuration, 30 FlashBlade//EXA blades equipped with 120 DirectFlash Modules dedicated to metadata operated alongside 30 Linux/NVMe data nodes, delivering approximately 866 TB of usable capacity through a unified file system. The system leverages NFSv4.1/pNFS over TCP for file layout negotiation and NFSv3 over RDMA (RoCEv2) for high-speed data movement. Consequently, data traverses the underlying network fabric directly, eliminating bottlenecks associated with centralized storage paths.
This architectural distinction grows increasingly important as AI clusters expand. The objective is not to build a storage system optimized for a single configuration that requires redesign when the AI environment evolves. Instead, it is to deploy an architecture where capacity and performance scale seamlessly alongside compute resources, without imposing additional complexity on engineering teams.
This approach aligns with Everpure’s strategic evolution beyond traditional storage arrays toward the Enterprise Data Cloud—a software-driven framework designed to help enterprises manage and leverage data across increasingly complex environments. AI serves as one of the most rigorous tests of this vision, as modern infrastructure can no longer be treated as a collection of siloed components. GPUs, networking, storage, and data must operate cohesively, with the performance of each increasingly dependent on the others. In this context, the ability to scale data movement predictably may prove just as critical as peak performance metrics.
Everpure (NYSE: P) empowers organizations to take control of their data through an industry-leading, continuously evolving storage and data management platform. The company helps enterprises unlock the full potential of their data by ensuring it remains accessible, intelligent, and optimized for performance in the AI era. Everpure streamlines data management while simultaneously scaling performance and significantly reducing energy consumption. Recognized with one of the highest Net Promoter Scores for over a decade, Everpure is the preferred partner for the world’s most innovative organizations. For more information, visit www.everpuredata.com.
Source: Mike Adams, Everpure