StorageReview.com

Veeam has launched the Veeam Intelligence MCP Server, an open-source server built on the Model Context Protocol that brings backup, recovery, malware, and compliance signals into broader enterprise IT workflows. Key details: - Read-only by default, with authenticated, authorized, and auditable queries - Deployed locally as a Docker container, fully customer or service provider controlled - No raw operational data sent to external services unless explicitly configured - Compatible with any MCP-compatible AI client, including local or self-hosted LLMs - Launches with support for Veeam Backup & Replication, Veeam ONE, and Veeam Service Provider Console Full details: https://lnkd.in/eaJ7BirC Veeam Software #DataProtection #MCP #Backup #EnterpriseIT #ModelContextProtocol

Our Dell Technologies PowerEdge R5715 review is live. Where the R4715 focuses on compute density in 1U, the R5715 is built around storage capacity and I/O expandability in a 2U single-socket footprint. We tested it with the EPYC 9015, the 8-core entry in the AMD EPYC 9005 family, paired with 384GB of DDR5 and a 12-bay 3.5-inch HDD configuration. The platform supports up to 288TB of raw storage in a single node, with four PCIe Gen5 slots, dual OCP 3.0 networking, and the same iDRAC10 management stack across the broader 17th Gen PowerEdge lineup. For file serving, backup targets, and video surveillance workloads where drive density matters more than core count, the R5715 makes a strong case. Both R4715 and R5715 reviews are now live. https://lnkd.in/ev6u24Zd #PowerEdge #AMEPYC #ServerReview #SMB #Storage #Infrastructure AMD Dell Technologies

We just published our review of the Dell Technologies PowerEdge R4715, the new 1U single-socket server built around 5th Gen AMD EPYC processors for SMB and midmarket environments. Our unit came configured with the EPYC 9335, the 32-core top-of-stack option for this platform, paired with 384GB of DDR5. Testing covered y-cruncher, Blender 4.5, and a full Phoronix suite including Apache, OpenSSL, kernel compile, 7-Zip, and Stream. The platform delivers strong compute density for virtualization, scale-out databases, and edge workloads, with iDRAC10, full PCIe Gen5 expansion, and enterprise security built in at a value-tier price point. The R4715 is one of two new AMD-based value platforms Dell just added to the 17th Gen PowerEdge family. Our companion R5715 review is also live. https://lnkd.in/ehvMw4Q4 #PowerEdge #AMEPYC #ServerReview #SMB #Infrastructure Dell Technologies AMD

For 25 years we've tested some of the most advanced server, networking, and storage infrastructure on the planet. We've never labeled any of it. That changes now, and it's amazing. Full Brady M511 portable label printer review coming soon. #Brady #DataCenter #Infrastructure #LabLife Brady Corporation

Your weekly news and reviews from StorageReview.

WEKA has announced general availability of NeuralMesh, its AI Data Platform built on the NVIDIA AI Data Platform reference architecture. The platform is aimed squarely at the problem that kept coming up all week at GTC: organizations that can run a successful AI pilot but stall when trying to get to production scale. NeuralMesh is built on over a decade of AI-native storage development and more than 170 patents. It is designed as a composable, exabyte-scale infrastructure that delivers AI-ready data to production environments, with consistent performance across early-stage workloads and full production deployments. WEKA positions it as a system that improves with scale, particularly in distributed AI environments where data access patterns and concurrency grow over time. Full details at the link. https://lnkd.in/eHpdFZ82 WEKA NVIDIA #storage #AI #datacenter #enterprise #AIinfrastructure

Not every piece of hardware needs to be practical to be interesting. The Dell UltraSharp U5226KW is a good example. At 52 inches and 6K, this isn’t about adding another display; it’s about replacing the entire concept of a multi-monitor setup. One screen, one cable, and enough room to keep everything in view at once. For most people, it’s hard to justify. But for workflows built around constant multitasking, large datasets, or timelines, the value becomes pretty clear. Sometimes the most compelling gear isn’t about what you need. It’s about what becomes possible when you remove the usual limits. Full review 👇 https://lnkd.in/eYGHJ3vJ #dell #monitor Dell Technologies

GTC 2026 is in the books. Time to break down the most interesting piece of the Vera Rubin platform. Divyansh Jain went deep on the NVIDIA Groq 3 LPX: how the architecture works, how it connects to Vera Rubin, what NVIDIA is actually offloading and why, and how many racks you'd need to run the biggest open-source models today. Full deep-dive is live on StorageReview. Here's the TL;DR: The Groq architecture is largely unchanged, and that's the point. Full determinism, compiler-scheduled data movement, MXM/VXM/SXM/MEM execution units, RealScale C2C inter-chip interconnect, dragonfly topology, cableless backplane: all intact. NVIDIA didn't redesign Groq. They updated it, scaled it up into a dense 256-LPU liquid-cooled MGX rack, and plugged it into Vera Rubin. Each LPX rack talks to a neighboring NVL72 via Spectrum-X Ethernet. NVIDIA Dynamo orchestrates the split: prefill and attention go to Rubin GPUs, FFN decode goes to the LPUs, on every single token. The activation tensors moving between GPU and LPU are tiny relative to the weight data, which is exactly the regime where Groq's near-zero-overhead networking shines. Why offload FFN specifically? In modern MoE models, FFN are more than two-thirds of all parameters, and are statically schedulable. Perfect fit for the LPU's deterministic SRAM-first architecture. Attention stays on the GPU because it needs the full KV cache, which grows dynamically with context and lives in HBM. This is also how NVIDIA solved Groq's biggest pre-acquisition limitation. Standalone Groq was expensive to scale for long-context models because KV cache would require racks of LPUs as context grew. By splitting the workload, attention on Rubin's 288 GB HBM4 per GPU and FFN on Groq's 150 TB/s SRAM per chip, NVIDIA eliminated that bottleneck. The LPU never touches the KV cache. It just does deterministic, ultra-fast matrix math on static weights. We also ran the sizing math. Each LPX rack has 128 GB of SRAM across 256 chips. GPT-OSS 120B (~53 GB FFN)? Fits in one rack. DeepSeek R1 (~623 GB FFN)? About 5 racks. GLM-5 in BF16 at 1.4 TB+? 10+ racks, or 5-6 at FP8. The front-panel C2C ports exist precisely to chain racks for models that exceed single-rack capacity. Shipping H2 2026 alongside Vera Rubin. Full breakdown with architecture diagrams, parameter calculations, and topology analysis at the link. https://lnkd.in/eqB5TMJ7 #nvidia #ai #gtc2026 NVIDIA

For the first time, end-to-end hardware-based in-flight encryption is available across the entire Fibre Channel data path, from server to storage array. Emulex SecureHBA technology is now integrated directly into Everpure FlashArray systems, with the FlashArray//XL130 R5 being the first primary storage array to natively support this capability. Future FlashArray models will ship with SecureHBA as the standard Fibre Channel adapter option. This is the most complete FC security architecture available today. If your organization operates Fibre Channel SANs in regulated industries or high-security environments, this is the solution the industry has been working toward. We evaluated it firsthand. Full review at the link. https://lnkd.in/ezKU2sHv Everpure Broadcom #storage #FibreChannel #SAN #security #enterprise #datacenter