Infortrend Launches High Throughput U.2 NVMe Storage and Dense JBOF

Infortrend’s latest U.2 NVMe announcements mark a clear push to move enterprise storage beyond incremental refreshes and into an era focused squarely on AI-scale throughput, NVMe-oF fabrics, and high-density U.2 economics—an aggressive strategy that pairs new flagship nodes with purpose-built all‑flash expansion to service media, HPC and model‑centric AI workflows.

Background / Overview​

Infortrend has refreshed its U.2 NVMe lineup with two headline-class developments: a new flagship unified storage array delivering markedly higher throughput and IOPS than prior generations, and a 2U all‑flash expansion (JBOF) that lets customers scale U.2 density without changing primary controllers. The vendor’s product pages and press releases describe the new storage node as a purpose-built platform for AI, HPC and media workloads, emphasizing high aggregate bandwidth, NVMe-oF and GPU‑direct integration for accelerated pipelines. The two product threads to track are:

• The EonStor GS 5024U (and related GS 50xxU family messaging) positioned as Infortrend’s top‑performing U.2 NVMe unified storage system, with claims of multi‑tens of gigabytes per second throughput, support for PCIe 5.0 SSDs, high IOPS counts and scale‑out options.
• The JB 4000U JBOF expansion enclosure: a compact 2U chassis that holds up to 24 U.2 NVMe SSDs and is positioned as a high‑density, lower‑cost way to add NVMe capacity and bandwidth to EonStor GS systems. Coverage and the vendor brief put its dual‑controller throughput at tens of gigabytes per second in suitable pairings.

Together these announcements signal Infortrend’s bid to provide an end‑to‑end U.2 NVMe strategy: high‑bandwidth controllers at the head of the stack and U.2 density shelves to scale capacity and parallelism.

---

Technical snapshot: what Infortrend is claiming​

EonStor GS 5024U — headline specs​

Infortrend’s release for its newest GS flagship emphasizes three core technical pillars:

• A modern host CPU platform drawn from Intel’s Xeon 6 family, intended to provide the CPU, PCIe lanes and accelerators needed for NVMe‑centric designs.
• Extreme aggregate throughput: the vendor claims up to 125 GB/s aggregate throughput (a 2.5× uplift over its prior top model) and 2.4 million IOPS, enabled by PCIe 5.0 U.2 SSDs and high‑speed fabric connectivity.
• High‑speed networking and fabric support: 200GbE class networking, NVMe‑over‑Fabric (NVMe‑oF) and GPUDirect Storage integration to accelerate GPU‑driven training and inference workflows. The system also claims compatibility with parallel file systems such as Lustre for HPC workflows.

JB 4000U — density and expansion​

Infortrend’s JB 4000U JBOF is described as a 2U, 24‑bay U.2 NVMe expansion enclosure with dual‑controller modes that can deliver ~24 GB/s in a dual‑controller configuration and up to 1.47 PB raw per chassis with appropriate high‑capacity U.2 SSDs. The vendor positions this enclosure as a way to increase U.2 density and aggregate NVMe parallelism for tasks like AI dataset staging and multi‑stream media editing.

Market context and component realities​

The market has been shifting toward larger‑capacity U.2 and E1.S modules (61.44 TB class SSDs, high‑density QLC and TLC enterprise SSDs are now available), which makes PB‑scale NVMe more attainable at rack scale. Vendors such as Solidigm, Western Digital and others publish enterprise parts at capacities in the 30–61.44 TB class that can populate 24‑bay U.2 enclosures for very high density. That capacity availability underpins Infortrend’s density claims for the JB 4000U and the GS family.

---

Why this matters: real workloads that benefit​

• AI training and preprocessing: Large models and massive datasets need rapid re‑reads, shuffling and checkpointing. The GS 5024U’s NVMe‑oF support, GPU‑direct claims and very high aggregate bandwidth are designed to reduce data movement bottlenecks between storage and GPU clusters.
• Media & Entertainment (M&E): Multi‑stream 4K/8K editing, color grading and multi‑camera ingest benefit from many parallel NVMe channels and dense SSD pools for high sustained write/read bandwidth; the JB 4000U offers a compact way to add many U.2 drives.
• HPC and parallel filesystems: Integration with Lustre and high‑speed RDMA networking positions the EonStor GS as a candidate for cluster storage frontends where aggregate throughput and low latency at scale matter.

---

Strengths — where Infortrend’s approach is compelling​

• End‑to‑end NVMe focus. Infortrend is doubling down on U.2 NVMe as the primary medium for both compute‑attached arrays and expansion enclosures. That design coherence simplifies procurement and lifecycle support compared with mixed SAS/NVMe stacks.
• High aggregate bandwidth claims. The 125 GB/s headline for the GS flagship represents a meaningful step for a unified array in this class—if sustained under real workloads, it materially narrows the storage‑side bottleneck for multi‑GPU nodes. This is especially relevant for customers who need model parallelism or large data staging without staging to remote object stores.
• Density + economics via U.2. Large U.2 QLC/TLC drives at 30–61.44 TB make PB‑scale NVMe deployments more financially plausible than they were just a few years ago. A 24‑bay U.2 JBOF filled with 61.44 TB drives gives the kind of density that previously required EDSFF or many 2.5" devices.
• NVMe‑oF and GPUDirect integration. Supporting NVMe‑oF and vendor ecosystems that allow GPUDirect I/O reduces unnecessary CPU copies, enabling faster host‑to‑GPU flows when software stacks are validated. For GPU‑heavy AI centers, this is a practical performance lever rather than a theoretical one.

---

Risks, unknowns and practical caveats​

No product announcement exists in a vacuum: several practical and technical caveats temper the headline claims.

• Vendor claims vs. sustained real‑world performance. Peak aggregate numbers (GB/s or IOPS) are useful marketing anchors but rarely translate into sustained throughput for complex, mixed workloads. Customers must validate sustained write behavior, p99/p999 latency under load, and thermal/thermal‑throttling curves on target host platforms and network fabrics. Treat peak numbers as directional until verified in your environment.
• NVMe‑oF and fabric complexity. NVMe‑oF gives tremendous benefits but brings design complexity: RDMA fabric planning, switch and HBA compatibility, congestion control, multipathing, and fault‑domain design are non‑trivial for many teams. Achieving the low latencies and aggregated throughput advertised requires a correctly designed network fabric and operational maturity.
• SSD endurance and SLAs. The economics of high‑capacity QLC vs TLC or enterprise TLC requires tradeoffs. QLC parts lower $/TB but come with weaker endurance (DWPD) and different thermal characteristics. For AI workloads that produce heavy write amplification (checkpoints, shuffles), choosing the wrong media mix can shorten drive life or trigger rebuild storms. Verify endurance ratings, warranty conditions, and how Infortrend’s algorithms manage SSD wear in mixed workloads.
• Thermals and power at rack scale. High aggregate NVMe density plus high IOPS drive power consumption — particularly with Gen5 devices — creates thermal and power provisioning challenges. Validate rack cooling, redundant power, and sustained power budgets; otherwise, you risk throttling or unplanned workload slowdowns.
• Firmware maturity and ecosystem support. New host controllers, Gen5 SSDs, and NVMe‑oF stacks often require firmware tuning across vendors. Early adopters should budget for firmware updates and interoperability testing. Where possible, require vendor‑validated lists for GPU, NIC, and HBA combos if you intend to use GPUDirect or do high‑performance NVMe‑oF.
• Meaningful verification of “GPUDirect Storage” claims. GPUDirect requires aligned driver/firmware stacks and validated host/GPU/NIC combinations; claiming GPUDirect support doesn’t remove the need for practical validation on your chosen GPU server platform. Ask vendors for validated configurations and test logs.

---

Deployment checklist — how to validate Infortrend’s claims for your site​

• Define workload profiles: collect realistic IO patterns (sequential vs random, read/write mix, QD distribution, checkpoint frequency).
• Request vendor‑supplied sustained performance graphs: long‑duration sustained write and mixed IO tests that mirror your workloads. If not available, insist on a loaner system for in‑house testing.
• Verify NVMe‑oF fabric capability: confirm switch, NIC (RDMA offload), and HBA compatibility. Design for congestion control and multipathing with clear QoS if needed.
• Validate GPUDirect: run a validated GPUDirect Storage test on the exact GPU, NIC and driver stack you plan to deploy. Don’t rely on generic statements of compatibility.
• Choose SSD media for the workload: prefer TLC / enterprise‑class parts for write‑heavy workloads; use high‑capacity QLC for read‑dominant archival or nearline tiers. Confirm endurance (DWPD and PBW) and warranty terms.

---

Practical configuration patterns and recommendations​

• For large‑scale model training clusters where GPUs are the bottleneck, design a tiered approach:
• Hot tier: GS 5024U frontends with high‑endurance TLC NVMe for model staging and checkpointing.
• Warm tier: JB 4000U shelves populated with high‑capacity QLC U.2 drives for dataset storage and archival staging.
• Cold tier: SAS HDD or QLC nearline shelves for long‑term archives.
  This preserves high endurance for write‑heavy operations while using dense QLC for capacity.
• For media post‑production workflows:
• Use JB 4000U racks as ingest pools for multi‑camera capture and editing teams; pair a GS frontend with redundant 100/200 GbE uplinks and NVMe‑oF to provide shared, low‑latency access.
• For HPC parallel filesystem frontends:
• Validate Lustre or similar integrations with Infortrend and test with representative IO patterns (IOZone, FIO mixed tests) at scale to ensure metadata and small‑IO patterns don’t become bottlenecks despite raw bandwidth.

---

Cost, TCO and operational considerations​

• Upfront hardware cost is only part of the equation. PB‑class NVMe adds recurring operational costs: higher power, denser cooling, drive replacement cycles (especially with QLC), and potential licensing/support for NVMe‑oF and GPUDirect software stacks.
• The JB 4000U is positioned as a cost‑effective density play versus SAS SSD enclosures; however, true TCO depends on drive selection (QLC vs TLC), expected write cycles, and the cost of overprovisioning to meet endurance targets. Vendors often show $/GB advantages for high‑capacity QLC — verify those numbers for your workload profile.
• Operational maturity with NVMe‑oF is a differentiator: teams with experience will extract full value; teams without RDMA/GPUDirect experience should budget for architecture and staff ramp.

---

Vendor claims to verify before buying (a quick checklist)​

• Confirm the exact CPU SKU and verify it provides the necessary PCIe lanes and platform features (CXL, channels) you require. Vendors sometimes list generational family names that hide actual SKU differences—ask for the SKU and microcode/firmware version.
• Get sustained performance graphs (not just peak numbers) across the mix of drives you plan to buy. Long‑duration tests are crucial for understanding thermal throttling and write amplification.
• Validate NPIV/HBA and NIC compatibility for NVMe‑oF and GPUDirect Storage with your exact server and GPU models. If you rely on vendor‑tested reference architectures, get them in writing.
• Confirm SSD models and their usable capacity (formatted vs raw) and endurance (DWPD). For advertised capacities like 61.44 TB, ask for the manufacturer part numbers and endurance data sheets.
• Ask for power and cooling delta measurements per fully populated JB 4000U and per GS chassis at typical operational IO. These numbers determine rack design and utility costs.

---

Broader industry perspective​

The industry trend toward larger U.2 modules and Gen‑5 capable platforms (and rebranded server CPU families such as Intel’s Xeon 6) is enabling vendors to advertise both higher raw throughput and denser per‑rack capacity. This shift is visible across multiple vendors and SSD suppliers and explains why vendors are announcing integrated NVMe‑oF + GPU‑direct solutions now—systems and networks finally have the parts to deliver meaningful throughput for AI and M&E pipelines. However, this also raises the bar for systems integration and infrastructure maturity in data centers; achieving vendor‑advertised numbers requires end‑to‑end validation.

---

Conclusion​

Infortrend’s push—pairing a high‑bandwidth flagship array with a dense U.2 JBOF—reflects sensible engineering choices for organizations that need both throughput and density for AI, M&E and HPC workloads. The technical building blocks (Xeon 6 host CPUs, PCIe 5.0 SSDs, NVMe‑oF, GPUDirect and 200GbE fabrics) are now mainstream enough to deliver the promised benefits, provided careful integration and validation.
Buyers should treat the headline numbers as the start of engineering conversations rather than as guaranteed outcomes. The real value will come from measured, sustained performance in your specific environment: test the exact drive SKUs you will run, validate NVMe‑oF and GPUDirect on your GPU servers, plan for thermal and power delta, and choose media (TLC vs QLC) aligned to endurance needs. When those steps are followed, Infortrend’s U.2‑centric approach can deliver a highly capable platform that meaningfully shortens data paths for AI and media pipelines while giving IT teams a pragmatic density option in the JB 4000U.

---

Source: TechPowerUp Infortrend Unveils Its Most Advanced U.2 NVMe SSD Storage Solution | TechPowerUp}