Supermicro Ships Petascale All-Flash Server With PCIe Gen 6
Supermicro's first PCIe Gen 6 storage system packs 3.93 PB of flash into 2U and claims up to double the sequential-read bandwidth of Gen 5, validated with eight software partners.

Why it matters
- Supermicro's new Petascale server holds up to 3.93 PB of flash in a 2U chassis and ships now, announced November 18, 2026.
- PCIe Gen 6 delivers up to 100% more sequential-read bandwidth than PCIe Gen 5, per Supermicro.
- The system uses sixth-generation AMD EPYC 9006 processors on the SP7 platform; SP8-based storage products are planned.
- The chassis supports up to 32 E3.S NVMe SSDs, reaching maximum capacity with 122 TB drives, plus up to four x16 networking cards.
- Validation covers CloudDC, GrandTwin and Hyper families and software from Hammerspace, IBM, MinIO, Nutanix, Scality, VAST Data, VDURA and WEKA.
Supermicro's new Petascale all-flash server packs 3.93 petabytes of flash storage into a 2U chassis, and the first systems are shipping now. The company announced the launch on November 18, 2026, confirming that this is its first storage product line to implement PCIe Gen 6, the latest generation of the peripheral interconnect that links SSDs, GPUs and CPUs inside a server.
The headline performance claim is concrete: Supermicro reports a sequential-read bandwidth gain of up to 100% over PCIe Gen 5 systems. That number deserves the usual scrutiny — "up to" claims describe best-case sequential reads, not the mixed random I/O that dominates real AI data pipelines. Still, a doubling of the storage interconnect's ceiling in a single generation is a meaningful step for workloads that saturate existing links.
The stakes are straightforward. AI training and inference clusters have shifted the bottleneck from compute to data movement: GPUs idle when storage cannot feed them fast enough. Supermicro is positioning this hardware squarely at that gap, targeting high-performance computing, AI training and inference deployments.
What is inside the Petascale server?
The initial Petascale server runs sixth-generation AMD EPYC 9006 processors on AMD's SP7 platform. Supermicro says more storage products will follow as AMD introduces SP8 CPU versions, which signals a roadmap tied to AMD's socket cadence rather than a one-off box.
The chassis accepts as many as 32 E3.S NVMe SSDs — E3.S being the enterprise flash form factor that replaced the older U.2 drive shape. Maximum capacity arrives with 122 TB drives, which is what produces the 3.93 PB figure in two rack units. On the networking side, the system supports up to four x16 high-performance cards, and Supermicro describes the I/O design as balanced so customers can adapt the system to different fabrics and workloads.
That networking detail is not incidental. Supermicro explicitly frames the design around the link between drive density and data movement: capacity alone does not address how quickly data can transfer between storage and the rest of an AI cluster. The hardware targets the sustained bandwidth required for AI data lakes, large-scale analytics and high-throughput storage networking.
The takeaway: this is a density-plus-bandwidth play, not a capacity play alone.
Who said what?
Chief business officer Vik Malyala framed the systems as a way to keep storage in step with current CPU and GPU clusters. Supermicro's Data Center Building Block Solutions architecture, he indicated, allows customization for those workloads and provides software-defined-storage partners with validated systems. Software-defined storage refers to storage management handled by software layers rather than fixed hardware controllers.
Vincent Hsu, IBM Fellow, vice president and CTO of IBM Storage, offered the most detailed partner statement. He said AI inference spans data ingest, model loading, retrieval and KV cache — the KV cache being the memory of previously computed attention results that large language models reuse during generation. IBM Storage Scale on the new Petascale server, he said, is designed to support the entire pipeline while maximizing GPU-utilization efficiency, with headroom for performance and capacity growth.
Hsu's framing is notable because it treats inference, not just training, as the storage-intensive phase. That aligns with the industry shift toward retrieval-heavy, long-context inference workloads.
Why does the software validation layer matter?
Supermicro validated the hardware across its CloudDC, GrandTwin and Hyper server families, and validated software from eight vendors: Hammerspace, IBM, MinIO, Nutanix, Scality, VAST Data, VDURA and WEKA. The breadth of that list is the strategic point. The partner statements position the hardware across several competing storage architectures rather than as a single fixed stack.
The partners describe genuinely different data-management approaches:
- Hammerspace: its Data Platform can fold the Petascale flash into one global namespace — a single unified file view — alongside a customer's existing lower-cost capacity systems.
- WEKA: its distributed NeuralMesh software runs across the new server without hardware tiers.
- VDURA: its platform can present Supermicro's all-flash and HDD systems within one namespace.
- IBM: Storage Scale supports the full inference pipeline from ingest through KV cache.
The pattern is clear. Supermicro sells the silicon and the enclosure; the software layer determines how that capacity gets organized, tiered and delivered to each stage of an AI workload. By validating eight vendors rather than betting on one, Supermicro hedges its position in a storage software market where no single architecture has won AI deployments.
The caveat: validation means the combinations are tested, not that every software stack extracts the full 100% bandwidth gain. Real-world throughput will depend on which partner stack a customer chooses and how its data is structured.
What comes next?
Two forward threads emerge from the announcement itself. First, the SP8-based storage products will follow AMD's next CPU platform, extending the family beyond the initial SP7 system. Second, the burden of proof shifts to the validated software partners: the hardware adds denser flash and higher-bandwidth I/O, but as the announcement itself concedes, the software layer determines how that capacity is organized and delivered to each stage of an AI workload.
For buyers, the practical question is no longer whether petascale flash fits in 2U — Supermicro says it now does — but whether their storage software can keep an AI cluster's GPUs fed at the new interconnect's ceiling.
By James Calloway
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
722 articles
Related articles
- Supermicro Ships 20-Plus H15 Servers Built on AMD EPYC 9006 Chips
- Nexalus and RENDA Launch Nova8, a Liquid-Cooled AI Server Built for Retrofit
- Lenovo Bets on Liquid Cooling and Services to Solve AI's Power Problem
- MiTAC Splits AI Inference Across Two Servers With 2 TB Decode Memory
- GIGABYTE AI TOP AERO fits 128GB in a roughly one-litre desktop