Supermicro ships PCIe Gen 6 storage servers
·

Supermicro ships PCIe Gen 6 storage servers

Supermicro has begun shipping all-flash storage servers using AMD EPYC 9006 processors and PCIe Gen 6 connectivity, offering up to 3.93PB of raw storage capacity in a 2U chassis.

Supermicro ships PCIe Gen 6 storage servers
Summary
  • Supermicro is shipping its first PCIe Gen 6 Petascale all-flash servers, powered by sixth-generation AMD EPYC 9006 processors.
  • The initial 2U system accommodates up to 32 E3.S NVMe drives, with a stated maximum raw capacity of 3.93PB and four high-performance network adapter slots.
  • The company reports up to twice the sequential read bandwidth of PCIe Gen 5 systems and has validated configurations with multiple software-defined storage partners.

Supermicro has begun shipping all-flash storage servers supporting PCIe Gen 6, introducing higher-bandwidth storage infrastructure for AI training, inference and high-performance computing.

The first product in the new Petascale family uses sixth-generation AMD EPYC 9006 Series processors, formerly developed under the Venice codename. The servers combine the newer processor architecture with PCIe Gen 6 connectivity and support for high-density NVMe solid-state drives.

Supermicro announced the systems on 18 November during Supercomputing 2026 in Chicago. It claims up to 100% greater sequential read bandwidth than comparable PCIe Gen 5 configurations, although the performance delivered by individual deployments will depend on their storage devices, networking and software.

The initial server occupies 2U of rack space and supports up to 32 E3.S NVMe drives. With 122TB-class drives installed, Supermicro specifies a maximum raw storage capacity of approximately 3.93 petabytes per chassis.

That figure represents the maximum stated physical drive capacity. Usable storage may be lower after accounting for data protection, formatting, metadata and the configuration of the installed storage software.

The server also supports up to four high-performance network adapters using x16 connections. These interfaces provide options for connecting the storage system to the wider computing environment, where bandwidth and latency can influence workload performance.

PCIe Gen 6 increases internal data movement

PCI Express connects processors with storage devices and other components inside a server. PCIe Gen 6 doubles the nominal signalling rate per lane compared with the previous generation, allowing compatible hardware to exchange more information over the available connections.

The specification increases the rate to 64 gigatransfers per second per lane. It also introduces changes to signalling and error correction required to support communication at the higher speed.

Greater interface capacity can reduce a bottleneck within storage servers, particularly where many fast drives compete for access to processors and network adapters.

However, the maximum signalling rate does not equal application throughput. Protocol overhead, device performance, processor resources and software behaviour all influence the amount of data that can be transferred during actual operation.

The four network adapter positions allow system designers to connect the server to high-capacity storage fabrics. Whether the additional bandwidth can be used depends on the adapters installed and the available external network infrastructure.

A server containing large numbers of NVMe drives may have considerable aggregate storage performance, but that capability cannot be delivered to other computing systems if its network connections are insufficient.

Conversely, a high-capacity external network offers limited benefit when the storage devices, controllers or software cannot supply information quickly enough.

AI computing environments are particularly sensitive to these relationships because expensive accelerator resources may remain idle while waiting for data to arrive.

Storage software determines how capacity is used

Supermicro has validated configurations involving several software-defined storage suppliers, including Hammerspace, IBM, MinIO, Nutanix, Scality, VAST Data, VDURA and WEKA.

These products provide different storage architectures and functions. Some concentrate on high-performance file systems, others on object storage, distributed data services or the management of storage across computing environments.

The inclusion of several suppliers gives customers options for configuring hardware around their applications. It does not establish identical performance, certification or availability across every possible hardware and software combination.

Software-defined storage determines how information is organised, distributed, protected and made available to applications. It can distribute data across multiple servers, replicate information for resilience or manage access through file and object interfaces.

The chosen architecture affects how efficiently the available storage hardware can be used. A system optimised for large sequential transfers may behave differently from one processing many smaller requests.

AI training systems frequently need to retrieve large datasets and write checkpoints that allow processing to resume after interruptions. Checkpointing can generate concentrated write activity, while training data access may require high sustained read throughput.

Inference environments create different demands, including model loading and access to supporting information. The performance requirements depend on the models, serving architecture and number of simultaneous requests.

The new servers are designed to support such workloads, but their suitability will depend on the storage software, network configuration and performance testing undertaken for the intended application.

Density introduces power and maintenance considerations

Installing up to 32 high-capacity NVMe drives within 2U concentrates storage capacity into a relatively small rack footprint. This can reduce the amount of physical space required for a specified raw capacity, although it also concentrates the associated electronics and thermal load.

Servers require adequate electrical supply and cooling airflow, with suitable access for replacing drives and maintaining supporting components.

Power consumption cannot be established from storage density alone. It depends on the installed drives, processor configuration, networking equipment and operating load.

Supermicro has not published an independently verified whole-facility energy saving associated with the new servers. A reduction in rack space may affect supporting infrastructure requirements, but the resulting benefit depends on the equipment replaced and the design of the data centre.

The company expects to extend the product family with systems using additional AMD EPYC 9006 processor configurations, including SP8 variants. Those products remain future introductions rather than part of the initial shipping configuration.

The confirmed milestone is the commercial shipment of a PCIe Gen 6 storage platform offering higher internal bandwidth and substantial flash capacity within a compact chassis. Operators will need to assess application performance, software compatibility, network capacity and operating requirements before determining its value within existing infrastructure.


Stay updated with the latest insights and trends in the data centre industry by subscribing to our newsletter.

← Back

Thank you for your response. ✨