IN Brief:
- Phison has introduced its X3 PS5303 enterprise SSD controller, targeting PCIe Gen6 x4 and NVMe 2.3.
- The controller's development targets include 28GB/s sequential throughput and 6.8 million random IOPS.
- Complementary computational storage and mixed-media architectures aim to reduce data movement and improve flash utilisation.
Phison Electronics has outlined three storage technology developments at Supercomputing 2026, including an enterprise SSD controller targeting sequential throughput of 28GB/s and random performance of 6.8 million input/output operations per second. The programme also includes computational storage hardware and a mixed-media architecture intended to address different constraints on data movement, processing, and flash capacity.
The X3 PS5303 enterprise controller supports PCIe Gen6 x4 and NVMe 2.3, extending Phison’s storage platform development towards a higher-bandwidth interface. Its stated throughput and IOPS figures are development targets rather than independently verified results achieved by commercially available production drives.
PCIe Gen6 doubles the transfer rate of each lane to 64 gigatransfers per second, providing additional bandwidth between host processors and connected devices. The increase creates opportunities for faster storage systems, although the performance achieved by a complete SSD also depends on its controller, NAND flash configuration, firmware, and available host resources.
The new interface uses pulse-amplitude modulation with four signalling levels, known as PAM4, to increase the amount of information transmitted through an electrical connection. Compared with conventional two-level signalling, the smaller separation between signal levels places additional demands on the quality of the transmission path.
Losses and interference along PCB traces, connectors, and other electrical interconnects can affect signal integrity. As transfer rates increase, system designers must pay greater attention to channel characteristics, impedance control, and the components used to maintain reliable communication.
Phison is developing PCIe Gen6 SSD controllers alongside redrivers and retimers, addressing signal transmission as part of the wider storage interface architecture. These devices perform different functions when compensating for electrical losses or regenerating high-speed data signals.
A redriver conditions a signal to compensate for attenuation and other channel effects, while a retimer recovers incoming data and transmits it using regenerated timing. Their use depends on the electrical configuration and requirements of the host system.
Phison identifies its PS7161 as the first PCIe 6.0 redriver included on the PCI-SIG Integrators List. The listing concerns the redriver component and should not be presented as evidence that the X3 controller or every future PCIe Gen6 SSD has completed system qualification.
The second development addresses the volume of data transferred between storage and processing hardware. Traditional computing arrangements frequently move datasets from SSDs into system memory before CPUs or accelerators examine the information.
Such transfers can consume substantial bandwidth when an application searches large amounts of stored data but ultimately requires only a relatively small result. Phison’s computational storage approach places selected processing capabilities within the storage device, allowing suitable operations to be performed closer to the information being held.
The architecture combines local computing resources with 64GB of memory and a stated memory bandwidth of 130GB/s. The company also identifies an integrated processing unit rated at 36,800 DMIPS, providing a measure of the processing capability incorporated into the design.
Phison says the approach could reduce data transfers by between ten and 100 times for applicable workloads, since the storage device can return selected results rather than transferring complete datasets to the host. These figures are manufacturer projections and depend on the processing task, data structure, and amount of information requested.
Computational storage does not remove the need for CPUs or GPUs. Operations requiring substantial general-purpose computing resources or large-scale accelerator processing may still be better executed elsewhere in the system. The proposed advantage concerns tasks that can be handled efficiently near the stored data.
Phison is also working with software frameworks including PyTorch, OpenVINO, vLLM, and llama.cpp. Software compatibility will influence the ability of developers to incorporate computational storage into established workflows without extensive changes to their applications.
The third development, Mixed Media, combines pseudo-single-level cell and quad-level cell flash within one SSD architecture. The design uses a sixteen-channel controller to manage the two flash types behind a common NVMe interface.
Pseudo-SLC operation stores less information per memory cell but can provide advantages in write performance and endurance. QLC flash stores more information per cell, supporting higher capacity at the expense of different performance and endurance characteristics.
Combining the technologies allows the controller to use each type for workloads suited to its characteristics. This could reduce the need to allocate separate SSDs exclusively to cache and capacity functions, depending on the requirements of the storage system.
The architecture also changes how high-performance cache resources can be distributed across a larger storage installation. Its effectiveness will depend on workload behaviour, firmware management, and the measures used to maintain data protection and availability.
Phison has not confirmed commercial shipment dates or independent performance results for all three developments. The X3 throughput figures, projected computational storage benefits, and mixed-media performance advantages require validation in completed systems.
The SC26 programme addresses three related storage constraints: host interface bandwidth, movement of information between storage and processing resources, and the allocation of different flash technologies. Their practical performance will depend on completed hardware development, system integration, and workload-specific testing.


