On November 17, 2026, at Supercomputing 2026 (SC26) in Chicago, Phison announced three new SSD technologies for data centers. They are the X3, a PCIe 6.0 controller targeting speeds of up to 28 GB/s; computational storage, which processes data inside the SSD and returns only the needed results to the server; and Mixed Media, which distributes the write cache across each SSD. In AI and high-performance computing (HPC), storage access can become a bottleneck for overall processing, and each of the three technologies addresses that problem in a different way.

AD

X3 supports PCIe 6.0, targeting up to 28 GB/s

The enterprise SSD controller X3 (PS5303) supports a four-lane (x4) PCIe 6.0 connection and NVMe 2.3. Phison's performance targets are up to 28 GB/s in sequential transfer speed and up to 6.8 million IOPS in random access. Sequential transfer speed indicates how many bytes of contiguous data can be transferred per second, while IOPS indicates how many input/output requests can be handled per second. Even at the same IOPS, actual transfer speed differs depending on how much data each I/O operation handles.

In the PCIe 6.0 specification published by PCI-SIG, the per-lane transfer rate doubles from PCIe 5.0's 32 GT/s to 64 GT/s. PCIe 6.0 adopts PAM4, which represents data using four signal levels, and also introduces forward error correction (FEC) and cyclic redundancy check (CRC) to handle signal errors. This expands the communication bandwidth available between the SSD and the host.

Working samples of the X3 were already shown in the COMPUTEX 2026 announcement. In that demo, four SSDs equipped with the X3 were run together with the PS7261, a retimer that adjusts and regenerates signal timing.

At SC26, in addition to the X3 performance targets, Phison announced the PS7161, a redriver that conditions signal waveforms. According to Phison, the PS7161 is the first PCIe 6.0 redriver listed on the PCI-SIG integrators list. The company develops not only SSD controllers but also the peripheral components needed to maintain the signal quality that high-speed communication requires.

An NPU and 64GB of memory inside the SSD, cutting data transfer by up to 100x

The second technology is computational storage, which gives the SSD itself processing capability.

Normally, when a server wants to extract needed information from a large amount of data on an SSD, it loads the data into host memory and performs searching and extraction on the CPU or GPU. For example, even when you want only the rows that match certain conditions in a database, depending on the processing method, a large volume of raw data may have to be transferred from the SSD.

Phison's computational storage scans and processes data in parallel inside the SSD and returns only the necessary results to the host. When the processed results are much smaller than the original data, the amount of data moving between the SSD and the host can be reduced. The company estimates that, for suitable workloads, transfer volume could be cut to between one-tenth and one-hundredth of the original.

The announced design puts an NPU for AI processing and 64GB of memory inside the SSD. The internal memory bandwidth is stated as 130 GB/s, but this figure refers to the SSD's internal memory bandwidth, not the transfer speed between the SSD and the host. Phison also plans support for AI software such as PyTorch, vLLM, OpenVINO, and llama.cpp.

The idea of processing data inside an SSD is not new. In 2022, Samsung announced its second-generation SmartSSD. That product used AMD's Xilinx Versal and Arm cores to run tasks such as database scanning and video transcoding on the SSD side. It was also designed to let customers combine their own software and intellectual property (IP).

By citing support for existing AI software such as PyTorch and vLLM, Phison may allow developers to use the SSD's compute functions from environments they already know. However, it has not been disclosed which specific workloads can be moved to the SSD side or how host-side software would invoke them.

Phison also cites a performance figure of 36,800 DMIPS, but this is a metric derived from Dhrystone, an integer-arithmetic benchmark, and does not directly represent the NPU's AI compute performance. Tom's Hardware points out that the number may indicate the performance of an auxiliary CPU paired with the NPU.

TechSpot also reported that while the 64GB memory capacity and 130 GB/s bandwidth were disclosed, the memory type, product name, and sample availability timing were not. Actual processing speeds and the range of AI workloads the design can handle will have to be judged from more detailed specifications released later.

AD

Combining pSLC and QLC to spread the cache across the whole SSD

The third technology, Mixed Media, is a storage design that combines pSLC (pseudo-SLC), which prioritizes write performance, with high-capacity QLC NAND flash. QLC increases storage capacity by recording 4 bits per memory cell, but it comes with constraints in write performance and endurance. pSLC runs NAND flash at 1 bit per cell, prioritizing write performance.

In Phison's Mixed Media, a single controller manages pSLC and QLC across 16 channels. By giving each SSD both a write cache and a large storage area, the design can use drive bays more efficiently than a configuration that adds separate dedicated cache SSDs. The company aims for a design that uses the standard NVMe interface and can be deployed without adding dedicated software on the host side.

There is precedent for SSDs that combine different recording methods. DapuStor's J5060 uses a design in which the capacity split between SLC and QLC can be adjusted. What Phison emphasized this time is that distributing the cache across each SSD limits the share of cache capacity lost when a drive fails.

In the 24-drive storage array Phison showed, if one SSD fails under the Mixed Media approach, only about 4% of total cache capacity is lost. Assuming each SSD is allocated the same amount of cache, the share for one drive is:

By contrast, with the dedicated-cache approach the company presented as a comparison, one failure loses 25% of cache capacity. If the cache is evenly allocated across four dedicated SSDs, losing one means losing a quarter. In other words, even with the same total number of SSDs in the array, the share lost on a failure differs depending on whether the cache is concentrated in a few drives or spread across every drive.

However, distributing the cache does not by itself determine sustained write performance. On QLC SSDs, data temporarily stored in the pSLC cache must be moved to the QLC area. Garbage collection, which frees up space, can also move valid data to other locations.

The phenomenon in which internal operations cause more data to be written to the NAND flash than the host actually wrote is called write amplification. According to Phison's technical explanation, garbage collection competes with host I/O and therefore also affects write speed and response time. To judge Mixed Media's real-world performance, it will be necessary to check how write speed and latency change under sustained loads that exhaust the cache.

The three technologies improve different metrics

Phison's official SC26 announcement presented three different approaches to improving SSD performance.

The X3's target of up to 28 GB/s is a goal for data transfer speed between the SSD and the host. The one-tenth to one-hundredth figure for computational storage is an estimate of how much the data transferred to the host can be reduced for specific workloads, not an improvement rate in processing speed itself. Meanwhile, Mixed Media's roughly 4% figure indicates the share of cache capacity lost when one drive fails in a 24-drive array.

Evaluating the three technologies therefore requires a different metric for each.

For the X3, it is important to measure actual transfer speed and IOPS with data sizes and read/write conditions held constant. For computational storage, the same task needs to be run on both the host side and the SSD side, comparing not only the amount of data transferred but also total processing time. For Mixed Media, the evaluation should cover write speed and response time during sustained writes that exhaust the cache, as well as performance during a drive failure, remaining capacity, and the effect on data protection.

At this point, no measured results under matched conditions, nor specific availability dates for X3-based products, have been published. According to Phison, many of the products and features introduced are in development, and availability may change depending on development progress and market conditions.

In storage for AI and HPC, not only the transfer speed to the host but also the volume of data transferred and the sustainability of write operations influence overall system performance. Each of these three technologies aims to address a different constraint, and how much improvement they deliver in real workloads will become clear through future productization and hands-on evaluation.