Phison Electronics unveiled three enterprise storage architectures at SC26, aimed at speeding data movement in AI servers. The lineup pairs PCIe Gen6 SSD controllers with drives that handle data internally and blend cache with bulk storage. Phison says the designs target server bottlenecks, but no commercial release dates have been announced.
PCIe Gen5 drives are hitting their bandwidth ceiling, and PCIe Gen6 doubles per‑lane speeds to 64 GT/s. Phison’s upcoming X3 (PS5303) enterprise controller is rated for 28 GB/s sequential throughput and 6,800,000 IOPS over a PCIe Gen6 x4 NVMe 2.3 link almost twice Gen5 performance while improving throughput per watt. Because faster signals decay quickly on circuit boards, Phison develops its controllers, retimers and redrivers as a package. Its PS7161 chip is already the first PCIe 6.0 redriver on the PCI SIG Integrators List.
Shuttling massive datasets between SSDs, memory and processors wastes both time and power. Phison sidesteps that by embedding compute hardware directly in the drive. The drives can scan and filter data in parallel, returning only the needed results instead of moving raw files. That reduces data transfers by a factor of 10 to 100 on supported workloads and frees the host CPU. Inside each drive sits an NPU rated at 36,800 DMIPS alongside 64 GB of memory capable of 130 GB/s. Developers can use familiar frameworks such as PyTorch, vLLM, OpenVINO and llama.cpp without learning new toolchains.
Traditionally, server builders separate high‑endurance cache drives from bulk‑storage SSDs. Phison Mixed Media combines pSLC and QLC flash behind a single controller across 16 channels, so every drive bay delivers usable capacity and eliminates dedicated cache hardware. Distributing cache across the array also cushions drive failures: losing one drive out of 24 cuts total cache by only 4 % instead of 25 %. Because it relies on standard NVMe interfaces, the solution works with existing servers without custom host software.
Michael Wu, GM and President of Phison US, said hardware progress now depends on smarter data handling rather than raw processor speed.
The next era of computing won’t be defined by processing power alone, but by how efficiently we can move, process and store the data that fuels it.
