Forward-looking: A lot of what AI servers do all day involves pulling data from their SSDs so a CPU or GPU can process it. Phison wants some of that processing to happen before the data ever leaves the drive. The controller maker has detailed an SSD design with its own NPU and 64GB of memory, designed to run PyTorch and llama.cpp directly on the drive.
The company showed off the design at Supercomputing 2026 (SC26) in Chicago this week. Rather than sending entire datasets to the host, the drive scans and processes data in parallel right where it sits. It then sends back only the results that matter. Phison claims this could cut data transfers by 10 to 100 times for suitable workloads. The host CPU would get some breathing room, too.
On paper, the drive's memory delivers 130 GB/s of bandwidth. That's more than four times the 28 GB/s that Phison's upcoming PCIe 6.0 controller can deliver to a host. In other words, the drive can read its own data far faster than it can transfer it to the server. The company also says developers won't need specialized programming tools because the drive supports familiar frameworks such as vLLM and OpenVINO.

Oddly, the only compute figure Phison shared for the NPU is 36,800 DMIPS. That's a decades-old CPU integer benchmark. As Tom's Hardware points out, the figure likely describes companion CPU cores on the drive rather than its AI processing power. There's no TOPS rating anywhere in the announcement.
The drive also has no product name or sampling date yet. Phison hasn't even said what type of memory it uses. In fact, the company says many of the products and features it showcased are still in development. For now, then, this is more of an architecture proposal than something a data center can actually order.
Drives that perform computations aren't new, either. Samsung put an FPGA on its SmartSSD back in 2018 and followed up with a second generation in 2022. But TechRadar reported last year that the drive had all but disappeared from Samsung's lineup. The outlet partly blamed the generative AI boom, which created much greater demand for raw storage capacity than for on-drive processing.
Samsung's drive also relied on software and IP that customers had to develop themselves. That makes Phison's promise to support the AI tools developers already use a significant difference. It could give the idea a better chance this time around, assuming the NPU turns out to be as capable as Phison claims.