Silicon Motion has started work on a PCIe 7 enterprise SSD controller, according to Alex Chou, the company executive who runs its enterprise storage unit. That is early by normal buyer standards, but not early by controller standards: the silicon has to be close to ready by the time server vendors and cloud operators decide they need it.
Chou told Tom’s Hardware that Silicon Motion has already defined the overall architecture for its PCIe 7 enterprise controller platform. He said the company is aiming for internal samples in the second half of 2027, with production planned in the same broad period.
The nearer product is PCIe 6. Chou said Silicon Motion’s PCIe 6 controller design is effectively finished, is running in FPGA form for algorithm emulation, and should tape out soon. If the schedule holds, first silicon would return in the second half of 2026.
Enterprise SSDs are splitting by workload
Silicon Motion is best known for SSD controllers, especially in client drives, but Chou said its enterprise PCIe 5 products are now moving into volume production with multiple OEM customers. He said the company remains early in its enterprise ramp and wants to eventually pass 10 percent share in what he described as a $4 billion enterprise SSD controller market.
Chou separated enterprise SSD demand into three buckets: conventional server and storage deployments, high-density AI and hyperscale storage, and storage placed closer to GPUs. The last category is where the old SSD spec-sheet religion of sequential throughput starts to look incomplete.
For GPU-adjacent storage, Chou said latency and quality of service matter more because systems may move large datasets or offloaded KV cache between accelerators, memory, and storage. Silicon Motion’s answer is a traffic-shaping engine that Chou said can prioritize and isolate workloads to reduce latency spikes when multiple tenants, applications, or virtual machines hit the same SSD.
That mechanism is not magic. It is controller scheduling with consequences. If a noisy workload can monopolize access to NAND or internal controller resources, the drive’s average bandwidth may look fine while tail latency turns ugly. Chou’s pitch is that Silicon Motion can keep those outliers under tighter control.
Nvidia’s Storage Next is shaping the roadmap
Chou said Nvidia’s Storage Next effort has become one of the industry developments Silicon Motion is watching most closely. The concept, as he described it, moves storage into a tighter path with GPUs rather than treating it as a distant block device feeding CPUs.
Nvidia has discussed very high IOPS targets for that class of storage. Chou said 100 million or 200 million IOPS may not be realistic soon, and that the target appears to be moving closer to about 50 million IOPS. He said Silicon Motion’s architects have been involved in Storage Next discussions from the start and are trying to keep future controller designs flexible enough to support it if the market forms.
That last caveat matters. Chou was less bullish on storage-class memory ideas such as Kioxia’s XL-Flash, saying Kioxia is effectively the only company still pushing that technology and that a broader ecosystem has not formed. He said conventional NAND remains attractive to memory vendors while prices are strong.
Future NAND still makes controller work harder. Chou said Silicon Motion is using 16KB LDPC collaborative codewords in its first enterprise PCIe 6 controller, the SM8466, and is more likely to refine that error-correction engine for PCIe 7 than expand it beyond 16KB. The problem set now includes interface speed, signal integrity, power, security, quality of service, and NAND generations that may behave differently from today’s parts.
Chou said Silicon Motion works with major NAND suppliers including Samsung, SK hynix, SanDisk, and Kioxia so its controllers and firmware can support multiple memory sources. For cloud buyers in a tight supply market, that is the useful part: fewer controller designs welded to one NAND vendor’s roadmap.
This story draws on original reporting from Tom's Hardware.
