Reply to thread

Message: [QUOTE="Shamiso, post: 92539, member: 160"] CXMT has discussed its planned NAND output with potential customers, including an unnamed startup building storage products for AI systems and supercomputers. Early customer talks matter, but they are nowhere close to a production deployment. The company has not announced a commercial NAND die, controller partner, finished SSD, qualification sample, endurance rating, or shipment date. [B][URL='https://goldmidi.com/community/threads/cxmt-prepares-beijing-nand-expansion-beyond-dram.77891/']CXMT’s reported NAND R&D line[/URL][/B] is still the step where engineers prove the flash itself can be manufactured consistently. A buyer would then have another stack of work. Bare NAND has to survive its own characterization, while a finished enterprise SSD has to prove the controller, firmware, power-loss behavior, error handling, latency, thermals, and host interface all behave under the workload that will actually hit the drive. [HEADING=2]A NAND die is not an enterprise SSD[/HEADING] Flash leaves the wafer fab as memory, not as a trustworthy server drive. An SSD vendor still has to combine NAND packages with a controller, firmware, error-correction logic, power management, mapping tables, spare capacity, and often capacitors for power-loss protection. Each piece changes how the flash behaves. Stronger error correction can extend usable life, overprovisioning can reduce write amplification, and firmware decides how aggressively the controller moves data, retires blocks, handles garbage collection, and reacts when cells become harder to read. This is why a good NAND die can still produce a bad SSD. Firmware bugs, controller problems, thermal throttling, unexpected write amplification, or poor recovery behavior can turn healthy flash into a drive an enterprise customer will reject. Large-scale field data backs up the distinction. [B][URL='https://www.usenix.org/conference/fast20/presentation/maneas']Enterprise SSD reliability in real deployments[/URL][/B] has shown that firmware version, drive family, NAND type, usage, and system context can all matter once drives leave controlled testing and spend years inside storage arrays. CXMT could therefore reach acceptable NAND yield without automatically becoming an enterprise SSD supplier. A storage startup buying its dies would inherit the integration job, and every controller or firmware change could force parts of the qualification cycle to run again. [HEADING=2]Qualification starts with retention and endurance[/HEADING] Enterprise endurance testing is stricter than a headline terabytes-written number. JEDEC-style enterprise qualification ties endurance to a defined workload while also requiring the drive to retain data after power is removed and stay below a specified uncorrectable bit error rate. For enterprise-class SSDs, the commonly used JESD218 framework sets the powered-off retention condition at three months at 40 degrees Celsius after endurance use. It also specifies an uncorrectable bit error rate no worse than one error per 10 quadrillion bits read. Workload shape matters just as much. Random small writes generally punish NAND harder than neat sequential transfers because they create more internal movement, garbage collection, and write amplification, so endurance claims only make sense beside the workload used to produce them. Raw NAND evaluation happens earlier and digs into cell behavior. Engineers watch program and erase cycling, retention drift, read disturbance, bad-block growth, error rates, voltage-window movement, and how much correction margin remains as the flash wears. A customer may also sample material across several production lots rather than trusting one impressive batch. Consistency is the point. A die that works beautifully on one wafer but shifts across lots creates a qualification headache for firmware tuning and long-term supply. [HEADING=2]Firmware and platform behavior can sink qualification[/HEADING] Current data-center SSD requirements go well beyond keeping bits readable. OCP’s data-center NVMe specification requires full power-loss protection for acknowledged data and metadata, along with health checks that verify the protection system still has enough stored energy to do its job. Latency gets similar scrutiny. Average throughput can look great while occasional stalls wreck a database or distributed storage system, so hyperscale buyers care about tail latency, quality of service, thermal behavior, and whether maintenance work inside the SSD creates ugly pauses. Firmware becomes part of the product contract. Buyers need predictable error reporting, telemetry, secure update behavior, namespace handling, recovery after abrupt resets, and compatibility with the exact server platforms and operating stacks they plan to deploy. Qualification can also expose boring failures that benchmark charts miss. A drive may mishandle surprise power removal, negotiate poorly with a host, behave differently after a firmware update, or fall outside latency targets only when nearly full and under sustained mixed writes. CXMT’s prospective customer discussions are therefore meaningful for a different reason than an early sales headline suggests. A willing storage partner can give the company real workload requirements and feedback while its NAND process is still young, letting controller and firmware work develop alongside the flash instead of starting after mass production. Commercial credibility arrives when samples survive repeated lots, endurance testing, retention requirements, power interruptions, controller integration, firmware validation, and sustained workload testing without nasty edge cases. Until those milestones appear, a customer conversation proves interest, not deployment readiness. [/QUOTE]

Name