Kioxia CD9P-R Review: Gen5 Speed Meets an E3.S Constraint

Kioxia’s 7.68TB CD9P-R exceeded its rated speeds in ServeTheHome Kioxia testing, reaching 14.9GB/s reads and 7.2GB/s writes. Those results put the drive near the front of a crowded PCIe Gen5 field. However, its read-intensive endurance and E3.S form factor make the buying decision more complicated than the headline numbers suggest.

The CD9P-R combines Kioxia’s BiCS FLASH generation 8 TLC NAND with an in-house controller and firmware. It supports PCIe 5.0 x4 and NVMe 2.0. Kioxia rates this capacity for 14,800MB/s sequential reads, 7,000MB/s writes, and 2.6 million random read IOPS.

The real contest is not Kioxia against one rival. It is peak Gen5 performance against workload fit. Samsung, Solidigm, Micron, Phison, SanDisk, and Kingston offer alternatives with different latency, endurance, capacity, and platform profiles. The CD9P-R looks strongest when buyers value read speed, predictable low-depth latency, and storage density more than sustained write endurance.

What the ServeTheHome Kioxia Test Actually Changed

The new evidence shows that the CD9P-R can meet its advertised Gen5 performance without depending on one familiar server architecture.

ServeTheHome tested the 7.68TB CD9P-R across platforms from AMD, Intel, Ampere, and Xsight Labs. That architectural spread matters because modern data centers no longer rely on one nearly uniform host platform. Storage devices increasingly sit behind varied CPUs, accelerators, PCIe switches, and data processing units.

The review measured 14.9GB/s in a sequential 128K read workload. It also recorded 7.2GB/s for sequential writes. Both results slightly exceeded Kioxia’s stated ratings of 14.8GB/s and 7GB/s for this model.

These are steady-state measurements, not a short consumer benchmark that can hide cache exhaustion. Enterprise tests generally precondition a drive before measuring it. That process pushes the SSD toward the sustained behavior expected after extended use.

The CD9P-R review also placed the drive at or near the top of its comparison set in several read-focused workloads. The comparison included Kioxia’s earlier products and drives from Samsung, Solidigm, and Phison.

That result supports a narrow but important conclusion. Kioxia’s rated performance does not appear inflated for the tested sample and conditions. It does not prove that every workload, firmware version, or server will produce identical numbers.

The review also examined performance across several host architectures. Differences were small, but they were present. Small variations matter when operators multiply them across eight drives, several servers, or a distributed storage tier.

ServeTheHome included Intel Xeon 6, AMD EPYC, AMD Threadripper, AmpereOne, and an Xsight Labs E1 DPU. A DPU is a programmable processor that moves infrastructure work away from the host CPU. Storage, networking, and security services are common DPU tasks.

The Xsight platform is especially relevant to AI infrastructure. Its E1 DPU is designed to connect eight PCIe Gen5 NVMe drives through 800Gbps network links. That design moves storage closer to a networked service than a traditional collection of local disks.

ServeTheHome said its team worked with the Xsight firmware group to make the test configuration operate correctly. That detail exposes an often-hidden part of enterprise storage adoption. A fast drive still depends on firmware, topology, cooling, and platform validation.

The review did not publish enough architecture-specific raw data to declare one CPU family the universal winner. Its broader finding is more useful. The drive maintained its central performance story across several increasingly different computing platforms.

That is what changed with this test. The CD9P-R moved from a specification sheet to a tested Gen5 component with evidence across multiple architectures. The result strengthens its case, but it also shifts attention toward deployment constraints.

E3.S Makes Density Part of the Performance Story

The CD9P-R is not simply a faster U.2 drive because its E3.S design changes server density, airflow, connectivity, and replacement planning.

E3.S belongs to the Enterprise and Data Center Standard Form Factor family, commonly called EDSFF. It uses a thinner rectangular design than a conventional 2.5-inch U.2 enclosure. The tested drive measures 7.5mm thick, 76mm wide, and 112.75mm long.

That shape lets server manufacturers place more drives across a chassis front panel. A thin carrier also leaves more room for airflow between storage devices. Both features matter when each active drive can consume 23 watts under Kioxia’s typical specification.

The connector is smaller than the familiar U.2 interface while still carrying a PCIe Gen5 x4 connection. ServeTheHome argues that this can improve cooling and signal integrity. Those benefits become more valuable as servers move toward faster interfaces.

PCIe 5.0 transfers data at 32 gigatransfers per second on each lane. Four lanes give the CD9P-R an aggregate interface rate of 128 gigatransfers per second. Encoding and protocol overhead mean usable throughput remains below that raw figure.

The compact connection also signals a broader transition. U.2 and U.3 preserve physical and electrical links to the older 2.5-inch storage ecosystem. EDSFF starts from data-center density, thermal behavior, and PCIe requirements instead.

That transition creates pressure for server vendors and operators. A buyer cannot insert an E3.S drive into a U.2 bay merely because both devices use NVMe. The chassis, backplane, carrier, cabling, firmware, and service procedures must support the new format.

Kioxia still offers 2.5-inch versions of the CD9P-R. Those models reach 61.44TB because the thicker enclosure provides more internal room for NAND packages. The E3.S family tops out at 30.72TB, according to Kioxia’s current product information.

The tested 7.68TB model sits at an interesting point in the lineup. It uses generation 8 BiCS FLASH, while the 1.92TB and 3.84TB E3.S models use generation 5 flash. Buyers should not assume every capacity has the same NAND or performance profile.

Kioxia’s E3.S specifications list 14,800MB/s sequential reads for the 7.68TB and 15.36TB capacities. The 30.72TB model drops to 13,500MB/s despite offering four times the tested capacity.

Random write performance follows a similar pattern. The 7.68TB and 15.36TB versions carry a rating of 450,000 IOPS. The 30.72TB model is rated for 270,000 IOPS.

This capacity wrinkle makes qualification more important. Operators often validate one model and assume a larger version behaves like the same drive with additional flash. Kioxia’s own specifications show that capacity changes can alter performance.

The E3.S format therefore serves as more than packaging. It supports dense arrays of drives that can feed CPUs, GPUs, and networked storage processors. Yet it also narrows the set of systems that can accept the device.

This creates the article’s central tradeoff. Kioxia offers near-interface-limit read performance in a format suited to newer dense servers. Buyers with established U.2 fleets must weigh that benefit against migration and interoperability work.

A greenfield AI cluster can make E3.S part of its chassis decision from the beginning. An existing cloud environment may need a slower replacement cycle. The same drive can be an obvious fit in one environment and an operational detour in another.

BiCS 8 Delivers Speed, but Workload Fit Decides the Winner

The CD9P-R’s strongest results come from matching BiCS 8 flash to read-heavy work, not from winning every enterprise benchmark.

BiCS FLASH generation 8 is Kioxia’s three-dimensional NAND design. Three-dimensional NAND stacks memory cells vertically to increase density without relying only on smaller planar features. The CD9P-R uses triple-level cell flash, which stores three bits in each cell.

Kioxia combines that NAND with its own controller and firmware. Vertical integration can help a vendor tune error correction, flash management, power behavior, and performance as one system. It does not eliminate the need for independent testing.

The ServeTheHome results align with other testing published during 2026. TweakTown measured 15,044MB/s sequential reads and 7,200MB/s writes from its 7.68TB sample. It also reported 2.758 million 4K random read IOPS.

Those results exceeded the relevant factory ratings under TweakTown’s configuration. Its Gen5 SSD testing also emphasized strong performance at low and moderate queue depths.

Queue depth describes how many input and output requests wait for processing. Very high queue depths can expose a drive’s maximum parallel throughput. Lower depths often resemble applications where individual request latency matters more than benchmark saturation.

A separate StorageReview test found the same distinction. The CD9P-R performed especially well in low-depth random reads. Competing drives caught up or passed it as concurrency increased in some workloads.

At the lowest tested 64K random read setting, StorageReview measured about 1,334MB/s from the CD9P-R. At maximum concurrency, it stabilized around 11GB/s to 12GB/s. Several competing drives reached between 13.5GB/s and 14.2GB/s there.

Its 128K sequential read latency measured 561.7 microseconds. The Phison Pascari X200P led that test by only 0.3 microseconds. Several Gen5 competitors finished within eight microseconds of the leader.

That narrow grouping suggests an interface ceiling in highly sequential workloads. Once several controllers and NAND designs can fill most of a Gen5 x4 link, raw bandwidth becomes less useful for separating products.

Random and mixed workloads expose larger differences. StorageReview found that the CD9P-R started its 4K random read sweep at about 32,300 IOPS. The drive lost the overall lead at peak concurrency, where SanDisk’s DC SN861 reached about 2.556 million IOPS.

The same latency analysis placed the CD9P-R in the middle of the group for several write-heavy tests. Micron’s 9550 Max often led those write-focused comparisons.

These results do not conflict with ServeTheHome’s positive conclusion. They define it more carefully. The CD9P-R is a fast read-intensive product, but its ranking changes with request size, queue depth, and read-write balance.

That distinction matters for databases. Online transaction processing rarely resembles one endless sequential read stream. It combines reads, writes, metadata operations, synchronization, and latency-sensitive transactions.

Content delivery and media streaming can place heavier emphasis on reads. Model loading, checkpoint distribution, and retrieval infrastructure may also reward high read bandwidth. However, AI training checkpoints can generate large bursts of writes.

Virtualization creates another mixed profile. Many virtual machines can produce small, irregular operations simultaneously. Average bandwidth alone will not predict tail latency or performance during busy periods.

Buyers should model the actual request mix before selecting a drive. They should record block sizes, queue depths, read-write ratios, write amplification, and latency percentiles. A benchmark winner under one profile can trail under another.

The CD9P-R’s appeal becomes clearest after this filtering. It offers strong reads and competitive latency without claiming the endurance of a write-focused product. That honest segmentation is useful, provided buyers respect it.

One Drive Write per Day Is the Hard Boundary

The 1-DWPD endurance rating separates an excellent read-tier SSD from a safe default for every data-center workload.

DWPD means drive writes per day. It estimates how many times a drive’s full stated capacity can be written daily during its warranty period. Kioxia rates the CD9P-R for one DWPD across five years.

For a 7.68TB drive, one complete logical drive write equals 7.68TB per day. Over five years, that corresponds to roughly 14 petabytes of host writes. Actual permitted totals depend on Kioxia’s warranty definitions and measurement rules.

That endurance level fits many read-intensive services. Content delivery, data distribution, search indexes, inference assets, and replicated read tiers can spend much of their time serving existing information.

It is less comfortable for sustained logging, write caches, high-churn databases, or checkpoint-heavy training pipelines. These uses can write far more data than their visible application payload suggests.

Write amplification explains part of that gap. NAND flash cannot overwrite existing cells directly. Controllers reorganize pages and erase larger blocks, so one host write can trigger additional physical writes inside the SSD.

Garbage collection, parity, compaction, database journaling, and storage replication can increase total activity. A workload averaging below 7.68TB of host writes can still create meaningful wear and latency pressure.

Kioxia offers the CD9P-V for mixed-use deployments. ServeTheHome notes that the V variant supports three DWPD over five years. Its additional spare capacity reduces the available capacity of the corresponding configuration to 6.4TB.

This is the clearest opponent map for the CD9P family. The R model maximizes usable capacity for read-intensive work. The V model sacrifices visible capacity to absorb heavier write cycles.

Neither route is universally better. A CD9P-R deployed into a stable read tier can provide more usable capacity per drive. The same product in a write-heavy service can create avoidable endurance and replacement risk.

Power also deserves qualification. Kioxia lists typical active consumption at 23 watts and ready consumption at five watts. A dense E3.S chassis multiplies both values across many devices.

Eight active drives at the typical figure account for 184 watts before CPUs, memory, networking, fans, and power conversion losses. Real consumption varies with workload, firmware, thermal state, and individual device behavior.

Cooling can affect sustained performance. E3.S improves airflow opportunities, but the form factor does not cool itself. Server designers still need correct carriers, airflow direction, fan curves, and inlet temperatures.

Kioxia specifies an operating range from zero to 75 degrees Celsius. That limit is not a recommended target temperature. Operators should follow the server vendor’s qualification and thermal guidance for sustained use.

Security and data protection features also require model-level verification. Kioxia lists power-loss protection and end-to-end data protection. It offers security options that include sanitize instant erase and self-encrypting drive configurations.

These options can involve different model numbers. Procurement teams should verify whether a quoted unit includes the required security behavior. A family-level feature list does not guarantee identical configuration across every order.

Platform support creates another source of uncertainty. The CD9P-R supports NVMe 2.0 and parts of the Open Compute Project Datacenter NVMe SSD specification version 2.5. Kioxia explicitly says not all OCP requirements are supported.

Firmware validation therefore remains essential. A server recognizing the drive is only the first step. Teams should test hot-plug behavior, telemetry, error reporting, firmware updates, power states, and recovery after unexpected resets.

ServeTheHome’s work with Xsight Labs illustrates this point. Cross-architecture operation can require firmware coordination even when the underlying connection follows PCIe and NVMe standards.

The skeptical conclusion is straightforward. Strong benchmark results reduce performance uncertainty, but they do not remove endurance, integration, or thermal risk. Those variables belong in the purchasing decision beside throughput.

The Next Three Signals Will Decide the CD9P-R’s Staying Power

The CD9P-R has cleared its first performance test, while fleet validation, rival latency, and Gen6 platform design will determine its longer-term position.

The first signal is broader platform qualification. More server vendors need to publish support for the 7.68TB and higher-capacity E3.S models. Buyers should look for validated firmware, hot-plug operation, telemetry, and predictable thermal behavior.

If that support expands, the ServeTheHome Kioxia result becomes easier to translate into production. The drive will no longer depend as heavily on individual engineering teams solving platform-specific issues.

Weak support would reduce its practical reach. An SSD can lead several benchmarks while remaining unsuitable for a fleet that standardizes on different backplanes or management tooling.

The second signal is performance under application-shaped workloads. Competitors already approach Gen5 x4 limits during sequential reads. Product separation will increasingly come from low-depth latency, mixed workloads, consistency, and tail behavior.

Micron’s write-focused performance shows where Kioxia still faces pressure. SanDisk, Phison, Solidigm, Kingston, and Samsung add further alternatives. Their next firmware and capacity revisions can change the ranking without changing the PCIe generation.

Independent results should therefore include database traces, virtualization mixes, AI checkpoint traffic, and retrieval workloads. Average throughput should appear beside high-percentile latency and performance-per-watt measurements.

If the CD9P-R maintains its low-depth advantage across these profiles, the read-intensive label becomes a competitive strength. If rivals match that latency while offering higher endurance, Kioxia’s usable-capacity advantage becomes less decisive.

The third signal is the design of upcoming PCIe Gen6 storage platforms. ServeTheHome expects EDSFF to become increasingly important as signal-integrity demands rise. Gen6 doubles the transfer rate again, making connector and board design harder.

A broad Gen6 shift toward E3.S would validate Kioxia’s form-factor direction. It would also let today’s E3.S buyers build operational experience before faster drives arrive.

Continued U.2 demand would preserve a split market. Kioxia already serves that demand with its 2.5-inch CD9P-R models. However, operators might delay E3.S adoption if their existing format remains competitive and easier to service.

The near-term verdict is favorable but specific. The 7.68TB CD9P-R delivers the advertised Gen5 read performance, and multiple laboratories have found strengths at practical queue depths.

It is not a universal storage answer. One-DWPD endurance excludes some write-heavy roles, while E3.S compatibility excludes older or differently configured servers. Peak speed cannot erase either boundary.

Infrastructure buyers should now test the drive with their own traces on the exact intended server. Measure sustained latency, write volume, thermals, power, firmware behavior, and recovery operations. Compare those results with at least one mixed-use drive and one direct read-intensive competitor.

That process turns the ServeTheHome Kioxia benchmark into a useful starting point rather than a purchasing shortcut. If your workload is read-heavy and your platform supports E3.S, the CD9P-R deserves a qualification slot.