Draft 03 | Primary source: Samsung Semiconductor's official technology blog. Verification date: 2026-11-15. Note: the org-ap-publish-cn mirror of this article could not be fetched, so verification used the same article on Samsung's global website, read in full; key specifications were cross-checked against Samsung's data center SSD overview page, the PM1743 product page, and the Chinese-language launch release.
What Gen5 Enterprise SSDs Actually Compete On

PCIe 5.0 x4 offers a theoretical ceiling of roughly 15.75 GB/s (calculated from 32 GT/s with 128b/130b encoding), and today's flagship enterprise drives already read sequentially at 12,000–14,000 MB/s — nearly touching the interface roof. The "faster sequential reads" card, in other words, has already been played. Samsung's PM9D3a blog opens with unusual honesty: the real constraints in a data center are rack power, space, and cooling, and storage has to deliver on power, density, and performance at the same time.
So the Gen5 enterprise fight has moved to four harder dimensions: random small-IO throughput and latency, performance per watt, form-factor density (E3.S and E1.S displacing 2.5-inch), and the invisible-but-operational capabilities — telemetry and security compliance — that ops teams end up paying for. On a many-core server, what decides the experience usually isn't how fast a big file copies; it's whether the drive holds steady when hundreds of concurrent threads all want 4K reads at once. The PM9D3a is aimed squarely at that fight.
The PM9D3a Spec Sheet: What the Official Blog Actually Says
| Item | Official figure | Source |
|---|---|---|
| Interface | PCIe Gen5 x4, NVMe | Blog + overview page |
| Flash | 8th-generation 3-bit V-NAND (V8 TLC) | Blog |
| Sequential read/write | Up to 12,000 / 6,800 MB/s | Blog + overview page |
| Random read/write | Up to 2,000K / 400K IOPS | Blog + overview page |
| Random-read gain | Up to 2.4× the previous generation | Blog claim |
| Power efficiency | ~45% better than the previous generation | Blog |
| Capacity | 960 GB–30.72 TB; M.2 from 240 GB–1.92 TB | Blog |
| Form factors | E1.S (25/15/9.5 mm), E1.L (9.5 mm), E3.S (1T/2T), U.2/U.3 2.5-inch (15 mm), M.2 | Blog |
| Endurance | 1 DWPD, 5 years | Overview page |
| MTBF | 2.5 million hours | Blog |
| Cooling | 2.5-inch 15 mm finned heatsink; E1 series with built-in heatsink + thermal interface material | Blog |
| Features | FDP optional; OCP v2.5-compliant version available on request | Blog |
| Security | TCG Storage Opal family test-validated; FIPS 140-3 validation in progress | Blog |
| Telemetry | Over a hundred data sets for analysis | Blog |
Two details are easy to skim past. First, the blog is explicit that peak performance figures apply to the U.2/E3.S form factors — the M.2 versions can't borrow those numbers. Second, the FIPS 140-3 wording is "undergoing" validation, not "certified." In a tender document, that one-word difference is a compliance risk.
Where the 2.4× Random-Read Claim Comes From — and What It Means for Many-Core Servers
Random reads aren't an interface-bandwidth contest; they're a contest of controller parallelism and NAND read latency. Run the math: 2,000K IOPS × 4 KiB ≈ 8.2 GB/s, only about half of Gen5 x4's theoretical bandwidth. The bottleneck isn't the link — it's how many dies the controller can schedule at once and how fast firmware works through the command queues. The OCP marketplace page discloses that the PM9D3a uses an 8-channel controller; paired with V8 NAND's interface speed and manufacturing consistency (what the blog calls "production efficiencies"), that's the physical basis for the random-read jump.
For many-core servers this metric is far more concrete than sequential reads. When a 64-core — or 128-plus-core — machine runs virtualization, containers, or databases, the flood of concurrent 4K reads directly sets the IO-wait share. Raising the random-read ceiling effectively hands every core a bigger IO budget, so CPU utilization stops being dragged down by storage.
But one point deserves real scrutiny. The OCP page names the PM9A3 as the PM9D3a's predecessor, and Samsung's own overview page lists the PM9A3's random-read maximum at 1,000K IOPS. Divide the two headline maxima — 2,000K ÷ 1,000K — and you get 2.0× (calculated). The blog's "up to 2.4×" implies a baseline of roughly 833K IOPS (2,000 ÷ 2.4, calculated), which suggests the comparison used a specific capacity or test configuration of the previous generation rather than its top model. The blog doesn't state the comparison conditions. At procurement time, ask the supplier for the baseline configuration and a SNIA-conformant test report. Every "up to" number is a peak under specific conditions; production environments should care more about sustained performance and the latency distribution (QoS).
Side by Side with the PM1743: Two Gen5 Routes
| Dimension | PM9D3a | PM1743 |
|---|---|---|
| Positioning | Data center scale-out (blog/overview) | Enterprise mission-critical (official enterprise page) |
| Sequential read | 12,000 MB/s | 13,000 MB/s at launch; current product page updated to up to 14,000 |
| Random read | 2,000K IOPS | 2,500K IOPS (launch release) |
| Sequential write | 6,800 MB/s | 6,600 MB/s (launch release) |
| Random write | 400K IOPS | 250K IOPS (launch release) |
| Port redundancy | Dual port not listed as a selling point in the blog | Dual port with failover on single-port failure |
| Efficiency claim | +45% vs previous generation | Product page: +48% vs PM1733a; launch release also cites 608 MB/s/W, ~+30% |
| Form factors | E1.S/E1.L/E3.S/U.2/U.3/M.2 | 2.5-inch + E3.S (7.5 mm) |
| Capacity ceiling | 30.72 TB | 15.36 TB (launch release; from 1.92 TB) |
The division of labor is actually clean. The PM1743 reads randomly faster and carries dual-port redundancy — the right fit for databases, finance, and other can't-stop workloads. The PM9D3a writes randomly at about 1.6× the rate (400K ÷ 250K, calculated), doubles the capacity ceiling, and spans a wider form-factor family — built for scale-out and AI data pipelines. Two cautions: the PM1743's launch release and current product page disagree (13,000 → 14,000 MB/s), meaning specs were updated after launch — cite the latest official page. And the two efficiency claims use different baselines (each against its own predecessor), so 45% and 48% cannot be compared head-to-head.
Selection and TCO: Straight Talk for Guangzhou and Pearl River Delta Buyers
Pick the form factor before you argue about performance. E3.S needs a new-generation backplane and chassis; if you're upgrading an installed 2.5-inch U.2 chassis smoothly, choose the PM9D3a's 2.5-inch version. M.2 is for boot duty only. A Gen5 drive in a Gen4 backplane won't throw an error — it will just waste half the money you spent.
Do the endurance math. At 1 DWPD, the 15.36 TB version can absorb roughly 15.36 TB of writes per day, about 28,032 TB over 5 years (15.36 × 365 × 5, calculated). For write-heavy log and cache workloads, measure average daily writes before choosing capacity — often buying one capacity tier up to dilute write pressure is cheaper than chasing a higher-endurance variant.
Efficiency and cooling are live issues in Pearl River Delta server rooms. Guangzhou is hot and humid, and rack power densities run high; a 45% efficiency gain means more usable IOPS inside the same power and cooling budget, and the finned and E1 built-in thermal designs can keep fan speeds down. FDP (Flexible Data Placement) has real potential to cut write amplification under multi-tenant mixed writes, but it needs software-stack support — validate before deployment rather than counting it as a default benefit.
On the operations side, pipe the blog's hundred-plus telemetry data sets into your monitoring for predictive maintenance, and confirm the firmware upgrade path and the OEM warranty entity with your supplier (OEM drives typically follow the system's warranty terms). Risk flags to carry into any order: peak figures are not sustained performance; the "2.4×" baseline needs clarification; FIPS 140-3 validation is not complete; and Gen5 backplanes, CPU PCIe lane budgets, and lead times all need to be locked down before you commit.
Sources
- Samsung Semiconductor, *Samsung's PM9D3a Solid State Drive* (official technology blog) — https://semiconductor.samsung.com/news-events/tech-blog/samsung-pm9d3a-solid-state-drive/ — page structured data gives a publication date of 2026-09-04; accessed and verified 2026-11-15.
- Samsung Semiconductor, *Datacenter SSD* official overview (PM9D3a / PM9A3 specification blocks) — https://semiconductor.samsung.com/ssd/datacenter-ssd/ — accessed 2026-11-15.
- Samsung Semiconductor, *PM1743* product page — https://semiconductor.samsung.com/ssd/enterprise-ssd/pm1743/ — accessed 2026-11-15.
- Samsung Semiconductor China, *Samsung Develops High-Performance PCIe 5.0 SSD for Enterprise Servers* (PM1743 launch release; states mass production from 2022-03) — https://semiconductor.samsung.cn/news-events/news/samsung-develops-high-performance-pcie-5-0-ssd-for-enterprise-servers/ — accessed 2026-11-15.
- Open Compute Project Marketplace, *Samsung PM9D3a PCIe Gen5 NVMe SSD* (supplementary: 8-channel controller, predecessor PM9A3) — https://www.opencompute.org/marketplace/it-infrastructure/samsung-pm9d3a-pcie-gen5-nvme-ssd/ — accessed 2026-11-15.
