Choosing the Right Storage

TL;DR

In sustained random write tests, DRAM-equipped drives maintain 50-70% higher IOPS after cache exhaustion compared to DRAM-less counterparts. For PostgreSQL workloads, that translates to 35-40% fewer operations per second and 7× worse P99 latency on DRAM-less drives. In this article well see why that DRAM cache matters, when you can skip it, and how to choose based on actual workload data.

---

Introduction: The Spec Sheet Lies

We’re in peak storage performance territory, PCIe Gen 5 NVMe drives pushing 14,500 MB/s sequential reads, Gen 4 flagships hitting 7,450 MB/s, and even budget NVMe drives breaking 5,000 MB/s. On paper, the storage bottleneck is dead. Except it’s not.

The dirty secret nobody puts on the spec sheet: those numbers only hold when the drive’s SLC cache isn’t full and its mapping tables are hot. Real-world sustained performance tells a different story, and the deciding factor isn’t PCIe generation or NAND type, it’s whether your SSD controller has dedicated DRAM cache.

In benchmark testing across consumer and enterprise workloads, DRAM-equipped drives showed 17-19% performance degradation after 60 seconds of sustained writes. DRAM-less drives? 57-60% performance drop. When PostgreSQL databases hammered these drives with INSERT-heavy workloads, DRAM-less drives delivered 12,000 operations per second versus 18,500 on DRAM equivalents, a 35% throughput loss. P99 latency spiked from 0.12 ms to 0.85 ms, making tail latencies 7× worse.

The question isn’t whether PCIe Gen 5 is faster than Gen 4. It’s whether you need DRAM cache to maintain that speed when it actually matters.

---

How DRAM Cache Works in SSDs

Here’s what DRAM doesn’t do in modern SSDs: it doesn’t buffer your user data. That’s what the SLC cache is for. DRAM’s job is far more critical, it holds the Flash Translation Layer (FTL) mapping table, the logical-to-physical address index that tells the controller where every block of data lives on the NAND chips.

The FTL Penalty Without DRAM

A typical 1TB SSD requires roughly 1GB of FTL metadata, a 1:1 ratio of 1MB metadata per 1GB NAND. Every read or write operation requires consulting this table. With DRAM onboard, that lookup happens in 0.01-0.1 milliseconds (DRAM access time). Without DRAM, the controller must read the FTL metadata from NAND flash itself, adding 1-3 milliseconds per operation.

Let’s visualize the data flow:

DRAM-Equipped SSD:
Host → Controller → DRAM (FTL lookup, 0.1ms) → NAND (data access)
Total latency: ~0.15ms for random 4K write

DRAM-less SSD (HMB or NAND-based FTL):
Host → Controller → NAND (FTL read, 2ms) → NAND (data write)
Total latency: ~2.5ms for random 4K write

Performance ratio: 16× slower on DRAM-less

This explains why benchmark data shows DRAM-less drives hitting 40,000-80,000 random 4K IOPS at QD1, while DRAM drives achieve 75,000-100,000 IOPS, a 20-55% gap. At saturated queue depths (QD32), the difference widens: DRAM drives sustain 800,000-950,000 IOPS, DRAM-less drops to 450,000-650,000 IOPS.

Host Memory Buffer: The Compromise

DRAM-less drives using Host Memory Buffer (HMB) technology borrow system RAM to cache FTL metadata. Sounds clever, right? The catch: HMB transactions traverse the PCIe bus, adding latency and consuming PCIe bandwidth. On older platforms (pre-Ryzen 3000, Intel 9th-gen and earlier), HMB support is flaky or absent, leaving DRAM-less drives to suffer full NAND-based FTL lookups.

Even with HMB working, measured latency shows the overhead: 6-13 μs extra at QD1, growing to 15-42 μs at QD32 as queue contention exposes the PCIe round-trip penalty.

---

The Sustained Performance Cliff: Where DRAM-less Drives Collapse

Sequential benchmarks are marketing candy, nobody copies 100GB files in a single burst. Real workloads mix reads, writes, small files, and large files. That’s where the SLC cache fills up and DRAM-less drives hit the wall.

Benchmark Data: Cache Exhaustion Performance

Testing methodology: 60+ seconds of sustained mixed random writes (4K-64K blocks), typical of database commits, VM snapshots, or photo library imports. Results measured in IOPS degradation and latency spikes:

|----------------------------|-----------------|------------------|

Metric	DRAM Drives	DRAM-less Drives
Performance drop after 60s	17-19%	57-60%
Latency during cache miss	0.015-0.018 ms	2.5-3.2 ms
Recovery time to peak	2-4 seconds	45-120 seconds
SLC cache duration	42-50 seconds	5-8 seconds

DRAM drives maintain performance because their controllers can quickly reorganize NAND blocks in the background, using instant FTL lookups to optimize garbage collection. DRAM-less drives serialize these operations, every GC cycle requires reading FTL metadata from NAND, creating cascading delays.

The fsync() of Databases

Database systems call fsync() after every transaction commit to ensure durability (ACID compliance). This forces the storage controller to flush write buffers and update metadata. Benchmark results from a 4-CPU server running PostgreSQL:

• DRAM NVMe drives: 285 fsync() operations/second, average latency 3.5 ms
• DRAM-less NVMe drives: 95 fsync() operations/second, average latency 10.5 ms

That 10.5 ms fsync() latency is equivalent to a 7200 RPM HDD (which averages 22 ms). The culprit: DRAM-less controllers must update NAND-based FTL tables on every flush, with no DRAM buffer to batch and coalesce writes.

---

Quantified Trade-offs: DRAM vs DRAM-less Economics

Let’s cut through the marketing and look at measurable advantages and costs.

DRAM-Equipped Drives: The Pros

Performance Consistency Under Load

• Sustained random writes: 800,000+ IOPS at QD32 (versus 450,000-650,000 for DRAM-less)
• QD1 latency: 75-100 μs (versus 100-200 μs)
• Post-cache performance retention: 81-83% (versus 40-43%)

Database Workload Performance (PostgreSQL benchmark on 4-CPU/64GB server)

• INSERT rate: 18,500 ops/sec (versus 12,000 on DRAM-less)
• SELECT rate: 31,000 ops/sec (versus 18,500)
• P95 latency: 0.08 ms (versus 0.35 ms)
• P99 latency: 0.12 ms (versus 0.85 ms, 7× worse on DRAM-less)

Endurance and Reliability

• Write amplification: 1.5-2.5× (versus 3-8× on DRAM-less)
• TBW ratings: Samsung 990 Pro 2TB = 1,200 TBW (versus 600 TBW on 980 DRAM-less equivalent)
• Annualized Failure Rate (AFR): 0.50-0.55% (versus 1.2-2.5% on consumer DRAM-less drives)

Lower write amplification directly extends drive lifespan. DRAM controllers batch writes and optimize garbage collection with instant metadata access, reducing unnecessary NAND wear.

DRAM-less Drives: The Advantages

Cost Per Gigabyte

• Typical premium: 18-20% cheaper (€0.070/GB versus €0.110/GB for DRAM equivalents)
• Budget SATA DRAM-less (Kingston A400): €0.052/GB, 35% cheaper than NVMe DRAM-less

Power Consumption

• Idle: 2.0-2.2 W (versus 2.5-3.2 W for DRAM drives)
• Active reads: 4.1 W (versus 5.2-5.8 W)
• Active writes: 5.2 W (versus 6.8-7.2 W)

The 15-25% power savings matter for laptops and mobile devices. For servers running 24/7 sustained workloads, the advantage disappears, DRAM-less drives stay active longer completing the same work, equalizing total energy consumption.

Simpler Bill of Materials

• Fewer components = lower manufacturing defect risk
• No DRAM chip compatibility issues
• Easier thermal management (one less heat source)

When the Trade-off Favors DRAM-less

If your workload is >80% sequential reads (media streaming, game installs, boot drives with light multitasking), the FTL lookup penalty barely registers. HMB-enabled DRAM-less drives on modern platforms (Ryzen 5000+, Intel 12th-gen+) perform within 10-15% of DRAM equivalents for burst workloads under 30 seconds.

For embedded systems or edge devices where power budget is critical (industrial IoT, surveillance NVRs, portable devices), the 1-2W savings justify the performance compromise.

---

Use Cases: Matching Workload to Storage Architecture

Database Servers (MySQL, PostgreSQL, MongoDB)

PostgreSQL INSERT performance dropped 35% on DRAM-less drives (18,500 ops/sec → 12,000). SELECT queries fell 40% (31,000 → 18,500). More critically, tail latency exploded: P99 latency went from 0.12 ms to 0.85 ms, the difference between imperceptible lag and user-noticeable delays.

SQL Server benchmarks from enterprise deployments show similar patterns: moving from SATA SSDs to high-performance NVMe yielded 35% higher throughput, but that gain evaporated with DRAM-less drives due to FTL thrashing during heavy transaction logging.

Recommendation: Samsung 990 Pro, WD Black SN850X, or equivalent DRAM-equipped NVMe. The €30-40 premium over DRAM-less pays for itself in the first month of production load.

AI/ML Training Workloads

Training datasets involve billions of small random reads (image patches, tokenized text, feature vectors). This is FTL lookup hell for DRAM-less drives. A typical ImageNet training run reads 1.3 million images across 1,000 classes, millions of random 4K-64K reads per epoch.

Benchmark data: DRAM drives deliver 75,000-100,000 random 4K read IOPS at QD1, versus 40,000-80,000 on DRAM-less. Over a 24-hour training run, that’s the difference between 18 epochs and 12 epochs, 33% slower model convergence.

For production ML pipelines: don’t even consider DRAM-less. The time wasted waiting on storage dwarfs the €40 you save per drive.

OS/Boot Drives and Casual Gaming

Booting Windows or Linux is primarily sequential reads. Game installs and level loads are also sequential-heavy. Benchmark data from Tom’s Hardware: in Final Fantasy XIV Stormblood benchmark, NVMe drives (both DRAM and DRAM-less) showed significantly shorter load times than SATA SSDs or HDDs, with minimal difference between DRAM and DRAM-less NVMe in this specific test.

For a gaming rig that boots once per day and loads games occasionally, a €70 Samsung 980 (DRAM-less) performs within 5-10% of a €110 990 Pro for actual user-perceptible tasks.

Virtualization Hosts and NAS Systems

VMs generate horrific random I/O patterns: multiple OS instances writing logs, cache files, temp data, all interleaved. A single VM snapshot operation on a DRAM-less drive can stall the entire host due to FTL metadata contention.

NAS systems running RAID5/6 with parity calculations need consistent low-latency writes. DRAM drives maintain 285 fsync()/second under load; DRAM-less drops to 95 fsync()/second.

---

Reliability: Backblaze Data

Backblaze’s Q3 2025 SSD reliability report (covering hundreds of thousands of drives) shows:

• Enterprise DRAM-equipped SSDs: 0.40-0.72% AFR
• Consumer DRAM drives (990 Pro, SN850X): 0.50-0.55% AFR
• Consumer DRAM-less (980, A400): 1.2-2.5% AFR

DRAM-less drives fail 2.2-4.5× more frequently. Why? Write amplification. Without DRAM to optimize garbage collection, DRAM-less controllers write more data to NAND for the same user workload, accelerating wear.

A 1TB DRAM drive rated for 1,200 TBW will last 5+ years under heavy use (50GB writes/day). A DRAM-less drive rated for 600 TBW reaches EOL in 2.5-3 years under identical load.

The PCIe Gen 5 Amplification Effect

Here’s the kicker: Gen 5 speeds make DRAM cache even more critical. A Crucial T705 pushing 14,500 MB/s sequential reads can saturate that bandwidth only if the FTL lookups keep pace. At those speeds, even a 1ms FTL penalty creates massive bubbles in the data pipeline.

Benchmark data shows Gen 4 DRAM drives already hitting PCIe 4.0 x4 bandwidth limits (~7,000 MB/s) in large sequential copies. Gen 5 doubles that ceiling, but only if the controller can feed data fast enough. DRAM cache becomes the bottleneck arbiter.

The HMB Compatibility Trap

DRAM-less drives relying on HMB (Host Memory Buffer) work great, when HMB is supported and enabled. On older platforms or certain BIOS configurations, HMB fails silently, leaving the drive to rely on NAND-based FTL lookups.

I’ve seen production servers running Ryzen 2000-series CPUs where HMB never initialized properly. Those DRAM-less NVMe drives performed worse than SATA SSDs because they were constantly thrashing NAND metadata. Upgrading to DRAM-equipped drives fixed it instantly.

If you’re building a system that will run for 3-5 years, betting on HMB compatibility is a gamble.

---

Practical Recommendations by Budget and Workload

High-Performance Tier (€200-280 for 2TB)

Target: Database servers, VM hosts, ML training, pro video editing

• Samsung 990 Pro 2TB (€220-225): Top-tier DRAM, 1,200 TBW, proven reliability
• WD Black SN850X 2TB: Comparable performance, slightly better gaming optimizations

Mid-Range Tier (€140-180 for 2TB)

Target: Development workstations, gaming with heavy mods, content creation

• WD Blue SN5000 2TB (€144-160): DRAM-less but decent HMB implementation, acceptable for mixed workloads
• Crucial P5 Plus 2TB: DRAM-equipped, lower-tier controller but solid performance

Budget Tier (€80-120 for 1TB)

Target: Boot drives, casual gaming, office productivity

• Samsung 980 1TB (€70): DRAM-less HMB, good enough for burst workloads
• Kingston NV2 1TB (€42-80): Cheapest NVMe option, fine for non-critical use

Archival/Backup Tier

Target: Media libraries, cold storage, network backups

• Kingston A400 960GB SATA (€50): €0.052/GB, acceptable for read-heavy archival
• Seagate BarraCuda HDD 8TB (€155): €0.019/GB for bulk cold storage

---