SSD vs HDD vs i-RAM: What Changed Since the Benchmarks

There is a particular pleasure in reading an old benchmark review, because you get to see which arguments held up. This one compared two early Samsung solid state drives, a Gigabyte i-RAM RAM-disk card and three of the best hard drives available, at the moment when nobody yet knew which storage technology would win.

The full original test data is preserved below as a record of that moment. Around it is what actually happened next: how NVMe and PCIe rewrote the performance picture, what endurance ratings mean now, why DRAM-less drives stopped being a warning sign, and why i-RAM turned out to be a dead end rather than a preview.

How to read this article. The benchmark charts and the drives in them are historical, from the late-2000s transition period. They are kept because the shape of the results still explains why SSDs behave as they do. Every claim about current hardware is in the modern sections and sourced separately — do not take the old numbers as a guide to buying anything today.

How SSDs and Hard Drives Actually Differ

A hard drive stores data on magnetic platters spinning at high speed, read and written by a block of heads that a microcontroller moves across the surface. Transfer rate depends on rotation speed and areal density, and the interface has never been the bottleneck for internal drives.

The cost is in movement. Data is scattered across the platters, so a random or mixed workload spends most of its time repositioning the heads. Firmware fights back with request reordering and read-ahead caching, which is why a hard drive’s real-world behavior depends so heavily on its cache algorithms. The rattle you hear under load is the heads shuttling between zones.

An SSD stores data in flash memory chips governed by a controller. Flash is a form of EEPROM built from NAND cells, and its defining asymmetry is that reading is fast — find the cell, read it — while writing is slow, because a cell must be erased before new data can go in. Multi-level cells make this worse: writing means reading the whole cell, modifying, erasing and rewriting.

The advantages follow directly. Access time for reads is very low; random read speed is therefore close to sequential read speed, because reaching any cell costs about the same; and with no moving parts there is no noise, no vibration sensitivity and no seek penalty at all.

The disadvantages listed at the time were slower sequential transfer than a hard drive and a limited number of write cycles per cell. The first was solved comprehensively. The second was real, and is now well quantified rather than feared.

What Changed: NVMe, PCIe Generations and Real Speeds

The single largest change is that consumer SSDs stopped pretending to be hard drives. SATA and the AHCI command set were designed around a mechanical device with one queue and high latency. NVMe replaced that with a protocol built for flash, running over PCIe lanes directly and supporting deep parallel queues.

Each PCIe generation has roughly doubled the ceiling. A modern x4 drive on PCIe 5.0 reaches peak bandwidth of up to 14 GB/s. The SATA drives in the archived test below topped out around 100 MB/s.

PCIe 5.0 will stay the consumer standard for a long time. Silicon Motion’s CEO told Tom’s Hardware that consumer PCIe 6.0 solutions will not appear until 2030, citing signal-integrity problems that shrink usable trace lengths, cost, and the absence of interoperability testing.

Which generation you actually need

Sequential bandwidth is the number on the box and the least relevant to how a laptop feels. Booting, launching applications and general use are dominated by small random reads at shallow queue depths, where the gap between PCIe 4.0 and PCIe 5.0 is small and the gap between either and SATA is enormous.

PCIe 5.0 drives also run hot enough to need substantial heatsinks, which is a real constraint in a thin laptop chassis. For most machines a good PCIe 4.0 drive is the sensible choice, and the money saved is better spent on capacity — particularly now, given how sharply RAM and SSD prices rose through 2026 as AI data-center demand absorbed NAND and DRAM production.

If you are replacing a mechanical drive in an older machine, even a SATA SSD transforms it, and the practical steps are covered in our walkthrough of replacing a laptop hard drive with an SSD.

DRAM-Cached Versus DRAM-Less Drives

An SSD controller keeps a mapping table translating logical addresses to physical flash locations. Traditionally that table lived in a dedicated DRAM chip on the drive. Removing it saves money, and for years “DRAM-less” was shorthand for a drive to avoid.

Host Memory Buffer changed that. HMB lets the drive borrow a small slice of the computer’s own RAM to hold the mapping table instead, which PCWorld describes as using the DRAM already installed in your computer as the cache rather than paying for dedicated memory on the drive.

Modern HMB drives are genuinely good. PCWorld notes they now deliver real-world sequential performance equivalent to the fastest PCIe 4.0 and even 5.0 drives, while random performance still trails onboard-DRAM designs.

The gap shows up in sustained heavy writes, in random-access work, and in situations where HMB is unavailable — PCWorld notes the PlayStation 5 does not support it. For an everyday laptop, PCWorld’s judgment is that the average user saves money while still getting the vast majority of the performance.

DRAM-less is not automatically fine, though. An HMB drive relies on the host supporting the feature and on the operating system granting the buffer. In an external USB enclosure, in a games console, or behind an adapter, HMB may simply not be available — and the drive then falls back to reading its mapping table from flash, which is markedly slower. Buy a DRAM-cached drive for anything that will live outside a normal PCIe slot.

Endurance, TLC, QLC and What TBW Really Tells You

The original review put flash endurance at around 100,000 rewrite cycles per cell, which was accurate for the single-level cell flash of that era. It is nowhere near accurate for the multi-level flash every consumer drive uses now, because packing more bits into each cell costs endurance directly.

The modern consumer choice is between TLC, storing three bits per cell, and QLC, storing four. A drive endurance guide summarizing the trade-off puts TLC at around 600 TBW per terabyte with better sustained performance, against 200 to 400 TBW per terabyte for QLC.

TBW — terabytes written — is the manufacturer’s endurance rating, and it is widely misread as a lifespan. The same guide is blunt about it: the rating marks where the warranty ends, not where the drive dies, and endurance testing has repeatedly shown consumer drives surviving several times their rated TBW.

Wear is not what kills most consumer SSDs. Typical desktop users write somewhere around 5 to 15TB a year, so a drive rated at 600 TBW lasts decades at that rate — and most drives several years old report only 3 to 8 percent of their wear consumed. Endurance matters for video editing, databases and heavy caching. For everything else, capacity and price are the real decision.

You can check your own drive rather than guessing. Every modern SSD reports total host writes and a wear percentage through SMART, readable with CrystalDiskInfo on Windows or smartctl elsewhere. That is also a useful step when a machine feels slow, alongside the other checks in our guide to running diagnostics on a laptop.

SLC caching and the write cliff

Almost every consumer drive operates part of its flash in fast single-bit mode as a write cache, then folds that data down into TLC or QLC in the background. This is why advertised write speeds look so high, and why they are only true for a burst.

Write more than the cache holds — a large video file, a game install, a full disk image — and speed drops to the drive’s native rate, which on a QLC drive can fall below a hard drive’s. The cache also shrinks as the drive fills, because it is carved out of free space, which is the real reason a nearly full SSD feels slow.

Keeping meaningful free space is therefore not superstition. It preserves the write cache and gives the controller room for garbage collection and wear leveling. Note that defragmenting is the opposite of helpful here — the reasoning is in our explanation of why defragmenting a laptop works differently on an SSD.

The Archived Comparison: Test Participants

Everything from here to the end of the results is the original comparison, preserved as published. The drives are long obsolete; the patterns in the charts are not.

The SSDs were compared against three of the fastest hard drives then available: a Hitachi 7K200 (2.5-inch, 7200rpm, 16MB buffer, SATA, 200GB), a Samsung SpinPoint F1 (3.5-inch, 7200rpm, 32MB buffer, SATA, 1TB), and a Fujitsu MBA3300RC (3.5-inch, 15,000rpm, 16MB buffer, SAS, 300GB).

Samsung SSDs: 32GB and 64GB

Both were built in the 2.5-inch hard drive form factor. The MCAQE32G5APP-0XA held 32GB over a UDMA66 interface.

The MCBQE64GBMPP-03A held 64GB and used SATA, in an aluminum case.

Both were identified by the operating system as ordinary hard drives, which is exactly why the SATA-and-AHCI approach persisted as long as it did — and exactly what NVMe later abandoned in order to go faster.

Gigabyte i-RAM

The concept is obvious from the photographs. Four DDR SDRAM slots on a PCI card, presented to the system as a SATA drive. RAM is enormously faster than either flash or magnetic storage, and the card drew power from the PCI bus while running.

SDRAM is volatile, so an onboard battery held the contents when the machine was off. The manufacturer claimed 16 hours — a claim nobody sensible tested, because everything on the card vanishes the moment the battery runs out. Anyone storing real data on one was advised to run a UPS.

The limitations were severe. It took only DDR, not the cheaper DDR2 of the day; four slots capped it at 4GB; and it used first-generation SATA at 150MB/s, which throttled a device that was otherwise vastly faster than its interface. Memory speed was irrelevant, because even the slowest module outran the port.

Archived Testbed and Methods

The utilities used were IOMeter 2003.02.15, WinBench 99 2.0, FC-Test 1.0, PCMark 2004, PCMark 2005 and PCMark Vantage.

• Albatron PX865PE Pro II motherboard
• Intel Pentium 4 2.4GHz
• IBM DTLA-307015 system disk, 15GB
• Radeon 7000 32MB graphics
• 256MB DDR2-533 SDRAM
• Microsoft Windows XP Professional SP2

Generic operating system drivers were used, with drives formatted as a single partition in FAT32 and NTFS at default cluster size, plus 32GB partitions for some tests. SATA came from a Promise SATA300 TX4302 controller and SAS from an LSI SAS3041E-R.

Two things about this configuration are worth noticing from a modern vantage point. There is no TRIM — the command did not exist yet in Windows XP — so the drives had no way to learn which blocks were free, and every write eventually became a read-modify-erase-write cycle. And 256MB of system RAM would make Host Memory Buffer unthinkable.

Archived Results: Intel IOMeter

Sequential read and write patterns

IOMeter sent a stream of read and write requests at queue depth 4, changing the block size each minute to show how sequential speed depends on request size.

The SSDs could not match the hard drives sequentially. They competed on small blocks and fell behind on large ones, with even the 2.5-inch Hitachi reaching a higher maximum — putting these SSDs roughly on a par with 5400rpm laptop drives for sequential reading.

Oddly, the 32GB drive on the slower UATA66 interface beat the 64GB SATA model, suggesting the larger drive used slower chips. The i-RAM trailed only the 15,000rpm drive, and only because its 150MB/s SATA port was the constraint.

Sequential writing repeated the pattern, though the 64GB model closed the gap and the 32GB model beat the Hitachi on small blocks and the SpinPoint on 512-byte blocks.

Disk response time

Ten minutes of 512-byte requests at queue depth 1, more than 60,000 operations, giving a sustained response time independent of buffer size.

Only the RAM-based i-RAM beat the SSDs on read access time. The 15,000rpm SAS drive — the fastest mechanical storage money could buy — was about thirty times slower, and ordinary 7200rpm drives slower still.

Writes went the other way. Hard drives were roughly twice as fast at writing as reading, because writes landed in cache and were committed to the platters in optimized order. The SSDs wrote directly to flash with no buffer at all, which is precisely the deficiency that DRAM caches and SLC write caches were later invented to solve.

Random read and write patterns

Two charts were needed for each test because the i-RAM’s lead was large enough to compress every other line into one. This is where the SSDs separated from the hard drives decisively, leading at every block size except the largest — and small blocks are what real applications generate.

Database pattern

Random 8KB blocks with the read/write ratio swept from 0 to 100 percent and queue depths from 1 to 256. Results are shown at queue depths of 1, 16 and 256.

The SSDs led on pure reading, then collapsed as soon as even 10 percent of the queue was writes. As the write share grew the 15,000rpm drive took the lead and the other hard drives followed, because deferred writing made them faster with more writes while the SSDs only got slower.

At queue depth 16 the Fujitsu overtook the SSDs at only 10 percent writes. The 32GB model was indifferent to queue depth, having nothing to optimize, while the 64GB model got slower on reads — a controller or driver problem rather than a flash one.

That last chart is the single most instructive thing in the whole archive. An early SSD controller that fell apart under a deep queue is the exact opposite of what flash should do, and fixing it is most of the story of the following decade. NVMe exists to exploit parallelism that this controller could not handle at all.

Multi-threaded read and write patterns

One to four simultaneous clients, outgoing queue depth varied from 1 to 8. The queue-depth-1 charts are the most illustrative.

With one thread the SSDs were slow, held back by sequential read speed. Adding a second thread cost the hard drives about a third of their throughput and cost the SSDs and i-RAM nothing at all, because there are no heads to move between zones.

The insensitivity to thread count is the property that mattered long-term. It is why an SSD makes a busy machine feel responsive even when its headline transfer rate is unremarkable, and it is why the metric that eventually replaced megabytes per second was IOPS.

Web-server, file-server and workstation patterns

Server and workstation load simulations, with results given as performance ratings. The Workstation pattern was run at full capacity and on a 32GB partition, with the queue capped at 32.

The SSDs beat the Samsung and Hitachi drives at shallow queues but were overhauled as the queue lengthened, with the 64GB model ending as the slowest device present. High IOPS at shallow depth still earned both a better overall rating than the 7200rpm drives.

The web-server pattern contains no writes, and the SSDs were transformed by that. The Fujitsu could not stay with even the slower 64GB model, and the 32GB drive was far ahead. Read-dominated server work was the first commercial niche flash genuinely owned.

Archived Results: WinBench 99

The apparent slowdown at the end of the 64GB SSD is a test program artifact — every cell has identical speed. The important result is the flat line itself: SSDs and the i-RAM held the same speed across the entire capacity, while hard drives slowed towards the inner tracks where the platter surface moves past the head more slowly.

That flatness is why the concept of a “fast partition” at the start of a disk disappeared along with mechanical drives, and why the short-stroking trick used on servers became pointless.

Archived Results: FC-Test

FC-Test creates, reads and copies real file sets, so its results sit closer to application behavior than pure synthetics. Two 32GB partitions were used, formatted NTFS then FAT32. The i-RAM’s 4GB capacity allowed only read and write testing, not the copy tests.

The Windows and Programs sets consist of many small files; ISO, MP3 and Install use a few large ones.

NTFS file system

The SSDs’ read speeds barely varied with file size, while the hard drives’ varied enormously — the Hitachi much faster on the ISO set and slower on Programs. Insensitivity to file size, rather than raw throughput, is the flash characteristic that made the difference in daily use.

FAT32 file system

Archived Results: PCMark 2004

PCMark04 replays four traced workloads — Windows XP startup, application loading, file copying and general disk usage — and combines them into an overall rating. Each test was run ten times and averaged.

The split between the synthetic results and these trace-based ones is the most important finding in the archive. The SSDs lost most of the synthetic tests and won the tests that replayed real usage, which is a lasting argument against judging storage by a single sequential number.

Archived Results: PCMark 2005

PCMark05 replaces the File Copying trace with a File Write trace and adds a Virus Scan trace. Again ten runs, averaged.

Archived Results: Power Consumption

Measured in idle, random read/write and sequential read/write. The i-RAM could not be measured because of how it draws power from the PCI bus, so the 64GB SSD was compared against two 7200rpm hard drives.

At idle the SSD used only slightly less than the 2.5-inch Hitachi, and slower-spinning laptop drives of the era used less still — so an SSD was not automatically the more economical choice. The 3.5-inch SpinPoint was far hungrier than either.

The asymmetry is neat: hard drives draw more power reading, SSDs more writing, because every flash write is preceded by an erase.

The i-RAM as a Historical Curiosity

The i-RAM won essentially every test in the archive, and it is completely gone. That is worth understanding, because it is a good example of a technically superior product losing for reasons that had nothing to do with performance.

Its problems were structural. Four gigabytes was too small to hold anything but a page file or a scratch disk. The data was volatile, protected only by a battery. It occupied a full-length expansion slot, ruling out servers and every laptop. It used unbuffered DDR with no error correction, which disqualified it from anything where a bit flip mattered. And its SATA 150MB/s port throttled a device whose whole appeal was speed.

What killed the category was not a better RAM disk but the removal of the reason to want one. System memory grew from the 256MB in that testbed to tens of gigabytes, so an operating system could simply cache aggressively in RAM it already had, with no card, no battery and no interface bottleneck. Software RAM disks make the same idea free.

The serious attempt at the middle ground between RAM and flash was Intel’s Optane, built on 3D XPoint, which offered latency far below NAND with persistence RAM does not have. It failed commercially anyway: Intel announced the wind-down in August 2022 alongside a $559 million inventory write-off.

Do not go looking for one. An i-RAM card is a collector’s item, not a usable device. It needs a PCI slot no current motherboard provides, DDR1 modules that are scarce and expensive, and a battery that has been dead for over a decade. Any modern NVMe drive beats it on every axis that matters, with a thousand times the capacity and no risk of the contents evaporating.

What the Original Conclusions Got Right and Wrong

The right calls were about physics. Flash would dominate anything read-heavy and anything needing vibration tolerance; hard drives would keep the price-per-gigabyte and sequential-throughput crown for a long time; and synthetic benchmarks alone were a poor guide, since the SSDs lost most of them and won the traced workloads.

The wish list at the end was almost prophetic. It asked for larger, faster SSDs with buffer memory and multi-channel access to the flash. Multi-channel controllers and onboard DRAM buffers are exactly what arrived, and they are exactly what fixed the write-performance collapse visible in the database charts.

The wrong calls were about limits. Sequential speed was treated as a durable weakness of flash — it lasted only until controllers addressed many chips in parallel and the interface moved off SATA. The prediction that a 64GB SSD could reasonably cost about $250 now reads as touchingly conservative.

The largest miss was the conclusion that hard drives remained the best combination of price, capacity and speed for a workstation. That was true when written and stopped being true within a few years, once flash prices fell far enough that the responsiveness advantage outweighed everything else for the system drive. The pattern in the PCMark traces was already pointing there.

Hard drives did not disappear — they retreated to bulk storage, where cost per terabyte still wins. But the notion that a fast machine might reasonably boot from a spinning disk did disappear completely.

Final Thoughts on Flash Storage and Its Dead Ends

The archived charts are useful precisely because those SSDs were bad at so many things. Every weakness they show — no write buffer, no TRIM, a controller that broke under load — names a feature that a modern drive has, and explains why it exists.

The i-RAM is the other lesson. It beat everything in the room and still lost, because capacity, form factor and durability decide what survives at least as often as raw speed does.

Keep exploring

Related Posts

Blog

Hackintosh Is Officially Dead: macOS Golden Gate Explained

Blog

What Is a Copilot+ PC? (And Do You Actually Need One)

Blog

Why Are Razer Laptops So Expensive? The Real Breakdown