SATA vs NVMe for Gaming: The 7,000 MB/s Marketing Lie

The Verdict: SSD Speed Doesn’t Move the Needle in Most Games

A drive on the old SATA interface turns in framerates that sit right alongside a modern NVMe SSD in most titles, which settles the question for anyone eyeing an upgrade purely for gaming: swapping SATA for NVMe rarely moves the average framerate at all. If a game isn’t asking for the bandwidth, having it sit unused doesn’t buy you anything. 

That pattern holds up the ladder too: the jump from one PCIe generation to the next produces gains too small to plan a purchase around, a point reviewers have landed on and readers watching the same benchmarks reached on their own.

The difference is negligible; you can easily play on PCIe 3 as well

None of this stops the box art from leading with sequential-speed numbers that have nothing to do with what a game session actually asks of a drive, and readers are catching on to the gap between the spec sheet and what they feel at the desk.

The fraud of the number that they keep “passing you” 😉

There’s a hardware reason storage interface stops mattering at the high end too: a flagship GPU loaded with VRAM keeps enough of the working set on-card that it rarely has to reach across the PCIe bus for anything, which is a variable that’s been missing from a lot of these arguments.

To be fair, the 5090 has so much VRAM that it rarely has to pull data over the PCIe bus, as compared to a lower end card.

A methodology gap also gets raised: pairing a PCIe 3.0 drive with DDR5 memory isn’t a rig anyone actually builds, since a board old enough to run PCIe 3.0 storage doesn’t pair with DDR5 in the first place, and that mismatch turns the comparison into a lab exercise rather than a real upgrade path.

But you don’t have DDR4 or DDR5 on PCI-E 3.0. The test is with DDR5 it’s just a “laboratory test” :D.

There’s a second methodology critique worth carrying forward into the game-by-game data: if you actually want to see a PCIe bottleneck show up, testing at 4K works against you, because the GPU-bound framerates there leave bandwidth headroom to spare. Push the resolution down to 1080p instead, where framerates climb high enough to demand more from the bus each second, and any real ceiling would show itself there first.

That’s the wrong way to test; if you want to test PCIe limitations, you test at 1080p because of the high frame rates.

Forspoken flips the script entirely: reviewers clocked eight to ten percent higher average framerates on a mechanical hard drive than on NVMe. That’s not the hard drive winning a bandwidth fight it has no business winning, it’s the game noticing how slow the drive is and quietly loading lower-resolution textures to keep up. Fewer texture bytes streaming in means less work for the GPU, and less work for the GPU means a higher framerate number on the chart, even though what’s actually on screen looks worse. Anyone chasing that number by buying a slower drive is trading fidelity for a stat that only looks like a win.

Content creation is where the calculus reverses. Video editing applications take real, measurable hits once you drop them onto slower slots, and DaVinci Resolve is the clearest example: performance falls twenty-five percent when it’s restricted to PCIe 4.0 x4. That’s not a rounding error hiding in a benchmark chart, it’s a quarter of your timeline scrubbing and export speed gone because the interface can’t move footage fast enough. Gamers can ignore Gen 5 with a clear conscience. Editors working off a single drive can’t.

Sequential Versus Random: Where the Advertised Numbers Fall Apart

The number on the box is a sequential speed, measured by shoving one giant continuous file through the drive as fast as it’ll go. It’s the figure manufacturers print because it’s the biggest one they’ve got. But loading a game, booting Windows, opening a level, none of that behaves like one giant file, it’s thousands of small scattered reads, and that’s random performance, which typically runs fifty to one hundred times slower than the sequential figure sitting on the packaging.

• Gen 4 NVMe: 7,000 MB/s sequential read, but only 70 to 80 MB/s random read, per OSCOO’s figures.
• SATA SSD: 550 MB/s sequential read, dropping to 40 to 50 MB/s random read, per OSCOO’s figures.
• Mechanical hard drive: random reads crawl to 0.5 to 1 MB/s, per OSCOO’s figures.

Loading a game level means pulling thousands of separate, scattered files off the drive, not one long stream, which is exactly the workload random speed measures and sequential speed doesn’t touch. That’s why a Gen 4 drive advertised at 7,000 MB/s and a SATA drive at 550 MB/s end up within shouting distance of each other once a level actually starts loading: both are working at their random speed, and there the gap between 70 to 80 MB/s and 40 to 50 MB/s is nowhere near the gulf printed on the box.

Forspoken: The First DirectStorage Game

Forspoken holds a title none of the others in this test suite can claim: it’s the first PC game built around DirectStorage 1.1, and the DLLs sit right there in the install folder to prove it. But having the API installed and using what it’s for are two different things. The version that shipped skips GPU decompression entirely and leans on the CPU instead, which is exactly why the game rarely comes close to pulling three gigabytes per second off the drive. The plumbing for a real next-gen storage pipeline is in the box. Nobody turned the tap.

Where Forspoken does draw a hard line is scene loading. PCIe 4 drives get a new scene up in one to two seconds, PCIe 3 drives trail by three seconds at most, and PCIe 5 generally sits ahead of both, though the gap between any of the three NVMe tiers is small enough that nobody’s picking a drive on it. A mechanical hard drive, meanwhile, takes almost a full minute to do the same job. That’s the one place in this game where the choice between NVMe and spinning rust isn’t academic.

Tom’s Hardware clocked peak bandwidth at 1080p at 485 MB/s on a SATA SSD against 2761 MB/s on PCIe 5.0, a gap of more than five and a half times on paper. That the game still rarely touches three gigabytes per second even on the fastest drive tells you which number actually governs the experience, and it isn’t the one on the spec sheet.

The average framerate might not move much, but the 1% lows do, and 1% lows are the number that actually tells you whether a game stutters. NVMe drives improve those lows over SATA across the board, and at 4K the gain reaches thirty-nine percent. That’s the real payoff of Forspoken’s storage work: not raw throughput, not a headline framerate, but fewer of the hitches that make a game feel worse than its average number suggests.

Dying Light 2: One Frame Separates Every PCIe Generation at 4K

At 4K, PCIe Gen3 x16 gives up roughly one percent versus the newer interfaces. That’s the entire cost of running a five-year-old motherboard slot in this game: a rounding error, not a reason to upgrade.

Gamers Nexus clocked 216.1 FPS on Gen5 x16 at 1440p against 215.5 on Gen4 and 215.4 on Gen3, the oldest interface trailing the newest by less than a single frame. Anyone buying a Gen5 board for this game is paying for a number on a spec sheet, not a number on their screen. 

Black Myth: Wukong: The Widest Gap in the Suite

Gamers Nexus clocked Black Myth: Wukong at 85.5 FPS on Gen5 at 4K, 84.7 FPS on Gen4, and 83.1 FPS on Gen3, a clean progression from oldest interface to newest that finally shows up in the numbers instead of getting lost in noise. Gen5 finishes three percent ahead of Gen3, which is the widest margin this article has shown yet and still nowhere near something a player would feel with a controller in hand.

At 1440p, Gamers Nexus put Gen5 at 129.9 FPS against 124.8 FPS on Gen3, a slightly wider gap than what showed up at 4K. But a Gen3 system old enough to still be running that slot is also carrying an older CPU, and at these framerates it’s the processor feeding the GPU, not the bandwidth between GPU and storage, that’s more likely setting the pace.

Drop to 1080p, where framerates run highest and any bottleneck should show up loudest, and Gamers Nexus still found Gen5 at only 159.8 FPS versus 153.3 FPS on Gen3. Even at the resolution most likely to expose an interface limit, the oldest slot tested keeps pace within a few percent of the newest one.

Turning on ray tracing doesn’t change the story. At 4K, Gamers Nexus found Gen4 and Gen5 turning in identical numbers, and at 1080p the gap opens to a four percent edge for Gen5. Ray tracing shifts load onto the GPU’s shaders and denoiser, not onto the storage interface, so it was never going to be where a PCIe generation proved its worth.

Resident Evil 4

GamersNexus put Resident Evil 4 at 4K at 206.7 FPS on Gen5 and 204.2 FPS on Gen3, an interface gap that’s already down to fractions of a percent at the top end of this suite. Whatever is limiting this game at 4K, it isn’t the PCIe slot.

Drop to 1080p and Gen4 and Gen5 land close enough to call it a tie, well inside the margin reviewers treat as noise. Gen4 does pull three percent ahead of Gen3 at that resolution, which is the only separation Resident Evil 4 offers up in this entire test, and it’s a gap you’d need a frame counter open to notice while playing.

F1 24

GamersNexus logged F1 24 at 4K running 291 FPS on Gen5 and 285 FPS on Gen3, a two percent edge for the newer interface despite Gen5 offering roughly four times the raw bandwidth of Gen3 on paper. That’s the gap between what a spec sheet promises and what a racing game actually asks a drive to do: nobody buying a Gen5 SSD to play F1 24 at 4K is getting anything a five-year-old interface wasn’t already delivering.

After Effects: Where Bandwidth Cuts Barely Register, Until They Do

Puget Systems ran the full spread from 64 GB/s down to 16 GB/s and came back with numbers that don’t separate from each other at all, which covers everything from a full Gen5 x16 link down to a Gen4 x4 slot. That’s a wide range of hardware configurations landing in the same bucket. The floor only shows up once you drop to PCIe 3.0 x8, where the results finally step outside the margin of error, and it gets worse from there: PCIe 3.0 x4 costs a full ten percent versus the higher-bandwidth configurations. Anybody running a modern Gen4 or Gen5 drive, even in a starved x4 or x8 slot on a recent board, isn’t losing anything here. It’s only the old x4-limited Gen3 setups, the kind you find on budget boards or squeezed out by a second GPU, that actually pay a price in this app.

DaVinci Resolve: The Application Where the Slot Finally Matters

Everything this suite has shown so far has been about games shrugging off restricted bandwidth. Resolve doesn’t shrug. A full Gen5 x16 link and a full Gen4 x16 link land in the same place, so nobody needs to chase the newer standard for its own sake. Drop to a 16 GB/s connection, though, and Puget Systems measured a ten percent penalty, the first real crack in this article’s pattern of PCIe indifference. Keep cutting bandwidth and the crack becomes a chasm: PCIe 3.0 x8 runs twenty-five percent slower than the full-bandwidth configurations, and PCIe 3.0 x4 delivers just fifty-four percent of what the fastest setup achieves. That’s not a rounding error like Dying Light 2’s one percent or Resident Evil 4’s fractional gap, that’s an editor losing nearly half their performance because a motherboard put the GPU on a narrow, old lane.

Unreal Engine

Two configurations sitting at opposite ends of the PCIe spec sheet, 5.0 x16 and 3.0 x16, land on the same number in Unreal Engine’s viewport, which means the interface generation itself buys nothing as long as you keep all sixteen lanes. What actually costs performance is cutting the lane count, not the generation: drop to 4.0 x4 and Puget Systems measured ninety-three percent of maximum speed, drop further to 3.0 x4 and that’s ninety percent. The full spread shows up at the extremes, 128.75 FPS GeoMean on 16x PCIe 5.0 against 116.58 FPS on 4x PCIe 3.0, a gap of about twelve frames that traces straight back to lane count rather than PCIe generation. Anyone running a full x16 slot, whatever generation it is, isn’t losing anything here; anyone stuck on a four-lane connection is the one who should be checking the number.

Blender: PCIe Bandwidth Doesn’t Touch the Render Time

Blender loads a scene onto the GPU once, then does the actual rendering entirely out of VRAM: there’s no repeated back-and-forth over the PCIe bus for the card to wait on. That’s why every configuration in this test lands within about five percent of the average, regardless of what interface is feeding the slot. If you’re buying a fast SSD or a Gen5 motherboard to speed up your Blender renders, this is the wrong line item to chase. 

V-Ray

V-Ray is the outlier in this whole suite, and not in the way you’d expect. A full 16x PCIe 5.0 link scores 10,613.00 in CUDA Mode – GPU, while a starved 4x PCIe 3.0 connection, a fraction of that bandwidth on paper, comes in at 10,405.20. That’s a gap of under two hundred points on a scale where the top and bottom of the entire lane-and-generation spread should, by spec sheet logic, be nowhere near each other. Every other render engine in this test either scales cleanly with bandwidth or ignores it completely once the workload sits in VRAM; V-Ray does neither, it just refuses to move much at all regardless of what’s feeding the card. Nobody buying a GPU renderer to run V-Ray needs to worry about their PCIe slot generation, the number barely notices it’s there.

OctaneBench: Another GPU Renderer, Another Non-Issue

Rendering workloads keep making the same point from a different angle, and OctaneBench is the cleanest version yet: every configuration tested lands within 2.5 percent of the average score. That’s a tighter spread than V-Ray showed, on a benchmark built specifically to hammer the GPU, which is exactly the kind of workload you’d expect to expose a starved PCIe link if one existed to expose.

Local AI Inference: VRAM Is the Only Number That Matters

Once a model is small enough to sit entirely in VRAM, Llama.cpp stops caring what feeds the GPU’s PCIe slot. Prompt processing and token generation numbers swing around without any pattern tied to lane count or generation, and the total spread across every configuration tested comes out to about six percent. A gap that size with no consistent direction isn’t a bandwidth ceiling, it’s noise: the kind of variance you’d see from run to run on the exact same hardware. If you’re buying storage or a motherboard with local AI workloads in mind, the PCIe slot isn’t where that decision gets made.

Alan Wake 2: The Heaviest Streamer in the Suite

Remedy built Alan Wake 2 around a fixed memory budget, which means the game never just loads a level and sits there: it’s constantly pulling new assets into that budget while shoving out whatever it doesn’t need anymore. That churn is exactly the kind of workload that should, in theory, put a real gap between a fast drive and a slow one.

Push the resolution up and the drive works harder to match it: Tom’s Hardware measured average bandwidth north of 200 MB/s at 4K, a sustained figure that’s genuinely high for a game rather than a synthetic benchmark.

That 5128 MB/s figure on PCIe 5.0 is nine times what SATA can hit, but a peak is a burst, not a sustained rate, and Alan Wake 2 only asks for that kind of throughput in short spikes rather than for seconds at a time. Framerates come out identical across every drive tested, PCIe 5.0 down to SATA, because a burst that fast finishes before it could ever become a bottleneck.

Ratchet and Clank: Rift Apart

Insomniac built the portal sequences around GPU decompression, which means the graphics card itself unpacks incoming assets instead of leaning on the CPU, and that only pays off if the drive behind it can actually keep the pipe full. It does: PCIe SSDs shave two to three seconds off every portal transition compared to SATA. That’s not a benchmark footnote, it’s time you can feel with a controller in your hand, since the whole point of those transitions is that they’re supposed to look instant.

Here’s the twist that makes average framerate a bad way to judge this game: at lower resolutions, SATA drives actually post slightly higher averages than PCIe drives. That’s not SATA winning anything, it’s SATA getting stuck longer inside the simpler, easy-to-render portal geometry while the faster drive has already loaded the heavier scene on the other side. A slow drive drags the average up by spending more time somewhere cheap to render.

Tom’s Hardware clocked peak bandwidth at 2160p at 532 MB/s on SATA against 2234 MB/s on PCIe 5.0, over four times the throughput, and that gap is exactly what a demanding portal moment goes looking for the instant the scene on the other side needs to be fully built before the camera lands. 

Step outside the portals and every drive in this test behaves the same, no measurable gap between SATA and PCIe in ordinary gameplay. Step through one at 4K and PCIe holds a 38 percent lead in one percent lows, which is the actual moment SATA drives stutter and PCIe drives don’t. That’s the whole verdict for this game in one line: buy NVMe if you care about the transitions, don’t bother if you don’t.

Spider-Man 2: GPU Decompression Without the Bandwidth to Match

Insomniac built Spider-Man 2 to unpack incoming assets on the GPU rather than the CPU, the same trick that made Rift Apart’s portals worth the SSD upgrade. Here it doesn’t matter: Tom’s Hardware measured a peak of 150 MB/s out of a SATA SSD at 1440p, and a PCIe 5.0 drive, capable of tens of gigabytes per second, peaked at 208 MB/s in the same test. That’s a drive rated for roughly a hundred times more bandwidth getting asked for barely more than the SATA drive gave up. 

None of the drives tested are put under any real load, and performance comes out virtually identical across every one of them. GPU decompression here is a feature with nothing to decompress fast enough to matter, and it buys the SATA drive owner nothing to worry about.

Starfield: Constant Streaming, Constant Bottleneck

Bethesda built Starfield around constant background reads as the player moves between areas, and that habit shows up as stuttering the moment a new zone loads. It’s not an occasional hiccup either: the drive frequently pegs itself at full utilization during these transitions, which is a different problem than not having enough bandwidth on tap.

Tom’s Hardware measured a SATA SSD peaking at 455 MB/s and a PCIe 5.0 drive peaking at 934 MB/s, better than double. On a spec sheet that looks like the kind of gap that should matter, but a PCIe 5.0 drive is capable of tens of thousands of megabytes per second, and getting nowhere near four figures points to low queue depths rather than a drive running out of headway. The practical result is that the faster NVMe drive buys nothing over SATA here.

Club386 pushed a Crucial T700, a Gen5 drive, to full utilization in Starfield, with peaks bouncing between 555 MB/s and 850 MB/s. Sitting at 100% utilization while peaking under a gigabyte a second already tells you the drive isn’t the bottleneck by bandwidth, and the average read speed during that stretch confirms it: around 20 MB/s, a fraction of even the peak figures. That’s the drive being fully busy doing almost nothing useful, and the GPU pays for it directly, with usage dropping to zero percent while the storage subsystem sorts itself out.

DOOM: The Dark Ages

id Software lists an NVMe SSD as a hard requirement for DOOM: The Dark Ages and built the engine around what it calls continuous sector streaming, the kind of language that implies a drive under constant load. The reality doesn’t match the pitch: this is one of the least demanding titles in the entire suite as far as the storage subsystem is concerned.

Tom’s Hardware measured a peak of just 192 MB/s out of a SATA SSD at 4K, against 287 MB/s on a PCIe 5.0 drive capable of moving tens of thousands of megabytes per second. The PCIe 5.0 number is higher, but both figures sit so far below what either interface can actually deliver that neither drive is working for a living here.

Every drive tested delivers identical gameplay, and the SATA SSD in particular produced no stutter, no texture pop-in, no visual artifacts of any kind. The NVMe requirement on the box is a gate id Software chose to put up, not a limit the game actually needs cleared.

Dead Space Remake: 100% Disk Utilization, Nowhere To Go

A drive pegged at full utilization looks like a bottleneck until you check what it’s actually being asked to move. Dead Space Remake issues its read requests at low queue depths, meaning it never has more than a handful of requests in flight at once, and a drive can show 100% busy while shuffling a trickle of data if it’s spending most of that time waiting on the next request rather than servicing a queue full of them. That’s a request pattern problem, not a bandwidth problem, and no amount of extra PCIe lanes fixes a queue that was never deep enough to fill them.

Whatever gap sits between 562 MB/s and 833 MB/s on paper, it doesn’t show up in the stuttering: every drive tested delivers the exact same performance in Dead Space Remake, SATA included. If you’re chasing a faster SSD to fix this game’s hitching, save the money.

Cyberpunk 2077 is next in the suite, and the answer is the same one this article keeps giving: whatever drive you put behind it, the game plays the same.

CD Projekt Red built this streaming system for a console generation that still shipped mechanical hard drives, and it shows: Cyberpunk 2077 pulls the lowest average bandwidth of any game in this entire test group. A streaming budget sized for a spinning platter was never going to notice whether the drive behind it is SATA or the fastest NVMe on the market, because it never asks for more than a spinning platter could give it in the first place.

Tom’s Hardware measured peak bandwidth at 4K and came back with numbers that make the point on their own: 121 MB/s on the SATA SSD, versus 111 MB/s on PCIe 5.0. The older, slower interface actually pulled more data at its peak than the drive with roughly a hundred times the bandwidth on paper, which only makes sense once you accept that neither drive is even close to being asked to work.

The Elder Scrolls IV: Oblivion Remastered

Nanite and Lumen are heavy asks for a GPU, chewing through geometry detail and lighting bounces that have nothing to do with the drive underneath them. None of the SSDs tested ever hit full peak utilization running this remaster, which means whatever load Unreal Engine 5’s fancier rendering puts on a system, it isn’t landing on storage.

Tom’s Hardware clocked the SATA SSD at 326 MB/s peak bandwidth at 1440p, ahead of the PCIe 5.0 drive’s 299 MB/s in the same test. A drive with a fraction of the interface bandwidth outrunning the fastest thing on the market only makes sense once you accept neither drive is close to being stressed: the SATA SSD delivers identical real-world performance to the NVMe options, Nanite and Lumen or not.

007 First Light: DirectStorage and Dynamic Loading Meet a Drive That Never Breaks a Sweat

IO Interactive built the Glacier engine’s streaming around DirectStorage and dynamic loading specifically to smooth out level transitions, but smoothing out transitions and taxing a drive are two different design goals, and this one only chases the first. Nothing in the pipeline asks a storage device to work for its living.

Tom’s Hardware measured a peak of 164 MB/s on a SATA SSD at 1080p against 171 MB/s on a PCIe 5.0 drive, a seven MB/s gap between an interface capable of hundreds of megabytes per second and one capable of tens of gigabytes. That’s not a bottleneck showing up faintly, it’s a bottleneck that never shows up at all: every SSD tested plays the game exactly the same, and anyone weighing a storage upgrade for 007 First Light specifically can put that money toward something else.

F1 25

Cars flashing past trackside cameras at 200 miles an hour need their surroundings streamed in fast, and F1 25 pulls more bandwidth doing it than most of the games in this suite. Tom’s Hardware clocked a peak of 554 MB/s out of a SATA SSD at 2160p, against 602 MB/s on PCIe 5.0, the smallest gap between the two interfaces this article has shown. That’s still nowhere close to what a SATA drive can choke on, so even F1 25’s heavier appetite leaves every drive in the test turning in identical performance.