A Gen5 SSD Loads Games Faster

A PCIe 5.0 drive does read faster than a PCIe 4.0 one, but a game load uses far less of that speed than the box suggests. Across my bench archive, PCIe 4.0 drives read 99% faster than PCIe 3.0 drives at the median, yet their replayed game-load traces moved only about 30% more data per second. I’ve got retail Gen5 drives too, though only write tests on them, so that half of the answer leans on those write tests, on pre-release samples, and on other people’s published tests.

Where the myth comes from

Each interface generation raises the top speed a drive can print. Tom’s Hardware put the raw throughput of a PCIe 5.0 x4 link at about four times that of a PCIe 3.0 x4 link, and that ratio is what ends up on a box. Makers print sequential read, and a game load doesn’t read that way.

Tom’s Hardware’s 11-game test explains why. Games mostly issue random reads in 32KB or 64KB requests at a low queue depthIOPS counts commands finished per second. Queue depth is how many commands are outstanding at once, and a drive needs a deeper queue to reach its top IOPS., while sequential speed needs deep queues and large blocks. Advertised speeds are sequential reads, so a game can keep a drive busy while moving much less data than that. Solidigm’s traces found an Elden Ring level load overwhelmingly sequential at a most common transfer size of 128KB, as higher sequential speed means a faster PC covers, so the pattern changes from game to game.

The left panel of Figure 1 is what a box measures, and the right panel is what Tom’s Hardware found most games ask for. A drive tuned for the left panel can still sit idle between the small, scattered requests on the right.

What my bench shows

My first archive, which I call bench set A, holds CrystalDiskMark and 3DMark Storage results for 20 retail PCIe 3.0 drives and 64 retail PCIe 4.0 drives. CrystalDiskMark gives sequential read of 1MiB blocks at a queue depth of 8 on one thread. It comes nearest to the figure on a box. 3DMark Storage stresses the drive with gaming activities, loading games among them, and reports bandwidth in MB/s for loading traces from Battlefield V, Call of Duty Black Ops 4, and Overwatch. Each CrystalDiskMark figure is the median of repeat runs and each 3DMark Storage figure the median of three runs; the tool versions and the exact test platform were not recorded with the data.

Figure 2 compares the medians. Sequentially, the median PCIe 3.0 drive reads 3,514MB/s and the median PCIe 4.0 drive 6,979MB/s, a gain of 99%. Those same two medians reach 723MB/s and 942MB/s on the Battlefield V loading trace, a gain of 30%. Call of Duty Black Ops 4 went from 614MB/s to 807MB/s, 31%, and Overwatch from 334MB/s to 436MB/s, also 31%. So on Battlefield V, the median PCIe 4.0 drive moves just 13% of its own sequential read speed.

Figure 3 plots every drive. Sequential read speed and the Battlefield V trace move together, loosely, with a correlation of 0.62 across the 84 drives. Among the 39 PCIe 4.0 drives that read at least 6,500MB/s sequentially, the trace ran from 735MB/s to 1,480MB/s. That is a twofold spread for nearly the same box number.

Set A’s only PCIe 5.0 drive is a pre-release engineering sample, so I’ve left it unnamed and drawn it as a single square. It reads 10,076MB/s sequentially, 44% above the PCIe 4.0 median. Its Battlefield V trace reaches 1,491MB/s, 58% above that median yet only 1% above the best retail PCIe 4.0 drive. A Gen5 drive can beat a typical PCIe 4.0 drive, then, and a good PCIe 4.0 drive already gets to about the same place.

Treat this as what it is. Each trace replays the drive’s side of a load. Those are the requests the game made, reported as throughput in MB/s. The trace leaves out the game’s own processing around those requests. Nothing in either of my archives is a measured load time, and the PCIe 5.0 result is one sample from one drive.

A second set of drives, with retail PCIe 5.0

My bench also holds set B, which has what set A lacked. It has three retail Gen5 drives at 1TB, 2TB, and 4TB, with same-capacity PCIe 4.0 drives to set them against. I keep the two sets apart, since they were measured separately, and no number below mixes them.

Those three drives have one kind of result. A write-saturation run is a long sequential write that fills the drive’s cache and keeps going. The write tool was not recorded with the data. There are no game traces or load times for them, so this section shows drive behavior and leaves game loading aside. They’re set against nine retail PCIe 4.0 drives, three at each capacity, all with DRAM and all from one controller family.

In the fastest stretch of the run’s first minute, the PCIe 5.0 drives wrote 1.56 to 1.83 times as fast as the PCIe 4.0 drives of the same capacity. Cache size matched at about 113GB for 1TB, 224GB for 2TB, and 446GB to 450GB for 4TB, within 1% across all nine pairings. After the cache filled, the lead shrank to 1.02 to 1.48 times, averaged over the rest of each run. At 4TB the PCIe 5.0 drive averaged 3,262MB/s after the cache, against 2,833MB/s to 3,185MB/s for the three PCIe 4.0 drives.

Figure 4 gives the range for each capacity. Each Gen5 drive differs from its PCIe 4.0 rivals in controller and flash as well as interface, so the gap isn’t the interface alone. Run lengths also varied, and the after-cache figure averages the rest of each run. Set B also has one DRAM-less PCIe 4.0 drive at 1TB, which had a 54GB cache and averaged 757MB/s after it, so the 1TB PCIe 5.0 drive led it by 2.96 times after the cache. I’ve left it out of the ranges because it’s a different design, and it shows that a drive’s design sets the floor more than its interface does.

Pre-release PCIe 5.0 samples sit in their own bucket in set B, and three of them have 3DMark Storage results. Their Battlefield V loading traces reach 783MB/s, 1,491MB/s, and 1,604MB/s, against 1,104MB/s for the 2TB retail PCIe 4.0 drive, which is 0.71, 1.35, and 1.45 times. The 1,491MB/s sample has the same results as the pre-release sample in set A, so I count it once. In PCMarkPCMark is UL’s benchmark of whole system performance, built from real-world apps and activities. Its storage benchmarks take a trace-based approach. 10’s storage test the three samples score 2,528, 4,915, and 5,202, against 3,296 for that PCIe 4.0 drive. At best a sample scores 1.58 times the PCIe 4.0 drive and at worst 0.77 times, a spread among the samples larger than the gap to PCIe 4.0.

Put together, Gen5 is clearly faster than PCIe 4.0 in the first seconds of a write, and faster by much less once the cache is full. Two pre-release samples beat PCIe 4.0 on a game-load trace, by 1.35 to 1.45 times, and a third was slower. Those are replayed traces measured in MB/s, and I still don’t have a measured game load time for any drive in either set.

What a game load is made of

Microsoft’s DirectX team has described the rest of the load. Game assets are compressed to keep the package small. When a game runs, they’re read into system memory, where the CPU decompresses them before they’re copied to the GPU, and in Microsoft’s account that transfer and decompression contribute heavily to load times.

Older storage paths also cost time per request. Previous-generation games streamed assets at about 50MB/s, according to Microsoft’s 2020 post, and filling a 2.4GB/s drive with 64KB requests takes more than 35,000 requests a second. Each request carries a small overhead in older storage APIs, harmless on hard drives, and at that rate it adds up and keeps games from using NVMe bandwidth.

DirectStorageDirectStorage is a Windows API that lets games load data from storage with less CPU overhead. Later versions add GPU decompression, so the graphics card unpacks game assets. is Microsoft’s answer, and games using DirectStorage 1.0 on an NVMe drive should expect load times to fall by up to 40% by Microsoft’s figure. That gain comes from a new storage API on the same drive. Version 1.1 added GPU decompression, with a Microsoft sample to show it.

Figure 5 shows the sample. It is an early preview whose numbers vary with workload and hardware. Its scene loaded in 0.8 seconds with GPU decompression and 2.36 seconds with CPU decompression, nearly 3 times faster, with the CPU largely freed. GPU decompression applies only when the destination is a GPU resource; requests that go to system memory are decompressed on CPU threads. NVIDIA’s own version, RTX IO, offloads what NVIDIA calls dozens of CPU cores’ worth of work. That is a vendor claim rather than a measurement.

What the press tests found

Two Tom’s Hardware pieces report load times rather than traces, and both are secondary sources. In 2023 it reported Compusemble’s Forspoken DirectStorage benchmark, where across seven scenes combined a PCIe 5.0 drive on a Phison E26 controller and a PCIe 3.0 drive differed by about 3 seconds in total load time. Tom’s Hardware called that gap unnoticeable.

Its own 2026 test of 11 games on SATA, PCIe 3.0, PCIe 4.0, and PCIe 5.0 SSDs reached the same kind of answer for loading. Between PCIe generations the differences were small, and flagship drives often gave only marginal gains in load time despite much higher sequential throughput. In Forspoken at 1080p, gameplay showed peaks of 1,324MB/s on the PCIe 3.0 drive, 2,956MB/s on PCIe 4.0, and 2,761MB/s on PCIe 5.0, but averages of 301MB/s, 374MB/s, and 307MB/s. Across the games, average disk bandwidth stayed similar from drive to drive, with only brief peaks.

In Figure 6 the peak columns climb with each generation while the average columns stay close together. The gameplay load never asks the faster drives to hold their peak.

What to do with this

Neither set gives me a retail PCIe 4.0 and PCIe 5.0 pair on the same controller family, so I can’t separate the interface from the drive’s controller and flash. There are no wall-clock load times either, only replayed traces from pre-release samples and write tests on the three retail Gen5 drives. The press numbers come from one DirectStorage game and one 11-game test, on specific systems. I wouldn’t read the 3 seconds as a rule for every game, and a game that decompresses on the CPU may weigh the drive differently.

Don’t pay extra for PCIe 5.0 because of load times. My bench says the trace tracks the drive more than the interface. The retail Gen5 drives I wrote to led same-capacity PCIe 4.0 drives by as little as 1.02 times once the cache was full, and the published load tests found small gaps between generations. If a game uses DirectStorage, the largest gain Microsoft shows comes from moving decompression to the GPU.

If you are choosing between drives for a game library, look past the sequential number. Check what the drive does after its cache fills in the speed on the box is the speed you get, and look for random read results at low queue depths. The Parallelism Lab shows why a drive needs many commands in flight before its dies can all work, which is what a low queue depth does not give them.