CrystalDiskMark Says 3700MB/s But Windows Copy Shows 1000? Here's Why

Your SSD isn't broken. CrystalDiskMark measures peak burst speed under ideal lab conditions — one direction, large blocks, no file system overhead. Windows file copy is a real-world operation that reads, writes, updates metadata, and runs error checks simultaneously. The gap between them is normal, expected, and unavoidable. Here's exactly what causes it and what you can do about it.

What Is CrystalDiskMark Actually Measuring?

CrystalDiskMark (CDM) is a synthetic benchmark. It tests your SSD's theoretical maximum by removing nearly all real-world complexity:

• One direction at a time. CDM tests read OR write — never both simultaneously. A file copy does both at once.
• Large sequential blocks. The default SEQ1M test reads/writes in 1MiB chunks — ideal for NVMe controllers to process efficiently.
• High queue depth and threading. The SEQ1M Q8T1 test sends 8 requests in parallel, keeping the SSD's command queue full. This is how NVMe drives achieve their rated speeds — but it's not how Windows Explorer copies files.
• Minimal file system interaction. CDM writes to a single test file. It doesn't create thousands of separate files, update directory entries, or write NTFS journal logs.
• Short bursts. The default 1GiB test file often completes within the SSD's SLC cache — meaning CDM may never hit the drive's slower native write speed.

The result is a number that reflects what your SSD can do under perfect conditions. It's useful for comparing drives and verifying hardware health, but it doesn't predict real-world copy speeds.

Source: CrystalDiskMark's official documentation notes that results depend on test file size, controller type, and CPU speed, and that small differences between tools are normal.

Why Windows File Explorer Isn't Built for Raw Speed

Windows Explorer's copy engine is designed for reliability and compatibility, not maximum throughput. Several architectural decisions directly limit its speed:

Single-threaded I/O scheduling. Explorer processes files one at a time, even when copying thousands of small files from scattered directories. On NVMe SSDs with parallel NAND channels, this leaves most of the drive's bandwidth unused. Analysis using Microsoft Sysinternals Process Monitor shows Explorer spends a significant portion of total copy time waiting for I/O queue arbitration — time that a multi-threaded copier would use to push more data.

Fixed buffer size. Explorer uses a relatively small I/O buffer regardless of your drive type, available RAM, or interface speed. Whether you're copying to a USB thumb drive or a PCIe Gen 5 NVMe SSD, Explorer uses the same buffer strategy. Third-party tools like TeraCopy dynamically adjust buffer sizes based on real-time throughput, reducing context switches and increasing effective bandwidth.

File system overhead on every file. For each file copied, Windows must:

• Read the source file data
• Allocate space on the destination drive
• Write the file data
• Update the NTFS Master File Table (MFT) — a hidden index of every file on the drive
• Write an NTFS journal entry for crash recovery
• Update directory metadata (timestamps, file size, attributes)

For a single large file, this overhead is negligible relative to the data transferred. For 10,000 small files, it becomes the dominant bottleneck — each file triggers the full cycle regardless of whether it's 4KB or 4GB.

Windows Defender real-time scanning. By default, Windows Defender scans every file as it's created. During a large copy operation, this adds CPU overhead and I/O latency on every single file. The actual performance impact varies by workload and system configuration.

Progress calculation overhead. Explorer pre-scans the entire file list before starting the copy, calculates estimated time remaining during the transfer, and updates the progress UI — all of which consume CPU and I/O resources that could otherwise go to the actual copy.

The SLC Cache Trap — Why Speed Drops Mid-Transfer

This is the single biggest reason for the dramatic gap between CDM scores and sustained copy speeds, and it has nothing to do with Windows.

Modern consumer SSDs use TLC (Triple-Level Cell) or QLC (Quad-Level Cell) NAND flash memory. Writing 3 or 4 bits per cell is cost-effective for capacity but inherently slower than writing 1 bit per cell. To mask this, manufacturers reserve a portion of the drive as pseudo-SLC (pSLC) cache — programming those cells to store just 1 bit, which is much faster.

Here's what happens during a file copy:

1. First few seconds: Data writes to the SLC cache at full rated speed (e.g., 3700MB/s). CrystalDiskMark's 1GiB default test often finishes entirely within this window.
2. Cache fills up: Depending on the drive, the SLC cache might be 50GB to 200GB (some drives allocate dynamically based on free space).
3. Speed cliff: Once the cache is full, the drive writes directly to slower TLC/QLC NAND while simultaneously trying to flush the cache in the background. Write speeds can plummet from several GB/s down to 200–500MB/s.

This is documented extensively. As How-To Geek explains: once the SLC cache is exhausted, "you will witness a sheer cliff drop in your transfer speeds, often plummeting from gigabytes per second down to a couple of hundred megabytes per second." No software tweak can change this — it's a physical limitation of the NAND.

Why CDM doesn't show this: CDM's default 1GiB test file is small enough to fit inside most SLC caches entirely. The benchmark finishes before the drive ever has to write to slower native NAND. To see your drive's post-cache speed, increase CDM's test size to 16GiB or 32GiB — you'll likely see write speeds drop significantly.

How DRAM Cache Helps

Not all SSDs handle this equally. Drives with a dedicated DRAM cache chip (separate from SLC cache) maintain a lookup table of where data is stored on the NAND. This allows the controller to manage writes more efficiently, especially during sustained workloads. DRAM-less drives rely on the slower Host Memory Buffer (HMB) protocol, borrowing a small portion of your system RAM instead.

If sustained write performance matters to you — video editing, game installation, large backups — a drive with DRAM cache will hold speed more consistently after the SLC cache is exhausted. For example, the OSCOO ON1000 PRO includes a dedicated DRAM cache alongside its NVMe Gen 4 controller, specifically to maintain more stable write performance under sustained loads.

Same-Drive Copy: The Worst-Case Scenario

Copying files from one folder to another on the same SSD is the slowest possible scenario:

• The controller must handle read and write requests simultaneously, splitting its bandwidth.
• On a single NVMe drive, this effectively halves your maximum throughput — or worse, because the controller also has to manage internal garbage collection, wear leveling, and cache flushing concurrently.
• With thousands of small files, the drive is bombarded with tiny random read and write requests. As HardForum users have documented, copying 41,000 small files totaling 1GB on the same NVMe drive can result in speeds below 1MB/s — not because the drive is broken, but because the per-file overhead completely dominates the actual data transfer.

Tip: If you need to reorganize files on the same drive, move instead of copy when staying on the same partition. A move within the same partition only updates the MFT pointers — no data is actually transferred, so it completes nearly instantly regardless of file size.

Other Hidden Bottlenecks

Even with a perfect SSD and efficient copy software, other factors can cap your speed:

Thermal throttling. NVMe SSDs without heatsinks can reach 80–100°C under sustained writes. The controller will automatically reduce speed to prevent damage. Adding a heatsink — most desktop motherboards include one — eliminates this in most cases.

PCIe generation mismatch. A Gen 4 SSD in a Gen 3 slot is limited to ~3,500MB/s theoretical maximum (less in practice). A Gen 5 SSD in a Gen 4 slot is similarly capped. Check your motherboard specifications.

USB bottleneck for external drives. If you're using an NVMe SSD in an external enclosure, the USB interface becomes the ceiling:

Interface	Max Theoretical	Real-World Max
USB 3.2 Gen 1 (5Gbps)	625 MB/s	~400–450 MB/s
USB 3.2 Gen 2 (10Gbps)	1,250 MB/s	~900–1,000 MB/s
USB 3.2 Gen 2x2 (20Gbps)	2,500 MB/s	~1,700–2,000 MB/s
USB4 / Thunderbolt 3 (40Gbps)	5,000 MB/s	~2,800–3,500 MB/s

If your CrystalDiskMark shows 3700MB/s on the internal drive but your external copy maxes out at 1000MB/s, the USB 3.2 Gen 2 port is likely the bottleneck — not the SSD. Upgrading to a USB4 enclosure like the OSCOO XT6 (40Gbps) removes this limitation for most Gen 4 drives.

Source drive speed. A file copy is only as fast as the slowest link in the chain. Copying from an older SATA SSD (max ~550MB/s) to a fast NVMe drive will be capped at ~550MB/s regardless of how fast the NVMe is.

Background processes. Windows Search indexing, OneDrive sync, antivirus scans, and system updates all compete for disk I/O. Check Task Manager → Performance → Disk during a copy to see if other processes are consuming bandwidth.

So Is My SSD Actually Working Fine?

Use this quick diagnostic:

Test	Healthy Result	Possible Issue
CDM Sequential Read	Within 10–15% of advertised spec	If significantly lower: check PCIe slot generation, driver, thermal throttling
CDM Sequential Write	Within 10–15% of advertised spec	Same as above; also check if drive is nearly full (reduces SLC cache)
Windows copy — single 10GB+ file (NVMe → different NVMe)	1,000–2,500 MB/s	If below 500MB/s: check for background processes, driver issues, thermal throttling
Windows copy — thousands of small files	100–500 MB/s	Normal — file system overhead dominates
Speed drops sharply mid-transfer	Normal — SLC cache exhaustion	Only concerning if it drops below 100MB/s on a modern NVMe

If your CrystalDiskMark results are close to the advertised specifications, your SSD is working correctly. The gap between benchmark scores and real-world copy speeds is inherent to how file systems, operating systems, and NAND flash memory work.

How to Get Faster Real-World Transfers

You can't eliminate the gap entirely, but you can narrow it:

1. Use a faster copy tool. FastCopy minimizes kernel overhead and uses asynchronous I/O — it's widely regarded as the fastest free file copier on Windows. TeraCopy integrates directly into the Windows shell (replacing Ctrl+C/V behavior) and uses adaptive buffer sizing. Both handle errors more gracefully than Explorer, skipping problem files instead of aborting the entire transfer.
2. Copy large files, not many small ones. If you're transferring a folder with thousands of small files, compress it into a single .zip or .7z archive first, copy the archive, then extract at the destination. This dramatically reduces per-file overhead.
3. Keep your SSD under 80% capacity. As an SSD fills up, the dynamic SLC cache shrinks (less empty space to repurpose as pseudo-SLC). A nearly full drive will hit the speed cliff much sooner during sustained writes.
4. Add a heatsink. Especially for desktop M.2 SSDs without thermal pads or heatsinks. Most modern motherboards include an M.2 heatsink — make sure it's installed.
5. Match your interfaces. Don't pair a Gen 4 NVMe SSD with a USB 3.0 enclosure. Make sure your USB cable, port, and enclosure all support the same maximum speed.
6. Temporarily disable real-time antivirus scanning during very large transfers. This can help in some scenarios, though the actual impact varies. Re-enable it immediately after the transfer completes.

Get more best practices

The Bottom Line

CrystalDiskMark tells you what your SSD is capable of in a controlled, synthetic test. Windows file copy tells you how fast your storage system works in the real world — including file system overhead, SLC cache limits, and software bottlenecks.

The 3700MB/s in CDM and the 1000MB/s in Windows Explorer aren't contradictory. They're measuring different things. Your SSD is almost certainly fine.

---

Want to learn more about SSD performance? Check out our guide to M.2 SSD sizes explained or browse OSCOO's full internal SSD lineup.